fix: address PR #63 review findings

1. Refactor err2 to use scoped loadErr variable (MINOR - sonnet-review-bot)
   The else-if branches are mutually exclusive, so the error variable
   should be scoped inside the block, not declared outside with err2.

2. Sanitize DisplayName before embedding in Markdown (MINOR - security-review-bot)
   Remote persona metadata is untrusted. Added sanitizeMarkdownText() to
   escape Markdown special characters and strip control characters.
   Applied to both the header title and the footer attribution.

3. Document YAML DoS mitigations (MINOR - security-review-bot)
   Added comprehensive comment in remote_persona.go explaining existing
   defenses: file size limit, file count cap, depth limit, node count cap,
   and alias cycle detection. These collectively mitigate billion-laughs
   and stack exhaustion attacks.

.gitea/actions/review/action.yml

@@ -26,18 +26,40 @@ inputs:
     required: false
     default: ''
   llm-base-url:
-    description: 'OpenAI-compatible LLM API base URL'
-    required: true
+    description: 'OpenAI-compatible LLM API base URL (not required for aicore provider)'
+    required: false
+    default: ''
   llm-api-key:
-    description: 'LLM API key'
-    required: true
+    description: 'LLM API key (not required for aicore provider)'
+    required: false
+    default: ''
   llm-model:
     description: 'LLM model name'
     required: true
   llm-provider:
-    description: 'LLM API provider: openai or anthropic (default openai)'
+    description: 'LLM API provider: openai, anthropic, or aicore (default openai)'
     required: false
-    default: 'openai' 
+    default: 'openai'
+  aicore-client-id:
+    description: 'SAP AI Core client ID (required for aicore provider)'
+    required: false
+    default: ''
+  aicore-client-secret:
+    description: 'SAP AI Core client secret (required for aicore provider)'
+    required: false
+    default: ''
+  aicore-auth-url:
+    description: 'SAP AI Core authentication URL (required for aicore provider)'
+    required: false
+    default: ''
+  aicore-api-url:
+    description: 'SAP AI Core API URL (required for aicore provider)'
+    required: false
+    default: ''
+  aicore-resource-group:
+    description: 'SAP AI Core resource group (default: default)'
+    required: false
+    default: 'default'
   conventions-file:
     description: 'Path to conventions file in the repo (e.g. CLAUDE.md)'
     required: false
@@ -74,6 +96,14 @@ inputs:
     description: 'Local file with additional system prompt instructions (e.g. security review focus)'
     required: false
     default: ''
+  persona:
+    description: 'Built-in persona name (security, architect, docs)'
+    required: false
+    default: ''
+  persona-file:
+    description: 'Path to custom persona JSON file'
+    required: false
+    default: ''
 
 runs:
   using: 'composite'
@@ -155,6 +185,13 @@ runs:
         LLM_PROVIDER: ${{ inputs.llm-provider }}
         UPDATE_EXISTING: ${{ inputs.update-existing }}
         SYSTEM_PROMPT_FILE: ${{ inputs.system-prompt-file }}
+        PERSONA: ${{ inputs.persona }}
+        PERSONA_FILE: ${{ inputs.persona-file }}
+        AICORE_CLIENT_ID: ${{ inputs.aicore-client-id }}
+        AICORE_CLIENT_SECRET: ${{ inputs.aicore-client-secret }}
+        AICORE_AUTH_URL: ${{ inputs.aicore-auth-url }}
+        AICORE_API_URL: ${{ inputs.aicore-api-url }}
+        AICORE_RESOURCE_GROUP: ${{ inputs.aicore-resource-group }}
       run: |
         ARGS=""
         if [ "${{ inputs.dry-run }}" = "true" ]; then

.gitea/workflows/ci.yml

@@ -18,7 +18,10 @@ jobs:
       - run: go vet ./...
       - run: go build -o review-bot ./cmd/review-bot
 
-  # Self-review: builds from source since we're pre-release
+  # Self-review using native SAP AI Core provider
+  # Models must match SAP AI Core deployments
+  # Available models: gpt-5, anthropic--claude-4.6-sonnet, anthropic--claude-4.6-opus
+  # Removed gpt-4.1, gpt-5-mini, gpt-4.1-mini - not deployed on AI Core
   review:
     runs-on: ubuntu-24.04
     if: github.event_name == 'pull_request'
@@ -28,10 +31,10 @@ jobs:
         include:
           - name: sonnet
             token_secret: SONNET_REVIEW_TOKEN
-            model: gpt-5
+            model: anthropic--claude-4.6-sonnet
           - name: gpt
             token_secret: GPT_REVIEW_TOKEN
-            model: gpt-4.1
+            model: gpt-5
           - name: security
             token_secret: SECURITY_REVIEW_TOKEN
             model: gpt-5
@@ -49,9 +52,13 @@ jobs:
           PR_NUMBER: ${{ github.event.pull_request.number }}
           REVIEWER_TOKEN: ${{ secrets[matrix.token_secret] }}
           REVIEWER_NAME: ${{ matrix.name }}
-          LLM_BASE_URL: ${{ secrets.LLM_BASE_URL }}
-          LLM_API_KEY: ${{ secrets.LLM_API_KEY }}
+          LLM_PROVIDER: aicore
           LLM_MODEL: ${{ matrix.model }}
+          AICORE_CLIENT_ID: ${{ secrets.AICORE_CLIENT_ID }}
+          AICORE_CLIENT_SECRET: ${{ secrets.AICORE_CLIENT_SECRET }}
+          AICORE_AUTH_URL: ${{ secrets.AICORE_AUTH_URL }}
+          AICORE_API_URL: ${{ secrets.AICORE_API_URL }}
+          AICORE_RESOURCE_GROUP: ${{ secrets.AICORE_RESOURCE_GROUP }}
           CONVENTIONS_FILE: "CONVENTIONS.md"
           PATTERNS_REPO: "rodin/go-patterns"
           PATTERNS_FILES: "README.md,patterns/"

.gitea/workflows/pr-ready-gate.yml

@@ -0,0 +1,38 @@
+name: PR Ready Gate
+
+on:
+  pull_request:
+    types: [synchronize]
+
+jobs:
+  clear-labels:
+    runs-on: ubuntu-24.04
+    # Always run - curl commands are safe if labels don't exist
+    steps:
+      - name: Remove ready and self-reviewed labels, reassign to author
+        env:
+          GITEA_TOKEN: ${{ secrets.RODIN_TOKEN }}
+        run: |
+          PR_NUMBER=${{ github.event.pull_request.number }}
+          AUTHOR=${{ github.event.pull_request.user.login }}
+          READY_LABEL_ID=38
+          SELF_REVIEWED_LABEL_ID=37
+          
+          # Remove ready label if present
+          curl -sS -X DELETE \
+            -H "Authorization: token $GITEA_TOKEN" \
+            "https://gitea.weiker.me/api/v1/repos/${{ github.repository }}/issues/${PR_NUMBER}/labels/${READY_LABEL_ID}" || true
+          
+          # Remove self-reviewed label if present
+          curl -sS -X DELETE \
+            -H "Authorization: token $GITEA_TOKEN" \
+            "https://gitea.weiker.me/api/v1/repos/${{ github.repository }}/issues/${PR_NUMBER}/labels/${SELF_REVIEWED_LABEL_ID}" || true
+          
+          # Reassign to author
+          curl -sS -X PATCH \
+            -H "Authorization: token $GITEA_TOKEN" \
+            -H "Content-Type: application/json" \
+            -d "{\"assignees\": [\"${AUTHOR}\"]}" \
+            "https://gitea.weiker.me/api/v1/repos/${{ github.repository }}/pulls/${PR_NUMBER}"
+          
+          echo "Cleared ready/self-reviewed labels and reassigned PR #${PR_NUMBER} to ${AUTHOR}"

CONVENTIONS.md

@@ -2,8 +2,26 @@
 
 ## Language & Dependencies
 
-- Go standard library only — no external dependencies.
 - Target the latest stable Go release.
+- **STRICT ALLOWLIST:** Only packages listed below may be imported. No exceptions.
+
+### Approved Third-Party Packages
+
+| Package | Use Case | Scope |
+|---------|----------|-------|
+| `gopkg.in/yaml.v3` | YAML parsing (persona files, config) | production |
+| `github.com/google/go-cmp` | Test comparisons (`cmp.Diff`) | test only |
+
+**Any import not in this table or the Go standard library is forbidden.**
+
+Transitive dependencies of approved packages are automatically allowed.
+
+To request a new dependency:
+1. Open a PR that ONLY updates this table
+2. Requires explicit approval from Aaron
+3. After merge, a separate PR may use the package
+
+*Enforcement: `scripts/check-deps.sh` parses this table — update only here.*
 
 ## Error Handling
 

Makefile

@@ -1,4 +1,4 @@
-.PHONY: build test test-integration lint clean coverage
+.PHONY: build test test-integration lint clean coverage check-deps precommit
 
 build:
 	go build -o review-bot ./cmd/review-bot/
@@ -12,9 +12,15 @@ test-integration:
 lint:
 	go vet ./...
 
+check-deps:
+	@./scripts/check-deps.sh
+
 clean:
 	rm -f review-bot
 
 coverage:
 	go test -coverprofile=coverage.out ./...
 	go tool cover -func=coverage.out
+
+# Precommit runs all checks required before pushing
+precommit: check-deps lint test

README.md

@@ -4,12 +4,12 @@ AI-powered code review bot for Gitea pull requests. Fetches diff + context, send
 
 ## Features
 
-- **Multi-provider**: OpenAI-compatible and Anthropic Messages API
+- **Multi-provider**: OpenAI-compatible, Anthropic Messages API, and SAP AI Core
 - **Context-aware**: Fetches full file content, conventions, language patterns, CI status
 - **Smart budget**: Automatically trims context to fit model token limits
 - **Idempotent reviews**: Posts new review, then cleans up stale ones (one review per bot)
 - **Custom prompts**: Load additional instructions from a file (e.g. security-focused review)
-- **Zero dependencies**: Go stdlib only
+- **Minimal dependencies**: Go stdlib + `gopkg.in/yaml.v3` only
 
 ## Quick Start: Composite Action
 
@@ -168,26 +168,56 @@ Prints the review to CI logs without posting to the PR. Useful for testing promp
     llm-provider: anthropic
 ```
 
+### Using SAP AI Core
+
+For SAP environments with AI Core deployments, use the `aicore` provider for native authentication:
+
+```yaml
+- uses: https://gitea.weiker.me/rodin/review-bot/.gitea/actions/review@v0.1.0
+  with:
+    reviewer-token: ${{ secrets.REVIEW_TOKEN }}
+    reviewer-name: aicore-review
+    llm-model: anthropic--claude-4.6-sonnet  # or gpt-5
+    llm-provider: aicore
+    aicore-client-id: ${{ secrets.AICORE_CLIENT_ID }}
+    aicore-client-secret: ${{ secrets.AICORE_CLIENT_SECRET }}
+    aicore-auth-url: ${{ secrets.AICORE_AUTH_URL }}
+    aicore-api-url: ${{ secrets.AICORE_API_URL }}
+    aicore-resource-group: default
+```
+
+AI Core handles OAuth token management and deployment discovery automatically. Model names must match the deployment name in AI Core (e.g. `anthropic--claude-4.6-sonnet`, `gpt-5`).
+
 ## Action Inputs
 
 | Input | Required | Default | Description |
 |-------|----------|---------|-------------|
 | `reviewer-token` | Yes | — | Gitea token for posting reviews (needs `write:issue`, `write:repository`) |
 | `reviewer-name` | No | `""` | Logical identity for this reviewer. Used as sentinel for idempotent cleanup. Set this when running multiple review bots on the same PR. |
-| `llm-base-url` | Yes | — | LLM API base URL |
-| `llm-api-key` | Yes | — | LLM API key |
+| `llm-base-url` | No* | `""` | LLM API base URL (required unless using aicore provider) |
+| `llm-api-key` | No* | `""` | LLM API key (required unless using aicore provider) |
 | `llm-model` | Yes | — | Model name |
-| `llm-provider` | No | `openai` | API provider: `openai` or `anthropic` |
+| `llm-provider` | No | `openai` | API provider: `openai`, `anthropic`, or `aicore` |
+| `aicore-client-id` | No** | `""` | SAP AI Core client ID |
+| `aicore-client-secret` | No** | `""` | SAP AI Core client secret |
+| `aicore-auth-url` | No** | `""` | SAP AI Core authentication URL |
+| `aicore-api-url` | No** | `""` | SAP AI Core API URL |
+| `aicore-resource-group` | No | `default` | SAP AI Core resource group |
 | `conventions-file` | No | `""` | Path to coding conventions file in the repo |
 | `patterns-repo` | No | `""` | Comma-separated repos with language patterns (e.g. `rodin/go-patterns`) |
 | `patterns-files` | No | `README.md` | Files/directories to fetch from pattern repos |
 | `system-prompt-file` | No | `""` | Local file with additional system prompt instructions |
+| `persona` | No | `""` | Built-in persona name (security, architect, docs) |
+| `persona-file` | No | `""` | Path to persona file (YAML or JSON) with custom review focus |
 | `temperature` | No | `0` | LLM temperature (0 = server default) |
 | `timeout` | No | `300` | LLM request timeout in seconds |
 | `dry-run` | No | `false` | Print review to stdout instead of posting |
 | `update-existing` | No | `true` | Delete previous review from same bot before posting. Accepts: true/1/yes or false/0/no |
 | `version` | No | `latest` | review-bot version to install |
 
+*Required for `openai` and `anthropic` providers, not for `aicore`.
+**Required only for `aicore` provider.
+
 ## Runner Requirements
 
 The composite action requires these tools on the runner:
@@ -329,3 +359,151 @@ budget/             Token estimation + context trimming
 ## License
 
 MIT
+
+## Review Personas
+
+Personas provide role-based review specialization. Instead of generic code review, each persona focuses on a specific domain (security, architecture, documentation) with tailored prompts and severity calibration.
+
+### Built-in Personas
+
+| Persona | Focus |
+|---------|-------|
+| `security` | Vulnerabilities, auth bypass, secrets exposure, injection attacks |
+| `architect` | Design patterns, code organization, API contracts, testability |
+| `docs` | Documentation quality, API clarity, error messages |
+
+### Using Built-in Personas
+
+```yaml
+- uses: rodin/review-bot/.gitea/actions/review@v1
+  with:
+    reviewer-name: security
+    persona: security
+    llm-model: claude-opus-4-20250514  # Security benefits from strong reasoning
+    ...
+```
+
+### Multiple Personas in Parallel
+
+```yaml
+jobs:
+  review:
+    strategy:
+      matrix:
+        include:
+          - name: security
+            persona: security
+          - name: architect
+            persona: architect
+    steps:
+      - uses: rodin/review-bot/.gitea/actions/review@v1
+        with:
+          reviewer-name: ${{ matrix.name }}
+          persona: ${{ matrix.persona }}
+          ...
+```
+
+Each persona posts independently with its own sentinel, so reviews don't interfere.
+
+
+### Custom Personas
+
+Create a YAML file with your domain-specific review focus:
+
+```yaml
+# .review/personas/trading.yaml
+name: trading
+display_name: Trading Domain Expert
+
+identity: |
+  You are a trading systems expert reviewing code for correctness.
+
+  Your expertise:
+  - Order lifecycle and state machines
+  - Fill handling and partial fills
+  - Position tracking and P&L calculations
+  - Event sourcing invariants
+
+focus:
+  - Order state machine correctness
+  - Fill handling edge cases (partial, overfill)
+  - Position and P&L calculation accuracy
+  - Event replay determinism
+  - Decimal precision for money
+
+ignore:
+  - Code style
+  - General performance
+  - Documentation formatting
+
+severity:
+  major: "Bugs that cause incorrect positions, fills, or money calculations"
+  minor: "Edge cases that could cause issues under unusual conditions"
+  nit: "Clarity improvements for domain logic"
+```
+
+Use it in CI:
+
+```yaml
+- uses: rodin/review-bot/.gitea/actions/review@v1
+  with:
+    reviewer-name: trading
+    persona-file: .review/personas/trading.yaml
+    ...
+```
+
+YAML is the recommended format for personas because it supports:
+- Multi-line strings with `|` blocks (cleaner identity definitions)
+- Comments for documentation
+- More readable arrays and nested structures
+
+JSON is also supported for backwards compatibility—just use `.json` extension.
+
+### Repository Personas (Auto-Discovery)
+
+Repositories can ship their own personas in `.review-bot/personas/`. When you specify `--persona <name>`, review-bot will:
+
+1. **Try to load from the target repo** — Checks `.review-bot/personas/<name>.yaml` (or `.yml`)
+2. **Fall back to built-in** — If not found in repo, uses the built-in persona
+
+This lets each repo define domain-specific personas without modifying CI config:
+
+```
+my-trading-repo/
+├── .review-bot/
+│   └── personas/
+│       ├── trading.yaml      # Custom trading persona
+│       └── regulatory.yaml   # Compliance-focused reviews
+├── lib/
+└── ...
+```
+
+```yaml
+# CI config (no persona-file needed)
+- uses: rodin/review-bot/.gitea/actions/review@v1
+  with:
+    reviewer-name: trading
+    persona: trading           # Will find .review-bot/personas/trading.yaml
+    ...
+```
+
+**Priority order:**
+1. Repo's `.review-bot/personas/<name>.yaml`
+2. Built-in persona with matching name
+3. Error if neither exists
+
+This allows repos to override built-in personas (e.g., a custom `security` persona that adds project-specific rules) while keeping the simple `persona: security` syntax in CI.
+
+
+### Persona vs system-prompt-file
+
+| Feature | `persona` / `persona-file` | `system-prompt-file` |
+|---------|---------------------------|----------------------|
+| Replaces base prompt | Yes | No (appends) |
+| Structured format | Yes (YAML/JSON) | No (freeform) |
+| Focus/ignore lists | Yes | Manual |
+| Severity calibration | Yes | Manual |
+| Header display name | Yes | No |
+| Built-in options | Yes | No |
+
+Use personas for domain-specialized reviews. Use `system-prompt-file` for minor tweaks to the generic review.

cmd/review-bot/main.go

@@ -69,7 +69,15 @@ func main() {
 	dryRun := flag.Bool("dry-run", false, "Print review to stdout instead of posting")
 	llmTemp := flag.Float64("llm-temperature", envOrDefaultFloat("LLM_TEMPERATURE", 0), "LLM temperature (0 = server default)")
 	llmTimeout := flag.Int("llm-timeout", envOrDefaultInt("LLM_TIMEOUT", 300), "LLM request timeout in seconds (default 300)")
-	llmProvider := flag.String("llm-provider", envOrDefault("LLM_PROVIDER", "openai"), "LLM API provider: openai or anthropic")
+	llmProvider := flag.String("llm-provider", envOrDefault("LLM_PROVIDER", "openai"), "LLM API provider: openai, anthropic, or aicore")
+	personaName := flag.String("persona", envOrDefault("PERSONA", ""), "Built-in persona name (security, architect, docs)")
+	personaFile := flag.String("persona-file", envOrDefault("PERSONA_FILE", ""), "Path to persona JSON file")
+	// AI Core specific flags (only used when provider=aicore)
+	aicoreClientID := flag.String("aicore-client-id", envOrDefault("AICORE_CLIENT_ID", ""), "SAP AI Core client ID (for provider=aicore)")
+	aicoreClientSecret := flag.String("aicore-client-secret", envOrDefault("AICORE_CLIENT_SECRET", ""), "SAP AI Core client secret (for provider=aicore)")
+	aicoreAuthURL := flag.String("aicore-auth-url", envOrDefault("AICORE_AUTH_URL", ""), "SAP AI Core auth URL (for provider=aicore)")
+	aicoreAPIURL := flag.String("aicore-api-url", envOrDefault("AICORE_API_URL", ""), "SAP AI Core API URL (for provider=aicore)")
+	aicoreResourceGroup := flag.String("aicore-resource-group", envOrDefault("AICORE_RESOURCE_GROUP", "default"), "SAP AI Core resource group (for provider=aicore)")
 
 	flag.Parse()
 
@@ -84,12 +92,32 @@ func main() {
 	slog.Info("review-bot starting", "version", version)
 
 	// Validate required fields
-	if *giteaURL == "" || *repo == "" || *prNum == "" || *reviewerToken == "" ||
-		*llmBaseURL == "" || *llmAPIKey == "" || *llmModel == "" {
+	// For aicore provider, llm-base-url and llm-api-key are not required
+	isAICore := llm.Provider(*llmProvider) == llm.ProviderAICore
+	if *giteaURL == "" || *repo == "" || *prNum == "" || *reviewerToken == "" || *llmModel == "" {
 		fmt.Fprintf(os.Stderr, "Error: missing required flags or environment variables\n\n")
-		fmt.Fprintf(os.Stderr, "Required: --gitea-url, --repo, --pr, --reviewer-token, --llm-base-url, --llm-api-key, --llm-model\n")
+		fmt.Fprintf(os.Stderr, "Required: --gitea-url, --repo, --pr, --reviewer-token, --llm-model\n")
 		os.Exit(1)
 	}
+	if !isAICore && (*llmBaseURL == "" || *llmAPIKey == "") {
+		fmt.Fprintf(os.Stderr, "Error: --llm-base-url and --llm-api-key are required for provider=%s\n", *llmProvider)
+		os.Exit(1)
+	}
+	if isAICore && (*aicoreClientID == "" || *aicoreClientSecret == "" || *aicoreAuthURL == "" || *aicoreAPIURL == "") {
+		fmt.Fprintf(os.Stderr, "Error: AI Core credentials required for provider=aicore\n\n")
+		fmt.Fprintf(os.Stderr, "Required: --aicore-client-id, --aicore-client-secret, --aicore-auth-url, --aicore-api-url\n")
+		os.Exit(1)
+	}
+
+	// Validate persona flags are mutually exclusive
+	if *personaName != "" && *personaFile != "" {
+		slog.Error("--persona and --persona-file are mutually exclusive")
+		os.Exit(1)
+	}
+
+	// Persona loading is deferred until after giteaClient is initialized,
+	// so we can try loading from the target repo first.
+	var persona *review.Persona
 
 	// Validate reviewer-name: only safe characters allowed in sentinel
 	if err := validateReviewerName(*reviewerName); err != nil {
@@ -125,8 +153,17 @@ func main() {
 	switch llm.Provider(*llmProvider) {
 	case llm.ProviderOpenAI, llm.ProviderAnthropic:
 		llmClient.WithProvider(llm.Provider(*llmProvider))
+	case llm.ProviderAICore:
+		llmClient.WithAICore(llm.AICoreConfig{
+			ClientID:      *aicoreClientID,
+			ClientSecret:  *aicoreClientSecret,
+			AuthURL:       *aicoreAuthURL,
+			APIURL:        *aicoreAPIURL,
+			ResourceGroup: *aicoreResourceGroup,
+		})
+		slog.Info("using SAP AI Core provider", "resource_group", *aicoreResourceGroup)
 	default:
-		slog.Error("invalid LLM provider", "provider", *llmProvider, "valid", "openai, anthropic")
+		slog.Error("invalid LLM provider", "provider", *llmProvider, "valid", "openai, anthropic, aicore")
 		os.Exit(1)
 	}
 	if *llmTimeout > 0 {
@@ -138,6 +175,45 @@ func main() {
 	ctx, cancel := context.WithTimeout(context.Background(), overallTimeout)
 	defer cancel()
 
+	// Load persona: try remote repo first, then fall back to built-in
+	if *personaName != "" {
+		// Try loading from target repo's .review-bot/personas/ directory
+		fetcher := &giteaFetcher{client: giteaClient}
+		remotePersonas, err := review.LoadRemotePersonas(ctx, fetcher, owner, repoName)
+		if err != nil {
+			slog.Warn("could not load remote personas", "repo", fmt.Sprintf("%s/%s", owner, repoName), "error", err)
+			// Assign empty map so the lookup below doesn't panic
+			remotePersonas = map[string]*review.Persona{}
+		}
+
+		if p, ok := remotePersonas[*personaName]; ok {
+			persona = p
+			slog.Info("loaded persona from target repo", "persona", persona.Name, "display", persona.DisplayName)
+		} else {
+			// Fall back to built-in persona
+			var err error
+			persona, err = review.LoadBuiltinPersona(*personaName)
+			if err != nil {
+				slog.Error("failed to load persona", "persona", *personaName, "error", err)
+				os.Exit(1)
+			}
+			slog.Info("loaded built-in persona", "persona", persona.Name, "display", persona.DisplayName)
+		}
+	} else if *personaFile != "" {
+		resolvedPath, err := validateWorkspacePath(*personaFile, "persona-file")
+		if err != nil {
+			slog.Error("invalid persona-file path", "error", err)
+			os.Exit(1)
+		}
+		loadedPersona, loadErr := review.LoadPersona(resolvedPath)
+		if loadErr != nil {
+			slog.Error("failed to load persona file", "file", *personaFile, "error", loadErr)
+			os.Exit(1)
+		}
+		persona = loadedPersona
+		slog.Info("loaded persona from file", "file", *personaFile, "persona", persona.Name)
+	}
+
 	slog.Info("reviewing pull request", "pr", prNumber, "repo", fmt.Sprintf("%s/%s", owner, repoName))
 
 	// Step 1: Fetch PR metadata
@@ -201,34 +277,14 @@ func main() {
 	// Step 6b: Load additional system prompt if specified
 	additionalPrompt := ""
 	if *systemPromptFile != "" {
-		workspace := os.Getenv("GITHUB_WORKSPACE")
-		if workspace == "" {
-			workspace, _ = os.Getwd()
-		}
-		absWorkspace, err := filepath.Abs(workspace)
+		resolvedPath, err := validateWorkspacePath(*systemPromptFile, "system-prompt-file")
 		if err != nil {
-			slog.Error("failed to resolve workspace path", "error", err)
-			os.Exit(1)
-		}
-		promptPath := filepath.Join(absWorkspace, *systemPromptFile)
-		promptPath = filepath.Clean(promptPath)
-		if !strings.HasPrefix(promptPath, absWorkspace+string(filepath.Separator)) && promptPath != absWorkspace {
-			slog.Error("system-prompt-file resolves outside workspace", "path", promptPath, "workspace", absWorkspace)
-			os.Exit(1)
-		}
-		// Resolve symlinks and re-validate to prevent symlink traversal
-		resolvedPath, err := filepath.EvalSymlinks(promptPath)
-		if err != nil {
-			slog.Error("failed to resolve system prompt file", "path", promptPath, "error", err)
-			os.Exit(1)
-		}
-		if !strings.HasPrefix(resolvedPath, absWorkspace+string(filepath.Separator)) && resolvedPath != absWorkspace {
-			slog.Error("system-prompt-file symlink resolves outside workspace", "resolved", resolvedPath, "workspace", absWorkspace)
+			slog.Error("invalid system-prompt-file path", "error", err)
 			os.Exit(1)
 		}
 		data, err := os.ReadFile(resolvedPath)
 		if err != nil {
-			slog.Error("failed to read system prompt file", "path", promptPath, "error", err)
+			slog.Error("failed to read system prompt file", "path", *systemPromptFile, "error", err)
 			os.Exit(1)
 		}
 		additionalPrompt = string(data)
@@ -236,7 +292,13 @@ func main() {
 	}
 
 	// Step 7: Budget-aware prompt assembly
-	systemBase := review.BuildSystemBase()
+	var systemBase string
+	if persona != nil {
+		systemBase = review.BuildPersonaSystemPrompt(persona)
+		slog.Debug("using persona system prompt", "persona", persona.Name)
+	} else {
+		systemBase = review.BuildSystemBase()
+	}
 	if additionalPrompt != "" {
 		systemBase += "\n\n## Additional Review Instructions\n\n" + additionalPrompt
 	}
@@ -254,30 +316,51 @@ func main() {
 		slog.Warn("context trimmed to fit budget", "trimmed", budgetResult.Trimmed)
 	}
 
-	// Step 8: Call LLM
+	// Step 8: Call LLM (with retry on parse failure)
 	slog.Info("sending request to LLM", "model", *llmModel)
 	messages := []llm.Message{
 		{Role: "system", Content: budgetResult.SystemPrompt},
 		{Role: "user", Content: budgetResult.UserPrompt},
 	}
 
-	response, err := llmClient.Complete(ctx, messages)
-	if err != nil {
-		slog.Error("LLM request failed", "model", *llmModel, "error", err)
-		os.Exit(1)
-	}
-	slog.Info("LLM response received", "bytes", len(response))
+	var response string
+	var result *review.ReviewResult
+	for attempt := 1; attempt <= 2; attempt++ {
+		if attempt > 1 {
+			slog.Warn("retrying LLM request after parse failure", "attempt", attempt)
+			time.Sleep(time.Second)
+		}
 
-	// Step 9: Parse response
-	result, err := review.ParseResponse(response)
-	if err != nil {
-		slog.Error("failed to parse LLM response", "error", err)
-		os.Exit(1)
+		response, err = llmClient.Complete(ctx, messages)
+		if err != nil {
+			slog.Error("LLM request failed", "model", *llmModel, "error", err, "attempt", attempt)
+			if attempt == 2 {
+				os.Exit(1)
+			}
+			continue
+		}
+		slog.Info("LLM response received", "bytes", len(response), "attempt", attempt)
+
+		// Step 9: Parse response
+		result, err = review.ParseResponse(response)
+		if err != nil {
+			slog.Error("failed to parse LLM response", "error", err, "attempt", attempt)
+			if attempt == 2 {
+				os.Exit(1)
+			}
+			continue
+		}
+		break
 	}
 	slog.Info("review parsed", "verdict", result.Verdict, "findings", len(result.Findings))
 
 	// Step 10: Format and post review
-	reviewBody := review.FormatMarkdown(result, *reviewerName)
+	var reviewBody string
+	if persona != nil && persona.DisplayName != "" {
+		reviewBody = review.FormatMarkdownWithDisplay(result, persona.DisplayName, *reviewerName)
+	} else {
+		reviewBody = review.FormatMarkdown(result, *reviewerName)
+	}
 
 	// Add commit footer so readers know which commit was evaluated
 	if pr.Head.Sha != "" {
@@ -299,6 +382,24 @@ func main() {
 
 	sentinel := fmt.Sprintf("<!-- review-bot:%s -->", *reviewerName)
 
+	// Stale check: verify HEAD hasn't moved since we started
+	evaluatedSHA := pr.Head.Sha
+	var currentSHA string
+	currentPR, err := giteaClient.GetPullRequest(ctx, owner, repoName, prNumber)
+	if err != nil {
+		slog.Warn("could not re-fetch PR for stale check", "pr", prNumber, "error", err)
+		// currentSHA stays empty — shouldSkipStaleReview will return false
+	} else {
+		currentSHA = currentPR.Head.Sha
+	}
+	if shouldSkipStaleReview(evaluatedSHA, currentSHA) {
+		slog.Warn("HEAD moved during review — skipping stale review",
+			"evaluated", evaluatedSHA,
+			"current", currentSHA,
+			"pr", prNumber)
+		return
+	}
+
 	// Map findings to inline comments for lines present in the diff
 	diffRanges := gitea.ParseDiffNewLines(diff)
 	var inlineComments []gitea.ReviewComment
@@ -553,6 +654,43 @@ func validateReviewerName(name string) error {
 	return nil
 }
 
+// validateWorkspacePath ensures a file path is within the workspace and resolves
+// symlinks to prevent traversal attacks. Returns the resolved absolute path or
+// an error if the path is outside the workspace.
+func validateWorkspacePath(path, pathName string) (string, error) {
+	workspace := os.Getenv("GITHUB_WORKSPACE")
+	if workspace == "" {
+		workspace, _ = os.Getwd()
+	}
+	absWorkspace, err := filepath.Abs(workspace)
+	if err != nil {
+		return "", fmt.Errorf("failed to resolve workspace path: %w", err)
+	}
+
+	// Join and clean the path
+	fullPath := filepath.Join(absWorkspace, path)
+	fullPath = filepath.Clean(fullPath)
+
+	// Check path is within workspace using filepath.Rel (more robust than HasPrefix)
+	rel, err := filepath.Rel(absWorkspace, fullPath)
+	if err != nil || strings.HasPrefix(rel, "..") {
+		return "", fmt.Errorf("%s resolves outside workspace: path=%s workspace=%s", pathName, fullPath, absWorkspace)
+	}
+
+	// Resolve symlinks and re-validate to prevent symlink traversal
+	resolvedPath, err := filepath.EvalSymlinks(fullPath)
+	if err != nil {
+		return "", fmt.Errorf("failed to resolve %s: %w", pathName, err)
+	}
+
+	relResolved, err := filepath.Rel(absWorkspace, resolvedPath)
+	if err != nil || strings.HasPrefix(relResolved, "..") {
+		return "", fmt.Errorf("%s symlink resolves outside workspace: resolved=%s workspace=%s", pathName, resolvedPath, absWorkspace)
+	}
+
+	return resolvedPath, nil
+}
+
 // buildSupersededBody creates the body for a superseded review: struck-through banner
 // with collapsed original content and the commit it was evaluated against.
 func buildSupersededBody(originalBody, commitSHA, newReviewURL, sentinel string) string {
@@ -650,3 +788,42 @@ func findAllOwnReviews(reviews []gitea.Review, sentinel string) []gitea.Review {
 	}
 	return result
 }
+
+// shouldSkipStaleReview reports whether to skip posting because HEAD moved.
+// Returns true (skip) if evaluatedSHA differs from currentSHA.
+// Returns false (don't skip) if:
+//   - SHAs match (no movement)
+//   - currentSHA is empty (re-fetch failed; prefer posting stale over failing)
+func shouldSkipStaleReview(evaluatedSHA, currentSHA string) bool {
+	if currentSHA == "" {
+		// Re-fetch failed; better to post potentially stale than fail
+		return false
+	}
+	return evaluatedSHA != currentSHA
+}
+
+// giteaFetcher adapts gitea.Client to review.PersonaFetcher interface.
+type giteaFetcher struct {
+	client *gitea.Client
+}
+
+func (f *giteaFetcher) ListContents(ctx context.Context, owner, repo, path string) ([]review.ContentEntry, error) {
+	entries, err := f.client.ListContents(ctx, owner, repo, path)
+	if err != nil {
+		return nil, err
+	}
+	// Convert gitea.ContentEntry to review.ContentEntry
+	result := make([]review.ContentEntry, len(entries))
+	for i, e := range entries {
+		result[i] = review.ContentEntry{
+			Name: e.Name,
+			Path: e.Path,
+			Type: e.Type,
+		}
+	}
+	return result, nil
+}
+
+func (f *giteaFetcher) GetFileContent(ctx context.Context, owner, repo, filepath string) (string, error) {
+	return f.client.GetFileContent(ctx, owner, repo, filepath)
+}

cmd/review-bot/main_test.go

@@ -6,6 +6,7 @@ import (
 	"log/slog"
 	"os"
 	"os/exec"
+	"path/filepath"
 	"strings"
 	"testing"
 
@@ -45,6 +46,114 @@ func TestValidateReviewerName(t *testing.T) {
 	}
 }
 
+func TestValidateWorkspacePath(t *testing.T) {
+	// Create a temp directory as our workspace
+	tmpDir := t.TempDir()
+
+	// Create a valid file inside the workspace
+	validFile := filepath.Join(tmpDir, "valid.json")
+	if err := os.WriteFile(validFile, []byte("{}"), 0644); err != nil {
+		t.Fatalf("failed to create test file: %v", err)
+	}
+
+	// Create a subdirectory with a file
+	subDir := filepath.Join(tmpDir, "subdir")
+	if err := os.MkdirAll(subDir, 0755); err != nil {
+		t.Fatalf("failed to create subdir: %v", err)
+	}
+	nestedFile := filepath.Join(subDir, "nested.json")
+	if err := os.WriteFile(nestedFile, []byte("{}"), 0644); err != nil {
+		t.Fatalf("failed to create nested file: %v", err)
+	}
+
+	// Create a symlink pointing outside the workspace
+	symlinkPath := filepath.Join(tmpDir, "evil-symlink.json")
+	if err := os.Symlink("/etc/passwd", symlinkPath); err != nil {
+		t.Fatalf("failed to create symlink: %v", err)
+	}
+
+	// Save and restore GITHUB_WORKSPACE
+	origWorkspace := os.Getenv("GITHUB_WORKSPACE")
+	defer os.Setenv("GITHUB_WORKSPACE", origWorkspace)
+
+	tests := []struct {
+		name      string
+		workspace string
+		path      string
+		wantErr   bool
+		errMatch  string
+	}{
+		{
+			name:      "valid relative path",
+			workspace: tmpDir,
+			path:      "valid.json",
+			wantErr:   false,
+		},
+		{
+			name:      "valid nested path",
+			workspace: tmpDir,
+			path:      "subdir/nested.json",
+			wantErr:   false,
+		},
+		{
+			name:      "path traversal attempt",
+			workspace: tmpDir,
+			path:      "../../../etc/passwd",
+			wantErr:   true,
+			errMatch:  "resolves outside workspace",
+		},
+		{
+			name:      "absolute path normalized to workspace-relative",
+			workspace: tmpDir,
+			path:      "/etc/passwd",
+			wantErr:   true,
+			// Go 1.21+ filepath.Join normalizes absolute paths: Join("/tmp/x", "/etc/passwd")
+			// becomes "/tmp/x/etc/passwd", which is within workspace but doesn't exist.
+			errMatch: "failed to resolve",
+		},
+		{
+			name:      "nonexistent file",
+			workspace: tmpDir,
+			path:      "nonexistent.json",
+			wantErr:   true,
+			errMatch:  "failed to resolve",
+		},
+		{
+			name:      "symlink escaping workspace",
+			workspace: tmpDir,
+			path:      "evil-symlink.json",
+			wantErr:   true,
+			errMatch:  "symlink resolves outside workspace",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			os.Setenv("GITHUB_WORKSPACE", tc.workspace)
+			resolved, err := validateWorkspacePath(tc.path, "test-file")
+
+			if tc.wantErr {
+				if err == nil {
+					t.Errorf("expected error for %q, got nil", tc.path)
+				} else if tc.errMatch != "" && !strings.Contains(err.Error(), tc.errMatch) {
+					t.Errorf("error %q should contain %q", err.Error(), tc.errMatch)
+				}
+			} else {
+				if err != nil {
+					t.Errorf("expected no error for %q, got %v", tc.path, err)
+				}
+				if resolved == "" {
+					t.Error("expected non-empty resolved path")
+				}
+				// Verify resolved path is within workspace
+				if !strings.HasPrefix(resolved, tc.workspace) {
+					t.Errorf("resolved path %q not within workspace %q", resolved, tc.workspace)
+				}
+			}
+		})
+	}
+}
+
 func makeReview(id int64, login, state string, stale bool, body string) gitea.Review {
 	r := gitea.Review{
 		ID:    id,
@@ -56,7 +165,6 @@ func makeReview(id int64, login, state string, stale bool, body string) gitea.Re
 	return r
 }
 
-
 func TestBuildSupersededBody(t *testing.T) {
 	original := "# Review\n\nLooks good.\n\n<!-- review-bot:sonnet -->"
 	sentinel := "<!-- review-bot:sonnet -->"
@@ -626,8 +734,8 @@ func TestExtractSentinelName_EdgeCases(t *testing.T) {
 		{"<!-- review-bot:sonnet --> rest", "sonnet"},
 		{"<!-- review-bot:gpt-review --> rest", "gpt-review"},
 		{"no sentinel here", "unknown"},
-		{"<!-- review-bot:", "unknown"},       // prefix but no suffix
-		{"prefix <!-- review-bot:abc --> end", "abc"}, // embedded in text
+		{"<!-- review-bot:", "unknown"},                   // prefix but no suffix
+		{"prefix <!-- review-bot:abc --> end", "abc"},     // embedded in text
 	}
 
 	for _, tc := range tests {
@@ -862,3 +970,53 @@ func TestFindAllOwnReviews(t *testing.T) {
 		}
 	}
 }
+
+func TestShouldSkipStaleReview(t *testing.T) {
+	tests := []struct {
+		name         string
+		evaluatedSHA string
+		currentSHA   string
+		wantSkip     bool
+	}{
+		{
+			name:         "matching SHAs",
+			evaluatedSHA: "abc123def456",
+			currentSHA:   "abc123def456",
+			wantSkip:     false,
+		},
+		{
+			name:         "different SHAs",
+			evaluatedSHA: "abc123def456",
+			currentSHA:   "xyz789abc123",
+			wantSkip:     true,
+		},
+		{
+			name:         "empty current SHA (re-fetch failed)",
+			evaluatedSHA: "abc123def456",
+			currentSHA:   "",
+			wantSkip:     false,
+		},
+		{
+			name:         "both empty (edge case)",
+			evaluatedSHA: "",
+			currentSHA:   "",
+			wantSkip:     false,
+		},
+		{
+			name:         "only current empty",
+			evaluatedSHA: "abc123",
+			currentSHA:   "",
+			wantSkip:     false,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			got := shouldSkipStaleReview(tc.evaluatedSHA, tc.currentSHA)
+			if got != tc.wantSkip {
+				t.Errorf("shouldSkipStaleReview(%q, %q) = %v, want %v",
+					tc.evaluatedSHA, tc.currentSHA, got, tc.wantSkip)
+			}
+		})
+	}
+}

docs/DESIGN-51-personas.md

@@ -0,0 +1,334 @@
+# Design: Role-based Review Personas (Issue #51)
+
+> **Note:** This design was revised during implementation to use JSON instead of YAML
+> to maintain the repository's zero-external-dependencies convention. All persona
+> files use JSON format. See "Design Revision" section at the end for details.
+
+## Problem
+
+Current review-bot performs generic code review. Every reviewer (regardless of `reviewer-name`) uses the same base prompt and evaluates the same concerns. This leads to:
+
+1. **Redundancy** — Two reviewers (e.g., GPT + Claude twins) often flag identical issues
+2. **Gaps** — Generic reviewers miss specialized concerns (security, domain logic, architecture)
+3. **Noise** — NITs about style mixed with critical security findings
+4. **No ownership** — Findings lack clear domain attribution
+
+## Constraints
+
+- Must work with existing CLI flags and CI workflow patterns
+- Must not break backwards compatibility (existing configs still work)
+- Must integrate cleanly with the budget system (personas add to context)
+- Multiple personas running in parallel must not interfere with each other
+- Each persona must have clear scope boundaries (no duplication)
+
+## Proposed Approach
+
+### 1. Persona Definition
+
+A persona is a named review role with:
+- **Identity** — Who am I? What's my expertise?
+- **Focus** — What do I look for?
+- **Scope boundaries** — What do I explicitly NOT comment on?
+- **Severity calibration** — What counts as MAJOR/MINOR/NIT for MY domain?
+
+Personas are defined in JSON files that can live:
+1. In the pattern repos (shared across projects)
+2. In the target repo (project-specific personas)
+3. Inline via a new `--persona-file` flag (JSON format)
+
+### 2. Persona File Format
+
+```json
+# .review/personas/security.yaml
+name: security
+display_name: Security Specialist
+model_preference: opus  # optional hint for expensive analysis
+
+identity: |
+  You are a security specialist reviewing code for vulnerabilities.
+  Your expertise: OWASP Top 10, injection attacks, auth/authz, secrets management,
+  event sourcing security (replay attacks, event injection).
+
+focus:
+  - Injection attacks (SQL, command, path traversal, template)
+  - Authentication and authorization gaps
+  - Secrets exposure (hardcoded credentials, tokens in logs)
+  - Input validation (unsanitized input, unsafe deserialization)
+  - Race conditions with security implications
+  - Event sourcing attack vectors
+
+ignore:
+  - Code style and naming conventions
+  - Performance (unless security-related)
+  - Documentation
+  - General code quality
+  - Test coverage
+
+severity:
+  critical: "Remote code execution, auth bypass, data exfiltration"
+  major: "Privilege escalation, information disclosure, DoS"
+  minor: "Missing rate limiting, verbose errors"
+  nit: "Theoretical risk with low exploitability"
+
+output_format: |
+  For each finding:
+  - Severity: [CRITICAL|MAJOR|MINOR|NIT]
+  - Attack vector: How could this be exploited?
+  - Evidence: Code snippet showing the vulnerability
+  - Recommendation: Specific fix
+```
+
+### 3. New CLI Flags
+
+```
+--persona-file PATH      Path to persona JSON file (local or in repo)
+--persona NAME           Built-in persona name (security, architect, domain)
+```
+
+Either flag sets the persona. If neither is provided, behavior is unchanged (generic review).
+
+### 4. Prompt Assembly
+
+Current flow:
+```
+SystemBase → Patterns → Conventions → [LLM]
+```
+
+New flow with persona:
+```
+PersonaPrompt (from YAML) → Patterns (filtered?) → Conventions → [LLM]
+```
+
+The persona's identity/focus/ignore/severity sections become the system prompt, replacing the generic "You are an expert code reviewer" base.
+
+### 5. Built-in Personas
+
+Ship with these built-in personas (loadable via `--persona NAME`):
+
+| Name | Focus |
+|------|-------|
+| `security` | Vulnerabilities, auth, secrets |
+| `architect` | Patterns, consistency, design |
+| `domain` | Business logic (requires repo-specific config) |
+| `docs` | Documentation, API clarity |
+
+Built-in personas live in `review/personas/` as embedded Go assets or YAML shipped with the binary.
+
+### 6. CI Workflow Integration
+
+Single persona:
+```yaml
+- uses: rodin/review-bot/.gitea/actions/review@v1
+  with:
+    reviewer-name: security
+    persona: security
+    ...
+```
+
+Multiple personas (parallel jobs):
+```yaml
+jobs:
+  review:
+    strategy:
+      matrix:
+        include:
+          - name: security
+            persona: security
+          - name: architect
+            persona: architect
+    steps:
+      - uses: rodin/review-bot/.gitea/actions/review@v1
+        with:
+          reviewer-name: ${{ matrix.name }}
+          persona: ${{ matrix.persona }}
+```
+
+Custom persona from repo:
+```yaml
+- uses: rodin/review-bot/.gitea/actions/review@v1
+  with:
+    reviewer-name: trading
+    persona-file: .review/personas/trading.yaml
+```
+
+### 7. Persona + Patterns Interaction
+
+Some personas benefit from filtered patterns:
+- Security → only security-related patterns
+- Architect → all patterns (structural focus)
+- Domain → domain docs, not language patterns
+
+For v1, keep it simple: all patterns are included regardless of persona. Future enhancement could add `patterns_filter` to persona YAML.
+
+### 8. Output Format Changes
+
+Persona name appears in the review header:
+```markdown
+# Security Review
+
+## Summary
+No critical vulnerabilities found in this change.
+
+## Findings
+| # | Severity | File | Line | Finding |
+...
+
+## Recommendation
+**APPROVE** — No security-relevant issues detected.
+
+---
+*Review by security*
+<!-- review-bot:security -->
+```
+
+## State/Data Model
+
+### Persona struct
+
+```go
+// review/persona.go
+type Persona struct {
+    Name          string   `yaml:"name"`
+    DisplayName   string   `yaml:"display_name"`
+    ModelPref     string   `yaml:"model_preference,omitempty"`
+    Identity      string   `yaml:"identity"`
+    Focus         []string `yaml:"focus"`
+    Ignore        []string `yaml:"ignore"`
+    Severity      Severity `yaml:"severity"`
+    OutputFormat  string   `yaml:"output_format,omitempty"`
+}
+
+type Severity struct {
+    Critical string `yaml:"critical"`
+    Major    string `yaml:"major"`
+    Minor    string `yaml:"minor"`
+    Nit      string `yaml:"nit"`
+}
+```
+
+### Loading precedence
+
+1. `--persona-file PATH` → load from local file system
+2. `--persona NAME` → load from embedded built-ins
+3. Neither → use generic system prompt (current behavior)
+
+## Error Cases
+
+| Error | Handling |
+|-------|----------|
+| Persona file not found | Fatal exit with clear message |
+| Invalid YAML in persona file | Fatal exit with parse error |
+| Both `--persona` and `--persona-file` specified | Fatal exit: mutually exclusive |
+| Unknown built-in persona name | Fatal exit with list of valid names |
+| Empty identity in persona | Warning, fall back to generic prompt |
+
+## Edge Cases
+
+- **Empty focus list**: Valid — persona relies on identity alone
+- **Empty ignore list**: Valid — no explicit scope exclusions
+- **No severity section**: Use default MAJOR/MINOR/NIT definitions
+- **Model preference set but budget insufficient**: Ignore preference, log warning
+- **Persona file in pattern repo**: Fetch like other pattern files
+
+## Testing Strategy
+
+### Unit tests
+- `persona_test.go`: Parse valid/invalid YAML, validate required fields
+- `prompt_test.go`: Verify persona prompt assembly
+- Integration with budget: persona prompts count toward token limit
+
+### Integration tests
+- End-to-end with `--persona security` (built-in)
+- End-to-end with `--persona-file custom.yaml`
+- Backwards compatibility: no flags = generic behavior
+
+### Manual verification
+- Run security persona on a PR with obvious vulnerability
+- Verify security persona ignores style issues
+- Verify non-security persona doesn't flag security issues
+
+## Implementation Phases
+
+### Phase 1: Persona types and loading
+- [ ] `review/persona.go`: Persona struct + YAML parsing
+- [ ] `review/persona_test.go`: Unit tests
+- [ ] Embed built-in personas in binary
+- [ ] Compiles clean, tests pass
+
+### Phase 2: Prompt generation
+- [ ] `review/prompt.go`: `BuildPersonaPrompt(p Persona) string`
+- [ ] Modify `BuildSystemBase()` to accept optional persona
+- [ ] Integrate persona prompt with budget system
+- [ ] Tests for prompt assembly
+
+### Phase 3: CLI integration
+- [ ] Add `--persona` and `--persona-file` flags
+- [ ] Flag validation (mutually exclusive, valid names)
+- [ ] Load persona based on flags
+- [ ] Pass persona to prompt builder
+
+### Phase 4: Action integration
+- [ ] Add `persona` and `persona-file` inputs to action.yml
+- [ ] Update README with persona examples
+- [ ] End-to-end CI test
+
+### Phase 5: Built-in personas
+- [ ] `security.yaml` built-in
+- [ ] `architect.yaml` built-in
+- [ ] `docs.yaml` built-in
+- [ ] Document each persona's focus
+
+## Open Questions
+
+1. **Persona file location in repo**: Should we support `--persona-file .review/security.yaml` where the file is fetched from the PR's repo (like conventions)? This adds complexity but enables project-specific personas without action changes.
+
+2. **Model preference enforcement**: If persona specifies `model_preference: opus` but the action uses a different model, should we warn? Override? Ignore? Current thinking: log warning, use the specified model (user controls model via action input).
+
+3. **Severity override output**: If persona defines custom severity levels (CRITICAL), should the JSON output include them, or map back to standard MAJOR/MINOR/NIT? Current thinking: keep standard output format, use severity calibration only for prompt guidance.
+
+## Completion Checklist
+
+1. Persona struct matches YAML schema exactly?
+2. Built-in personas embedded in binary (not external files)?
+3. `--persona` and `--persona-file` are mutually exclusive?
+4. Unknown persona name produces clear error with valid options?
+5. Empty persona file fields have sensible defaults?
+6. Persona prompt integrates with budget system (token counting)?
+7. Backwards compatibility: no flags = current behavior?
+8. Review header shows persona display name?
+9. Sentinel still uses reviewer-name (not persona name)?
+10. Unit tests cover parse errors, missing fields, valid YAML?
+
+## Design Review Findings (Self-Review)
+
+### Finding 1: Severity Mapping
+The persona YAML allows `critical` severity, but the LLM output parser (`review/parser.go`) only accepts MAJOR/MINOR/NIT. 
+
+**Resolution:** Keep standard output format. Persona severity section is ONLY for calibrating the LLM's judgment (prompt guidance). Output must still use MAJOR/MINOR/NIT. Document this clearly in persona format docs.
+
+### Finding 2: Embedding Built-in Personas
+Go doesn't natively embed YAML. Must use `//go:embed` directive (Go 1.16+).
+
+**Resolution:** Create `review/personas/` directory with YAML files and use:
+```go
+//go:embed personas/*.yaml
+var embeddedPersonas embed.FS
+```
+
+### Finding 3: display_name vs reviewer-name
+Design says header shows "persona display name" but sentinel uses "reviewer-name". This is correct - they serve different purposes:
+- `display_name` → human-readable header ("Security Specialist Review")
+- `reviewer-name` → machine sentinel for cleanup (`<!-- review-bot:security -->`)
+
+When persona is used, `display_name` takes precedence for the header title, but `reviewer-name` (CLI flag) is still used for the sentinel.
+
+## Design Revision: YAML with gopkg.in/yaml.v3
+
+**Decision:** Add `gopkg.in/yaml.v3` as a dependency.
+
+YAML is preferred over JSON for persona files because:
+- Multi-line strings are cleaner (no escaping quotes in identity/focus text)
+- Comments are supported for documentation
+- More human-readable for complex persona definitions
+
+The implementation supports both YAML (`.yaml`, `.yml`) and JSON (`.json`) for backwards compatibility, with YAML as the default for built-in personas.

docs/DESIGN-57-yaml-persona.md

@@ -0,0 +1,108 @@
+# Design: YAML Support for Persona Files (#57)
+
+## Problem
+
+JSON is awkward for persona files that contain multi-line text (identity, severity descriptions). YAML supports cleaner multi-line strings and comments, improving readability and maintainability.
+
+## Constraints
+
+- Backwards compatibility: existing JSON personas must continue to work
+- Security: protect against DoS via deeply nested YAML (AIKIDO-2024-10486)
+- Consistency: use `.yaml` extension (not `.yml`)
+- Library: use `gopkg.in/yaml.v3` (approved in CONVENTIONS.md) with explicit depth limiting
+
+## Proposed Approach
+
+1. **Update `parsePersona`** to detect format from file extension
+2. **Add YAML parsing** with explicit depth limit (defense in depth)
+3. **Keep JSON as fallback** for files without `.yaml`/`.yml` extension
+4. **Convert built-in personas** to YAML format
+5. **Update embed directive** to include both formats
+
+### File Extension Detection
+
+```go
+func parsePersona(data []byte, source string) (*Persona, error) {
+    isYAML := strings.HasSuffix(source, ".yaml") || strings.HasSuffix(source, ".yml")
+    if isYAML {
+        return parseYAML(data, source)
+    }
+    return parseJSON(data, source)
+}
+```
+
+### YAML Parsing with Depth Protection
+
+```go
+func unmarshalYAMLWithDepthLimit(data []byte, out any, maxDepth int) error {
+    var node yaml.Node
+    dec := yaml.NewDecoder(bytes.NewReader(data))
+    if err := dec.Decode(&node); err != nil {
+        return err
+    }
+    if err := checkYAMLDepth(&node, 0, maxDepth); err != nil {
+        return err
+    }
+    return node.Decode(out)
+}
+
+func checkYAMLDepth(node *yaml.Node, depth, maxDepth int) error {
+    if depth > maxDepth {
+        return fmt.Errorf("YAML nesting depth exceeds maximum (%d)", maxDepth)
+    }
+    // Handle alias nodes by following the Alias pointer
+    if node.Kind == yaml.AliasNode && node.Alias != nil {
+        return checkYAMLDepth(node.Alias, depth, maxDepth)
+    }
+    for _, child := range node.Content {
+        if err := checkYAMLDepth(child, depth+1, maxDepth); err != nil {
+            return err
+        }
+    }
+    return nil
+}
+```
+
+The `gopkg.in/yaml.v3` library does not have built-in depth protection, so we implement explicit depth checking by first decoding into a `yaml.Node`, walking the tree to verify depth (including alias resolution), then decoding into the target struct.
+
+## State/Data Model
+
+No new state. Same `Persona` struct, just different parsing.
+
+## Error Cases
+
+| Error | Handling |
+|-------|----------|
+| Invalid YAML syntax | Return parse error with source file |
+| Deeply nested YAML | Library rejects (v1.16.0+ fix) |
+| Unknown extension | Fall back to JSON parsing |
+| Missing required fields | Validation rejects after parse |
+
+## Edge Cases
+
+- File with `.json` extension but YAML content → JSON parse fails, user sees error
+- File with no extension → defaults to JSON
+- Embedded persona reference like `builtin:security` → detect by embed path (`personas/X.yaml`)
+
+## Testing Strategy
+
+1. Unit tests for YAML parsing (valid, invalid, deeply nested)
+2. Unit tests for extension detection
+3. Integration test for built-in personas (now YAML)
+4. Backwards compat test: verify JSON still works for external files
+
+## Completion Checklist
+
+1. [ ] `go-yaml` dependency added at v1.16.0+
+2. [ ] Extension detection uses case-insensitive comparison
+3. [ ] YAML parse errors include source file name
+4. [ ] JSON parsing still works for `.json` files
+5. [ ] Built-in personas converted to YAML with readable multi-line strings
+6. [ ] Embed directive updated to include `*.yaml`
+7. [ ] Test for deeply nested YAML rejection
+8. [ ] All existing tests pass
+
+## Open Questions
+
+- Should we support both `.yaml` AND `.yml`? Issue says `.yaml` only for consistency, but some users expect `.yml`. **Decision:** Support both for reading, recommend `.yaml` in docs.
+- Should we add a "format" field to detect mismatched extension/content? **Decision:** No, keep it simple. Extension determines format.

go.mod

@@ -1,3 +1,5 @@
 module gitea.weiker.me/rodin/review-bot
 
 go 1.26.2
+
+require gopkg.in/yaml.v3 v3.0.1

go.sum

@@ -0,0 +1,4 @@
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
+gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=

llm/aicore.go

@@ -0,0 +1,391 @@
+package llm
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/url"
+	"strings"
+	"sync"
+	"time"
+)
+
+// AICoreOpenAIAPIVersion is the API version used for OpenAI models through AI Core.
+// Update this when SAP AI Core releases a new stable version.
+const AICoreOpenAIAPIVersion = "2024-12-01-preview"
+
+// maxErrorBodyLen limits the length of response bodies included in error messages
+// to prevent leaking potentially sensitive upstream details in logs.
+const maxErrorBodyLen = 200
+
+// AICoreConfig holds SAP AI Core authentication and connection settings.
+type AICoreConfig struct {
+	ClientID      string
+	ClientSecret  string
+	AuthURL       string
+	APIURL        string
+	ResourceGroup string
+}
+
+// AICoreClient wraps AI Core authentication and deployment discovery.
+// Thread-safe for concurrent use after construction.
+//
+// Design: The deployment cache is populated once and never invalidated. This is
+// acceptable for short-lived CI runner processes, but longer-lived deployments
+// may want to add a TTL or re-fetch on errors.
+type AICoreClient struct {
+	config AICoreConfig
+	http   *http.Client
+
+	mu          sync.RWMutex
+	token       string
+	tokenExpiry time.Time
+	deployments map[string]string // model name -> deployment URL
+}
+
+// NewAICoreClient creates a new AI Core client with the given configuration.
+// The client uses a default 5-minute timeout; use WithTimeout to customize.
+func NewAICoreClient(cfg AICoreConfig) *AICoreClient {
+	return &AICoreClient{
+		config:      cfg,
+		http:        &http.Client{Timeout: 5 * time.Minute},
+		deployments: make(map[string]string),
+	}
+}
+
+// WithTimeout sets the HTTP request timeout for AI Core calls.
+// This should be called during construction, before concurrent use.
+func (c *AICoreClient) WithTimeout(d time.Duration) *AICoreClient {
+	c.http.Timeout = d
+	return c
+}
+
+// truncateBody truncates a response body for inclusion in error messages.
+// This prevents leaking potentially sensitive upstream response details in logs.
+func truncateBody(body []byte) string {
+	if len(body) <= maxErrorBodyLen {
+		return string(body)
+	}
+	return string(body[:maxErrorBodyLen]) + "..."
+}
+
+// getToken returns a valid OAuth token, refreshing if necessary.
+func (c *AICoreClient) getToken(ctx context.Context) (string, error) {
+	c.mu.RLock()
+	if c.token != "" && time.Now().Add(5*time.Minute).Before(c.tokenExpiry) {
+		token := c.token
+		c.mu.RUnlock()
+		return token, nil
+	}
+	c.mu.RUnlock()
+
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	// Double-check after acquiring write lock
+	if c.token != "" && time.Now().Add(5*time.Minute).Before(c.tokenExpiry) {
+		return c.token, nil
+	}
+
+	token, expiry, err := c.fetchToken(ctx)
+	if err != nil {
+		return "", err
+	}
+	c.token = token
+	c.tokenExpiry = expiry
+	return token, nil
+}
+
+func (c *AICoreClient) fetchToken(ctx context.Context) (string, time.Time, error) {
+	tokenURL := strings.TrimRight(c.config.AuthURL, "/") + "/oauth/token"
+
+	data := url.Values{}
+	data.Set("grant_type", "client_credentials")
+	data.Set("client_id", c.config.ClientID)
+	data.Set("client_secret", c.config.ClientSecret)
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, tokenURL, strings.NewReader(data.Encode()))
+	if err != nil {
+		return "", time.Time{}, fmt.Errorf("create token request: %w", err)
+	}
+	req.Header.Set("Content-Type", "application/x-www-form-urlencoded")
+
+	resp, err := c.http.Do(req)
+	if err != nil {
+		return "", time.Time{}, fmt.Errorf("token request: %w", err)
+	}
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return "", time.Time{}, fmt.Errorf("read token response: %w", err)
+	}
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return "", time.Time{}, fmt.Errorf("token request failed (status %d): %s", resp.StatusCode, truncateBody(body))
+	}
+
+	var tokenResp struct {
+		AccessToken string `json:"access_token"`
+		ExpiresIn   int    `json:"expires_in"`
+	}
+	if err := json.Unmarshal(body, &tokenResp); err != nil {
+		return "", time.Time{}, fmt.Errorf("parse token response: %w", err)
+	}
+
+	if tokenResp.AccessToken == "" {
+		return "", time.Time{}, fmt.Errorf("empty access token in response")
+	}
+
+	expiry := time.Now().Add(time.Duration(tokenResp.ExpiresIn) * time.Second)
+	return tokenResp.AccessToken, expiry, nil
+}
+
+// getDeploymentURL returns the deployment URL for a model, fetching deployments if needed.
+// getDeploymentURL returns the deployment URL for a model, fetching deployments if needed.
+// Also returns a valid token for use by the caller, avoiding redundant getToken calls.
+//
+// Note: The token is fetched before acquiring the write lock to avoid holding the lock
+// during network I/O. In rare cases where multiple goroutines race and one waits a long
+// time for the write lock, the token could theoretically expire. The 5-minute refresh
+// buffer in getToken makes this extremely unlikely in practice.
+func (c *AICoreClient) getDeploymentURL(ctx context.Context, model string) (deployURL, token string, err error) {
+	c.mu.RLock()
+	if u, ok := c.deployments[model]; ok {
+		c.mu.RUnlock()
+		// Still need a token for the caller
+		token, err = c.getToken(ctx)
+		if err != nil {
+			return "", "", fmt.Errorf("get token: %w", err)
+		}
+		return u, token, nil
+	}
+	c.mu.RUnlock()
+
+	// Fetch token first (before acquiring write lock to avoid holding lock during I/O)
+	token, err = c.getToken(ctx)
+	if err != nil {
+		return "", "", fmt.Errorf("get token for deployments: %w", err)
+	}
+
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	// Double-check after acquiring write lock
+	if u, ok := c.deployments[model]; ok {
+		return u, token, nil
+	}
+
+	if err := c.fetchDeployments(ctx, token); err != nil {
+		return "", "", err
+	}
+
+	if u, ok := c.deployments[model]; ok {
+		return u, token, nil
+	}
+	return "", "", fmt.Errorf("no deployment found for model %q", model)
+}
+
+func (c *AICoreClient) fetchDeployments(ctx context.Context, token string) error {
+	deployURL := strings.TrimRight(c.config.APIURL, "/") + "/v2/lm/deployments"
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, deployURL, nil)
+	if err != nil {
+		return fmt.Errorf("create deployments request: %w", err)
+	}
+	req.Header.Set("Authorization", "Bearer "+token)
+	req.Header.Set("AI-Resource-Group", c.config.ResourceGroup)
+
+	resp, err := c.http.Do(req)
+	if err != nil {
+		return fmt.Errorf("deployments request: %w", err)
+	}
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return fmt.Errorf("read deployments response: %w", err)
+	}
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return fmt.Errorf("deployments request failed (status %d): %s", resp.StatusCode, truncateBody(body))
+	}
+
+	var deployResp struct {
+		Resources []struct {
+			DeploymentURL string `json:"deploymentUrl"`
+			Status        string `json:"status"`
+			Details       struct {
+				Resources struct {
+					BackendDetails struct {
+						Model struct {
+							Name string `json:"name"`
+						} `json:"model"`
+					} `json:"backend_details"`
+				} `json:"resources"`
+			} `json:"details"`
+		} `json:"resources"`
+	}
+	if err := json.Unmarshal(body, &deployResp); err != nil {
+		return fmt.Errorf("parse deployments response: %w", err)
+	}
+
+	for _, r := range deployResp.Resources {
+		if r.Status != "RUNNING" {
+			continue
+		}
+		modelName := r.Details.Resources.BackendDetails.Model.Name
+		if modelName == "" {
+			continue
+		}
+		c.deployments[modelName] = r.DeploymentURL
+	}
+
+	return nil
+}
+
+// CompleteAnthropic sends a request to an Anthropic model via AI Core.
+func (c *AICoreClient) CompleteAnthropic(ctx context.Context, model string, messages []Message, maxTokens int, temperature float64) (string, error) {
+	deployURL, token, err := c.getDeploymentURL(ctx, model)
+	if err != nil {
+		return "", err
+	}
+
+	// Extract system message
+	var system string
+	var userMessages []anthropicMsg
+	for _, m := range messages {
+		if m.Role == "system" {
+			system = m.Content
+		} else {
+			userMessages = append(userMessages, anthropicMsg{
+				Role:    m.Role,
+				Content: m.Content,
+			})
+		}
+	}
+
+	reqBody := anthropicRequest{
+		AnthropicVersion: "bedrock-2023-05-31", // SAP AI Core uses Bedrock format
+		// Model omitted - AI Core deployment already specifies model
+		MaxTokens: maxTokens,
+		System:    system,
+		Messages:  userMessages,
+	}
+	if temperature > 0 {
+		reqBody.Temperature = temperature
+	}
+
+	data, err := json.Marshal(reqBody)
+	if err != nil {
+		return "", fmt.Errorf("marshal request: %w", err)
+	}
+
+	// AI Core uses /invoke for Anthropic models
+	invokeURL := strings.TrimRight(deployURL, "/") + "/invoke"
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, invokeURL, bytes.NewReader(data))
+	if err != nil {
+		return "", fmt.Errorf("create request: %w", err)
+	}
+	req.Header.Set("Authorization", "Bearer "+token)
+	req.Header.Set("AI-Resource-Group", c.config.ResourceGroup)
+	req.Header.Set("Content-Type", "application/json")
+
+	resp, err := c.http.Do(req)
+	if err != nil {
+		return "", fmt.Errorf("AI Core request: %w", err)
+	}
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return "", fmt.Errorf("read response: %w", err)
+	}
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return "", fmt.Errorf("AI Core API error (status %d): %s", resp.StatusCode, truncateBody(body))
+	}
+
+	var anthropicResp anthropicResponse
+	if err := json.Unmarshal(body, &anthropicResp); err != nil {
+		return "", fmt.Errorf("parse response: %w", err)
+	}
+
+	if len(anthropicResp.Content) == 0 {
+		return "", fmt.Errorf("no content in response")
+	}
+
+	var sb strings.Builder
+	for _, block := range anthropicResp.Content {
+		if block.Type == "text" {
+			sb.WriteString(block.Text)
+		}
+	}
+	result := sb.String()
+	if result == "" {
+		return "", fmt.Errorf("no text content in response")
+	}
+	return result, nil
+}
+
+// CompleteOpenAI sends a request to an OpenAI model via AI Core.
+func (c *AICoreClient) CompleteOpenAI(ctx context.Context, model string, messages []Message, temperature float64) (string, error) {
+	deployURL, token, err := c.getDeploymentURL(ctx, model)
+	if err != nil {
+		return "", err
+	}
+
+	reqBody := ChatRequest{
+		Model:       model,
+		Temperature: temperature,
+		Messages:    messages,
+	}
+
+	data, err := json.Marshal(reqBody)
+	if err != nil {
+		return "", fmt.Errorf("marshal request: %w", err)
+	}
+
+	// AI Core uses /chat/completions?api-version=<version> for OpenAI models
+	chatURL := strings.TrimRight(deployURL, "/") + "/chat/completions?api-version=" + AICoreOpenAIAPIVersion
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, chatURL, bytes.NewReader(data))
+	if err != nil {
+		return "", fmt.Errorf("create request: %w", err)
+	}
+	req.Header.Set("Authorization", "Bearer "+token)
+	req.Header.Set("AI-Resource-Group", c.config.ResourceGroup)
+	req.Header.Set("Content-Type", "application/json")
+
+	resp, err := c.http.Do(req)
+	if err != nil {
+		return "", fmt.Errorf("AI Core request: %w", err)
+	}
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return "", fmt.Errorf("read response: %w", err)
+	}
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return "", fmt.Errorf("AI Core API error (status %d): %s", resp.StatusCode, truncateBody(body))
+	}
+
+	var openaiResp ChatResponse
+	if err := json.Unmarshal(body, &openaiResp); err != nil {
+		return "", fmt.Errorf("parse response: %w", err)
+	}
+
+	if len(openaiResp.Choices) == 0 {
+		return "", fmt.Errorf("no choices in response")
+	}
+	return openaiResp.Choices[0].Message.Content, nil
+}
+
+// IsAnthropicModel returns true if the model name indicates an Anthropic model.
+// SAP AI Core uses "anthropic--" prefix for Anthropic models (e.g., "anthropic--claude-3-5-sonnet").
+func IsAnthropicModel(model string) bool {
+	return strings.HasPrefix(model, "anthropic--")
+}

llm/aicore_test.go

@@ -0,0 +1,535 @@
+package llm
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+func TestAICoreClient_TokenFetch(t *testing.T) {
+	tokenCalls := int32(0)
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/oauth/token" {
+			atomic.AddInt32(&tokenCalls, 1)
+			if r.Method != http.MethodPost {
+				t.Errorf("expected POST for token, got %s", r.Method)
+			}
+			if r.Header.Get("Content-Type") != "application/x-www-form-urlencoded" {
+				t.Errorf("expected form content type")
+			}
+			w.Header().Set("Content-Type", "application/json")
+			json.NewEncoder(w).Encode(map[string]interface{}{
+				"access_token": "test-token-123",
+				"expires_in":   3600,
+			})
+			return
+		}
+		t.Errorf("unexpected path: %s", r.URL.Path)
+	}))
+	defer server.Close()
+
+	client := NewAICoreClient(AICoreConfig{
+		ClientID:      "test-id",
+		ClientSecret:  "test-secret",
+		AuthURL:       server.URL,
+		APIURL:        server.URL,
+		ResourceGroup: "default",
+	})
+
+	token, err := client.getToken(context.Background())
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if token != "test-token-123" {
+		t.Errorf("expected token 'test-token-123', got %q", token)
+	}
+
+	// Second call should use cached token
+	token2, err := client.getToken(context.Background())
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if token2 != "test-token-123" {
+		t.Errorf("expected cached token")
+	}
+	if atomic.LoadInt32(&tokenCalls) != 1 {
+		t.Errorf("expected 1 token call (cached), got %d", tokenCalls)
+	}
+}
+
+func TestAICoreClient_DeploymentFetch(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/oauth/token" {
+			w.Header().Set("Content-Type", "application/json")
+			json.NewEncoder(w).Encode(map[string]interface{}{
+				"access_token": "test-token",
+				"expires_in":   3600,
+			})
+			return
+		}
+		if r.URL.Path == "/v2/lm/deployments" {
+			if r.Header.Get("Authorization") != "Bearer test-token" {
+				t.Errorf("expected Bearer auth")
+			}
+			if r.Header.Get("AI-Resource-Group") != "default" {
+				t.Errorf("expected resource group header")
+			}
+			w.Header().Set("Content-Type", "application/json")
+			json.NewEncoder(w).Encode(map[string]interface{}{
+				"resources": []map[string]interface{}{
+					{
+						"id":            "deploy-123",
+						"deploymentUrl": "https://example.com/v2/inference/deployments/deploy-123",
+						"status":        "RUNNING",
+						"details": map[string]interface{}{
+							"resources": map[string]interface{}{
+								"backend_details": map[string]interface{}{
+									"model": map[string]interface{}{
+										"name": "anthropic--claude-4.6-sonnet",
+									},
+								},
+							},
+						},
+					},
+					{
+						"id":            "deploy-456",
+						"deploymentUrl": "https://example.com/v2/inference/deployments/deploy-456",
+						"status":        "STOPPED",
+						"details": map[string]interface{}{
+							"resources": map[string]interface{}{
+								"backend_details": map[string]interface{}{
+									"model": map[string]interface{}{
+										"name": "gpt-5",
+									},
+								},
+							},
+						},
+					},
+					{
+						"id":            "deploy-789",
+						"deploymentUrl": "https://example.com/v2/inference/deployments/deploy-789",
+						"status":        "RUNNING",
+						"details": map[string]interface{}{
+							"resources": map[string]interface{}{
+								"backend_details": map[string]interface{}{
+									"model": map[string]interface{}{
+										"name": "gpt-5",
+									},
+								},
+							},
+						},
+					},
+				},
+			})
+			return
+		}
+		t.Errorf("unexpected path: %s", r.URL.Path)
+	}))
+	defer server.Close()
+
+	client := NewAICoreClient(AICoreConfig{
+		ClientID:      "test-id",
+		ClientSecret:  "test-secret",
+		AuthURL:       server.URL,
+		APIURL:        server.URL,
+		ResourceGroup: "default",
+	})
+
+	// Should find running deployment
+	url, _, err := client.getDeploymentURL(context.Background(), "anthropic--claude-4.6-sonnet")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if url != "https://example.com/v2/inference/deployments/deploy-123" {
+		t.Errorf("unexpected URL: %s", url)
+	}
+
+	// Should find running gpt-5, not stopped one
+	url, _, err = client.getDeploymentURL(context.Background(), "gpt-5")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if url != "https://example.com/v2/inference/deployments/deploy-789" {
+		t.Errorf("unexpected URL: %s", url)
+	}
+
+	// Should error on unknown model
+	_, _, err = client.getDeploymentURL(context.Background(), "unknown-model")
+	if err == nil {
+		t.Error("expected error for unknown model")
+	}
+}
+
+func TestAICoreClient_CompleteAnthropic(t *testing.T) {
+	// baseURL is set after server creation; captured by closure in handlers
+	var baseURL string
+	mux := http.NewServeMux()
+	mux.HandleFunc("/oauth/token", func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(map[string]interface{}{
+			"access_token": "test-token",
+			"expires_in":   3600,
+		})
+	})
+	mux.HandleFunc("/v2/lm/deployments", func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(map[string]interface{}{
+			"resources": []map[string]interface{}{
+				{
+					"id":            "deploy-anthropic",
+					"deploymentUrl": baseURL + "/deployments/anthropic",
+					"status":        "RUNNING",
+					"details": map[string]interface{}{
+						"resources": map[string]interface{}{
+							"backend_details": map[string]interface{}{
+								"model": map[string]interface{}{
+									"name": "anthropic--claude-4.6-sonnet",
+								},
+							},
+						},
+					},
+				},
+			},
+		})
+	})
+	mux.HandleFunc("/deployments/anthropic/invoke", func(w http.ResponseWriter, r *http.Request) {
+		if r.Header.Get("Authorization") != "Bearer test-token" {
+			t.Errorf("expected Bearer auth on invoke")
+		}
+		var req anthropicRequest
+		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+			t.Fatalf("decode request: %v", err)
+		}
+		if req.AnthropicVersion != "bedrock-2023-05-31" {
+			t.Errorf("expected bedrock anthropic_version in request")
+		}
+		if req.System != "You are helpful" {
+			t.Errorf("expected system prompt: %q", req.System)
+		}
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(map[string]interface{}{
+			"content": []map[string]interface{}{
+				{"type": "text", "text": "Hello from AI Core!"},
+			},
+		})
+	})
+
+	server := httptest.NewServer(mux)
+	baseURL = server.URL
+	defer server.Close()
+
+	client := NewAICoreClient(AICoreConfig{
+		ClientID:      "test-id",
+		ClientSecret:  "test-secret",
+		AuthURL:       server.URL,
+		APIURL:        server.URL,
+		ResourceGroup: "default",
+	})
+
+	result, err := client.CompleteAnthropic(context.Background(), "anthropic--claude-4.6-sonnet", []Message{
+		{Role: "system", Content: "You are helpful"},
+		{Role: "user", Content: "Hello"},
+	}, 8192, 0)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if result != "Hello from AI Core!" {
+		t.Errorf("expected 'Hello from AI Core!', got %q", result)
+	}
+}
+
+func TestAICoreClient_CompleteOpenAI(t *testing.T) {
+	var baseURL string
+	mux := http.NewServeMux()
+	mux.HandleFunc("/oauth/token", func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(map[string]interface{}{
+			"access_token": "test-token",
+			"expires_in":   3600,
+		})
+	})
+	mux.HandleFunc("/v2/lm/deployments", func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(map[string]interface{}{
+			"resources": []map[string]interface{}{
+				{
+					"id":            "deploy-openai",
+					"deploymentUrl": baseURL + "/deployments/openai",
+					"status":        "RUNNING",
+					"details": map[string]interface{}{
+						"resources": map[string]interface{}{
+							"backend_details": map[string]interface{}{
+								"model": map[string]interface{}{
+									"name": "gpt-5",
+								},
+							},
+						},
+					},
+				},
+			},
+		})
+	})
+	mux.HandleFunc("/deployments/openai/chat/completions", func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Query().Get("api-version") != AICoreOpenAIAPIVersion {
+			t.Errorf("expected api-version %s, got %s", AICoreOpenAIAPIVersion, r.URL.Query().Get("api-version"))
+		}
+		var req ChatRequest
+		if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+			t.Fatalf("decode request: %v", err)
+		}
+		if req.Model != "gpt-5" {
+			t.Errorf("expected model gpt-5, got %s", req.Model)
+		}
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(ChatResponse{
+			Choices: []struct {
+				Message struct {
+					Content string `json:"content"`
+				} `json:"message"`
+			}{
+				{Message: struct {
+					Content string `json:"content"`
+				}{Content: "Hello from GPT-5!"}},
+			},
+		})
+	})
+
+	server := httptest.NewServer(mux)
+	baseURL = server.URL
+	defer server.Close()
+
+	client := NewAICoreClient(AICoreConfig{
+		ClientID:      "test-id",
+		ClientSecret:  "test-secret",
+		AuthURL:       server.URL,
+		APIURL:        server.URL,
+		ResourceGroup: "default",
+	})
+
+	result, err := client.CompleteOpenAI(context.Background(), "gpt-5", []Message{
+		{Role: "user", Content: "Hello"},
+	}, 0)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if result != "Hello from GPT-5!" {
+		t.Errorf("expected 'Hello from GPT-5!', got %q", result)
+	}
+}
+
+func TestIsAnthropicModel(t *testing.T) {
+	tests := []struct {
+		model    string
+		expected bool
+	}{
+		// SAP AI Core uses "anthropic--" prefix for Anthropic models
+		{"anthropic--claude-4.6-sonnet", true},
+		{"anthropic--claude-4.6-opus", true},
+		{"anthropic--claude-3-5-sonnet", true},
+		// Non-prefixed model names are not detected as Anthropic
+		// (SAP AI Core always uses the prefix for Anthropic models)
+		{"claude-sonnet-4", false},
+		{"gpt-5", false},
+		{"gpt-4.1", false},
+		{"llama-3", false},
+		{"my-claude-model", false}, // Avoid false positives on "claude" substring
+	}
+
+	for _, tt := range tests {
+		got := IsAnthropicModel(tt.model)
+		if got != tt.expected {
+			t.Errorf("IsAnthropicModel(%q) = %v, want %v", tt.model, got, tt.expected)
+		}
+	}
+}
+
+func TestAICoreClient_TokenExpiry(t *testing.T) {
+	tokenCalls := int32(0)
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/oauth/token" {
+			call := atomic.AddInt32(&tokenCalls, 1)
+			w.Header().Set("Content-Type", "application/json")
+			json.NewEncoder(w).Encode(map[string]interface{}{
+				"access_token": fmt.Sprintf("token-%d", call),
+				"expires_in":   1, // 1 second expiry
+			})
+			return
+		}
+	}))
+	defer server.Close()
+
+	client := NewAICoreClient(AICoreConfig{
+		ClientID:      "test-id",
+		ClientSecret:  "test-secret",
+		AuthURL:       server.URL,
+		APIURL:        server.URL,
+		ResourceGroup: "default",
+	})
+
+	// First call
+	token1, err := client.getToken(context.Background())
+	if err != nil {
+		t.Fatalf("first getToken: %v", err)
+	}
+
+	// Force token expiry by manipulating expiry time
+	client.mu.Lock()
+	client.tokenExpiry = time.Now().Add(-time.Hour)
+	client.mu.Unlock()
+
+	// Should fetch new token
+	token2, err := client.getToken(context.Background())
+	if err != nil {
+		t.Fatalf("second getToken: %v", err)
+	}
+
+	if token1 == token2 {
+		t.Error("expected different tokens after expiry")
+	}
+	if atomic.LoadInt32(&tokenCalls) != 2 {
+		t.Errorf("expected 2 token calls, got %d", tokenCalls)
+	}
+}
+
+func TestAICoreClient_WithTimeout(t *testing.T) {
+	client := NewAICoreClient(AICoreConfig{
+		ClientID:      "test-id",
+		ClientSecret:  "test-secret",
+		AuthURL:       "https://auth.example.com",
+		APIURL:        "https://api.example.com",
+		ResourceGroup: "default",
+	})
+
+	// Default timeout is 5 minutes
+	if client.http.Timeout != 5*time.Minute {
+		t.Errorf("expected default timeout 5m, got %v", client.http.Timeout)
+	}
+
+	// WithTimeout should update the timeout
+	client.WithTimeout(10 * time.Minute)
+	if client.http.Timeout != 10*time.Minute {
+		t.Errorf("expected timeout 10m, got %v", client.http.Timeout)
+	}
+}
+
+func TestClient_WithAICore(t *testing.T) {
+	client := NewClient("http://example.com", "key", "model")
+	if client.provider != ProviderOpenAI {
+		t.Errorf("expected default provider openai, got %s", client.provider)
+	}
+
+	client.WithAICore(AICoreConfig{
+		ClientID:      "id",
+		ClientSecret:  "secret",
+		AuthURL:       "https://auth.example.com",
+		APIURL:        "https://api.example.com",
+		ResourceGroup: "default",
+	})
+
+	if client.provider != ProviderAICore {
+		t.Errorf("expected provider aicore, got %s", client.provider)
+	}
+	if client.aicore == nil {
+		t.Error("expected aicore client to be set")
+	}
+}
+
+func TestClient_WithTimeout_PropagatestoAICore(t *testing.T) {
+	client := NewClient("http://example.com", "key", "model").
+		WithAICore(AICoreConfig{
+			ClientID:      "id",
+			ClientSecret:  "secret",
+			AuthURL:       "https://auth.example.com",
+			APIURL:        "https://api.example.com",
+			ResourceGroup: "default",
+		})
+
+	// Default should be 5 minutes (inherited from parent client)
+	if client.aicore.http.Timeout != 5*time.Minute {
+		t.Errorf("expected aicore default timeout 5m, got %v", client.aicore.http.Timeout)
+	}
+
+	// WithTimeout should propagate to AI Core client
+	client.WithTimeout(15 * time.Minute)
+	if client.http.Timeout != 15*time.Minute {
+		t.Errorf("expected parent timeout 15m, got %v", client.http.Timeout)
+	}
+	if client.aicore.http.Timeout != 15*time.Minute {
+		t.Errorf("expected aicore timeout 15m, got %v", client.aicore.http.Timeout)
+	}
+}
+
+func TestClient_CompleteAICore(t *testing.T) {
+	var baseURL string
+	mux := http.NewServeMux()
+	mux.HandleFunc("/oauth/token", func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(map[string]interface{}{
+			"access_token": "test-token",
+			"expires_in":   3600,
+		})
+	})
+	mux.HandleFunc("/v2/lm/deployments", func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(map[string]interface{}{
+			"resources": []map[string]interface{}{
+				{
+					"id":            "deploy-test",
+					"deploymentUrl": baseURL + "/deployments/test",
+					"status":        "RUNNING",
+					"details": map[string]interface{}{
+						"resources": map[string]interface{}{
+							"backend_details": map[string]interface{}{
+								"model": map[string]interface{}{
+									"name": "gpt-5",
+								},
+							},
+						},
+					},
+				},
+			},
+		})
+	})
+	mux.HandleFunc("/deployments/test/chat/completions", func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(ChatResponse{
+			Choices: []struct {
+				Message struct {
+					Content string `json:"content"`
+				} `json:"message"`
+			}{
+				{Message: struct {
+					Content string `json:"content"`
+				}{Content: "AI Core via Client works!"}},
+			},
+		})
+	})
+
+	server := httptest.NewServer(mux)
+	baseURL = server.URL
+	defer server.Close()
+
+	client := NewClient("", "", "gpt-5").WithAICore(AICoreConfig{
+		ClientID:      "test-id",
+		ClientSecret:  "test-secret",
+		AuthURL:       server.URL,
+		APIURL:        server.URL,
+		ResourceGroup: "default",
+	})
+
+	result, err := client.Complete(context.Background(), []Message{
+		{Role: "user", Content: "Hello"},
+	})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if !strings.Contains(result, "AI Core via Client works!") {
+		t.Errorf("unexpected result: %s", result)
+	}
+}

llm/client.go

@@ -1,6 +1,6 @@
 // Package llm provides clients for LLM chat completion APIs.
 //
-// Supports OpenAI-compatible (default) and Anthropic Messages API providers.
+// Supports OpenAI-compatible (default), Anthropic Messages API, and SAP AI Core providers.
 package llm
 
 import (
@@ -22,6 +22,8 @@ const (
 	ProviderOpenAI Provider = "openai"
 	// ProviderAnthropic uses the Anthropic Messages API endpoint.
 	ProviderAnthropic Provider = "anthropic"
+	// ProviderAICore uses SAP AI Core with OAuth authentication.
+	ProviderAICore Provider = "aicore"
 )
 
 // Client calls an LLM chat completion API.
@@ -35,6 +37,7 @@ type Client struct {
 	temperature float64
 	provider    Provider
 	http        *http.Client
+	aicore      *AICoreClient // Only set when provider is aicore
 }
 
 // NewClient creates a new LLM client. Default provider is OpenAI-compatible.
@@ -49,8 +52,12 @@ func NewClient(baseURL, apiKey, model string) *Client {
 }
 
 // WithTimeout sets the HTTP request timeout for LLM calls (default 5 minutes).
+// When using AI Core, this also sets the timeout on the AI Core client.
 func (c *Client) WithTimeout(d time.Duration) *Client {
 	c.http.Timeout = d
+	if c.aicore != nil {
+		c.aicore.WithTimeout(d)
+	}
 	return c
 }
 
@@ -60,12 +67,21 @@ func (c *Client) WithTemperature(t float64) *Client {
 	return c
 }
 
-// WithProvider sets the API provider format (openai or anthropic).
+// WithProvider sets the API provider format (openai, anthropic, or aicore).
 func (c *Client) WithProvider(p Provider) *Client {
 	c.provider = p
 	return c
 }
 
+// WithAICore configures the client to use SAP AI Core for authentication.
+// This sets the provider to aicore automatically.
+// The AI Core client inherits the current HTTP timeout from this client.
+func (c *Client) WithAICore(cfg AICoreConfig) *Client {
+	c.provider = ProviderAICore
+	c.aicore = NewAICoreClient(cfg).WithTimeout(c.http.Timeout)
+	return c
+}
+
 // Message represents a chat message.
 type Message struct {
 	Role    string `json:"role"`
@@ -75,12 +91,66 @@ type Message struct {
 // Complete sends a chat completion request and returns the assistant's response content.
 // The first message with role "system" is treated as the system prompt.
 func (c *Client) Complete(ctx context.Context, messages []Message) (string, error) {
-	switch c.provider {
-	case ProviderAnthropic:
-		return c.completeAnthropic(ctx, messages)
-	default:
-		return c.completeOpenAI(ctx, messages)
+	var result string
+	var err error
+
+	for attempt := 0; attempt < 2; attempt++ {
+		switch c.provider {
+		case ProviderAnthropic:
+			result, err = c.completeAnthropic(ctx, messages)
+		case ProviderAICore:
+			result, err = c.completeAICore(ctx, messages)
+		default:
+			result, err = c.completeOpenAI(ctx, messages)
+		}
+
+		if err == nil {
+			return result, nil
+		}
+
+		// Only retry on response body read errors (transient network issues).
+		// Do not retry on context cancellation, status errors, or parse errors
+		// that indicate a structural API problem.
+		if !isRetryableError(err) {
+			return "", err
+		}
+
+		if attempt == 0 && ctx.Err() == nil {
+			// Brief pause before retry to allow transient issues to resolve.
+			time.Sleep(500 * time.Millisecond)
+		}
 	}
+
+	return "", err
+}
+
+// completeAICore routes to AI Core using the appropriate endpoint based on model type.
+func (c *Client) completeAICore(ctx context.Context, messages []Message) (string, error) {
+	if c.aicore == nil {
+		return "", fmt.Errorf("AI Core client not configured")
+	}
+
+	if IsAnthropicModel(c.model) {
+		return c.aicore.CompleteAnthropic(ctx, c.model, messages, 8192, c.temperature)
+	}
+	return c.aicore.CompleteOpenAI(ctx, c.model, messages, c.temperature)
+}
+
+// isRetryableError returns true for transient errors worth retrying.
+func isRetryableError(err error) bool {
+	if err == nil {
+		return false
+	}
+	s := err.Error()
+	// Body read failures (connection reset, truncation)
+	if strings.Contains(s, "read response") {
+		return true
+	}
+	// Unexpected body length (our content-length validation)
+	if strings.Contains(s, "body length mismatch") {
+		return true
+	}
+	return false
 }
 
 // --- OpenAI-compatible implementation ---
@@ -136,11 +206,12 @@ func (c *Client) completeOpenAI(ctx context.Context, messages []Message) (string
 // --- Anthropic Messages API implementation ---
 
 type anthropicRequest struct {
-	Model       string            `json:"model"`
-	MaxTokens   int               `json:"max_tokens"`
-	System      string            `json:"system,omitempty"`
-	Messages    []anthropicMsg    `json:"messages"`
-	Temperature float64           `json:"temperature,omitempty"`
+	AnthropicVersion string         `json:"anthropic_version,omitempty"`
+	Model       string         `json:"model,omitempty"`
+	MaxTokens   int            `json:"max_tokens"`
+	System      string         `json:"system,omitempty"`
+	Messages    []anthropicMsg `json:"messages"`
+	Temperature float64        `json:"temperature,omitempty"`
 }
 
 type anthropicMsg struct {
@@ -231,6 +302,12 @@ func (c *Client) doRequest(req *http.Request, parse func([]byte) (string, error)
 		return "", fmt.Errorf("read response: %w", err)
 	}
 
+	// Validate body length against Content-Length header when present.
+	// A mismatch indicates the response was truncated in transit.
+	if cl := resp.ContentLength; cl > 0 && int64(len(body)) < cl {
+		return "", fmt.Errorf("body length mismatch: Content-Length=%d, received=%d", cl, len(body))
+	}
+
 	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
 		return "", fmt.Errorf("LLM API error (status %d): %s", resp.StatusCode, string(body))
 	}

llm/client_test.go

@@ -3,6 +3,7 @@ package llm
 import (
 	"context"
 	"encoding/json"
+	"fmt"
 	"net/http"
 	"net/http/httptest"
 	"testing"
@@ -295,3 +296,131 @@ func TestWithProvider(t *testing.T) {
 		t.Errorf("expected provider anthropic, got %s", client.provider)
 	}
 }
+
+func TestComplete_RetryOnBodyReadError(t *testing.T) {
+	attempts := 0
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		attempts++
+		if attempts == 1 {
+			// First attempt: send headers then close connection abruptly
+			// Simulate by writing partial response and flushing with wrong Content-Length
+			w.Header().Set("Content-Length", "1000")
+			w.WriteHeader(http.StatusOK)
+			w.Write([]byte(`{"choices":[{"message":{"con`))
+			// The test HTTP server will close the connection after handler returns,
+			// but Content-Length mismatch means client gets fewer bytes than expected
+			return
+		}
+		// Second attempt: succeed
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(ChatResponse{
+			Choices: []struct {
+				Message struct {
+					Content string `json:"content"`
+				} `json:"message"`
+			}{{Message: struct {
+				Content string `json:"content"`
+			}{Content: "success"}}},
+		})
+	}))
+	defer server.Close()
+
+	client := NewClient(server.URL, "key", "model")
+	got, err := client.Complete(context.Background(), []Message{{Role: "user", Content: "Hi"}})
+	if err != nil {
+		t.Fatalf("expected retry to succeed, got error: %v", err)
+	}
+	if got != "success" {
+		t.Errorf("expected %q, got %q", "success", got)
+	}
+	if attempts != 2 {
+		t.Errorf("expected 2 attempts, got %d", attempts)
+	}
+}
+
+func TestComplete_ContentLengthMismatch(t *testing.T) {
+	attempts := 0
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		attempts++
+		if attempts == 1 {
+			// Claim Content-Length is larger than actual body
+			w.Header().Set("Content-Length", "500")
+			w.Header().Set("Content-Type", "application/json")
+			w.WriteHeader(http.StatusOK)
+			// Write less than 500 bytes
+			w.Write([]byte(`{"choices":[{"message":{"content":"partial"}}]}`))
+			return
+		}
+		// Second attempt succeeds
+		w.Header().Set("Content-Type", "application/json")
+		json.NewEncoder(w).Encode(ChatResponse{
+			Choices: []struct {
+				Message struct {
+					Content string `json:"content"`
+				} `json:"message"`
+			}{{Message: struct {
+				Content string `json:"content"`
+			}{Content: "complete"}}},
+		})
+	}))
+	defer server.Close()
+
+	client := NewClient(server.URL, "key", "model")
+	got, err := client.Complete(context.Background(), []Message{{Role: "user", Content: "Hi"}})
+	if err != nil {
+		t.Fatalf("expected retry to succeed on content-length mismatch, got: %v", err)
+	}
+	if got != "complete" {
+		t.Errorf("expected %q, got %q", "complete", got)
+	}
+}
+
+func TestComplete_NoRetryOnAPIError(t *testing.T) {
+	attempts := 0
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		attempts++
+		w.WriteHeader(http.StatusBadRequest)
+		w.Write([]byte(`{"error":"bad request"}`))
+	}))
+	defer server.Close()
+
+	client := NewClient(server.URL, "key", "model")
+	_, err := client.Complete(context.Background(), []Message{{Role: "user", Content: "Hi"}})
+	if err == nil {
+		t.Fatal("expected error for 400, got nil")
+	}
+	if attempts != 1 {
+		t.Errorf("should not retry on API errors, got %d attempts", attempts)
+	}
+}
+
+func TestIsRetryableError(t *testing.T) {
+	tests := []struct {
+		name     string
+		err      string
+		expected bool
+	}{
+		{"nil formatted", "", false},
+		{"read response error", "read response: unexpected EOF", true},
+		{"body length mismatch", "body length mismatch: Content-Length=1000, received=500", true},
+		{"API error", "LLM API error (status 400): bad request", false},
+		{"parse error", "parse response: unexpected end of JSON input", false},
+		{"request error", "LLM request: connection refused", false},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			if tt.err == "" {
+				if isRetryableError(nil) {
+					t.Error("nil error should not be retryable")
+				}
+				return
+			}
+			err := fmt.Errorf("%s", tt.err)
+			got := isRetryableError(err)
+			if got != tt.expected {
+				t.Errorf("isRetryableError(%q) = %v, want %v", tt.err, got, tt.expected)
+			}
+		})
+	}
+}

review/formatter.go

@@ -2,15 +2,63 @@ package review
 
 import (
 	"fmt"
+	"regexp"
 	"strings"
 )
 
 // FormatMarkdown formats a ReviewResult into the markdown body for a Gitea review.
 func FormatMarkdown(result *ReviewResult, reviewerName string) string {
+	return FormatMarkdownWithDisplay(result, reviewerName, reviewerName)
+}
+
+// GiteaEvent converts the verdict to the Gitea API event string.
+func GiteaEvent(verdict string) string {
+	switch verdict {
+	case "APPROVE":
+		return "APPROVED"
+	case "REQUEST_CHANGES":
+		return "REQUEST_CHANGES"
+	default:
+		return "COMMENT"
+	}
+}
+
+// markdownSpecialChars matches characters that have special meaning in Markdown.
+// We escape these to prevent untrusted input from breaking formatting.
+// Uses a quoted string since raw strings can't contain backticks.
+var markdownSpecialChars = regexp.MustCompile("([\\\\*_`\\[\\]()#<>|~])")
+
+// sanitizeMarkdownText escapes special Markdown characters in untrusted text.
+// This prevents markdown injection attacks where a malicious display name could
+// break formatting, inject links, or create unexpected rendering.
+func sanitizeMarkdownText(s string) string {
+	// First, remove any control characters and null bytes
+	cleaned := strings.Map(func(r rune) rune {
+		if r < 32 && r != '\t' && r != '\n' {
+			return -1 // drop the character
+		}
+		return r
+	}, s)
+	// Escape special Markdown characters by prepending backslash
+	return markdownSpecialChars.ReplaceAllString(cleaned, `\$1`)
+}
+
+// FormatMarkdownWithDisplay formats a ReviewResult with separate display name and sentinel name.
+// displayName is sanitized to prevent Markdown injection from untrusted remote persona metadata.
+// sentinelName is used for the cleanup sentinel comment (machine-readable, not rendered).
+// If displayName is empty, sentinelName is used for both.
+func FormatMarkdownWithDisplay(result *ReviewResult, displayName, sentinelName string) string {
 	var sb strings.Builder
 
-	if reviewerName != "" {
-		title := strings.ToUpper(reviewerName[:1]) + reviewerName[1:]
+	// Use display name for header, or fall back to sentinel name
+	headerName := displayName
+	if headerName == "" {
+		headerName = sentinelName
+	}
+
+	if headerName != "" {
+		// Sanitize the header name to prevent Markdown injection
+		title := CapitalizeFirst(sanitizeMarkdownText(headerName))
 		sb.WriteString(fmt.Sprintf("# %s Review\n\n", title))
 	}
 
@@ -33,23 +81,12 @@ func FormatMarkdown(result *ReviewResult, reviewerName string) string {
 	sb.WriteString("## Recommendation\n\n")
 	sb.WriteString(fmt.Sprintf("**%s** — %s\n", result.Verdict, result.Recommendation))
 
-	if reviewerName != "" {
-		sb.WriteString(fmt.Sprintf("\n---\n*Review by %s*\n", reviewerName))
+	if sentinelName != "" {
+		// Sanitize headerName for the footer as well
+		sb.WriteString(fmt.Sprintf("\n---\n*Review by %s*\n", sanitizeMarkdownText(headerName)))
 		// Hidden sentinel for identifying this bot's reviews during cleanup
-		sb.WriteString(fmt.Sprintf("\n<!-- review-bot:%s -->\n", reviewerName))
+		sb.WriteString(fmt.Sprintf("\n<!-- review-bot:%s -->\n", sentinelName))
 	}
 
 	return sb.String()
 }
-
-// GiteaEvent converts the verdict to the Gitea API event string.
-func GiteaEvent(verdict string) string {
-	switch verdict {
-	case "APPROVE":
-		return "APPROVED"
-	case "REQUEST_CHANGES":
-		return "REQUEST_CHANGES"
-	default:
-		return "COMMENT"
-	}
-}

review/formatter_test.go

@@ -159,3 +159,126 @@ func TestFormatMarkdown_RoleTitle(t *testing.T) {
 		t.Error("should not contain role title header when reviewer name is empty")
 	}
 }
+
+func TestFormatMarkdownWithDisplay(t *testing.T) {
+	result := &ReviewResult{
+		Verdict:        "APPROVE",
+		Summary:        "Test summary",
+		Findings:       nil,
+		Recommendation: "Test recommendation",
+	}
+
+	t.Run("with display name", func(t *testing.T) {
+		body := FormatMarkdownWithDisplay(result, "Security Specialist", "security")
+
+		// Header should use display name
+		if !strings.Contains(body, "# Security Specialist Review") {
+			t.Error("header should use display name")
+		}
+
+		// Sentinel should use sentinel name
+		if !strings.Contains(body, "<!-- review-bot:security -->") {
+			t.Error("sentinel should use sentinel name")
+		}
+
+		// Footer "Review by" should use display name
+		if !strings.Contains(body, "*Review by Security Specialist*") {
+			t.Error("footer should use display name")
+		}
+	})
+
+	t.Run("without display name", func(t *testing.T) {
+		body := FormatMarkdownWithDisplay(result, "", "reviewer")
+
+		// Should fall back to sentinel name for header
+		if !strings.Contains(body, "# Reviewer Review") {
+			t.Error("header should fall back to sentinel name")
+		}
+
+		if !strings.Contains(body, "<!-- review-bot:reviewer -->") {
+			t.Error("sentinel should use sentinel name")
+		}
+	})
+
+	t.Run("empty both names", func(t *testing.T) {
+		body := FormatMarkdownWithDisplay(result, "", "")
+
+		// Should not have header
+		if strings.Contains(body, "# ") && strings.Contains(body, " Review") {
+			t.Error("should not have header when both names empty")
+		}
+
+		// Should not have sentinel
+		if strings.Contains(body, "<!-- review-bot:") {
+			t.Error("should not have sentinel when sentinel name empty")
+		}
+	})
+}
+
+func TestSanitizeMarkdownText(t *testing.T) {
+	tests := []struct {
+		name  string
+		input string
+		want  string
+	}{
+		{
+			name:  "plain text unchanged",
+			input: "Security Specialist",
+			want:  "Security Specialist",
+		},
+		{
+			name:  "escapes asterisks",
+			input: "**bold** attack",
+			want:  `\*\*bold\*\* attack`,
+		},
+		{
+			name:  "escapes brackets for links",
+			input: "[click me](http://evil.com)",
+			want:  `\[click me\]\(http://evil.com\)`,
+		},
+		{
+			name:  "escapes backticks",
+			input: "`code` injection",
+			want:  "\\`code\\` injection",
+		},
+		{
+			name:  "escapes angle brackets",
+			input: "<script>alert(1)</script>",
+			want:  `\<script\>alert\(1\)\</script\>`,
+		},
+		{
+			name:  "escapes hash for headers",
+			input: "# Fake Header",
+			want:  `\# Fake Header`,
+		},
+		{
+			name:  "escapes pipe for tables",
+			input: "col1 | col2",
+			want:  `col1 \| col2`,
+		},
+		{
+			name:  "removes control characters",
+			input: "hello\x00world\x1f",
+			want:  "helloworld",
+		},
+		{
+			name:  "preserves tabs and newlines",
+			input: "line1\n\tindented",
+			want:  "line1\n\tindented",
+		},
+		{
+			name:  "escapes tilde for strikethrough",
+			input: "~~strikethrough~~",
+			want:  `\~\~strikethrough\~\~`,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := sanitizeMarkdownText(tt.input)
+			if got != tt.want {
+				t.Errorf("sanitizeMarkdownText(%q) = %q, want %q", tt.input, got, tt.want)
+			}
+		})
+	}
+}

review/parser.go

@@ -29,7 +29,19 @@ func ParseResponse(response string) (*ReviewResult, error) {
 
 	var result ReviewResult
 	if err := json.Unmarshal([]byte(cleaned), &result); err != nil {
-		return nil, fmt.Errorf("parse LLM response as JSON: %w\nRaw response: %s", err, response)
+		// LLMs sometimes produce JSON with unescaped quotes inside string values.
+		// Try to repair before giving up.
+		repaired := repairJSON(cleaned)
+		if err2 := json.Unmarshal([]byte(repaired), &result); err2 != nil {
+			// Include diagnostic info: lengths help identify truncation
+			rawLen := len(response)
+			cleanedLen := len(cleaned)
+			preview := cleaned
+			if len(preview) > 200 {
+				preview = preview[:100] + "..." + preview[len(preview)-100:]
+			}
+			return nil, fmt.Errorf("parse LLM response as JSON: %w\nRaw length: %d, cleaned length: %d\nCleaned preview: %s", err, rawLen, cleanedLen, preview)
+		}
 	}
 
 	// Validate verdict
@@ -74,3 +86,230 @@ func extractJSON(s string) string {
 	s = strings.TrimSpace(s)
 	return s
 }
+
+// repairJSON attempts to fix common LLM JSON issues:
+// - Unescaped double quotes inside string values
+//
+// Strategy: walk the JSON structurally. Object keys are parsed normally (LLMs
+// get those right). For string VALUES, we find all candidate closing quotes and
+// pick the LAST one that leaves valid JSON structure afterward — maximizing
+// string content, which is the correct bias for the "LLM put unescaped quotes
+// in a string value" failure mode.
+func repairJSON(s string) string {
+	runes := []rune(s)
+	var out strings.Builder
+	out.Grow(len(s) + 64)
+
+	i := 0
+	for i < len(runes) {
+		c := runes[i]
+
+		if c != '"' {
+			out.WriteRune(c)
+			i++
+			continue
+		}
+
+		// We hit an opening quote. Determine if this is a key or a value.
+		// Keys: the standard JSON parser in LLMs gets keys right, so we parse
+		// them normally (first unescaped quote closes).
+		// Values: may contain unescaped quotes — use the repair heuristic.
+		isValue := isValuePosition(runes, i)
+
+		if !isValue {
+			// Parse key/simple string normally
+			out.WriteRune('"')
+			i++
+			for i < len(runes) {
+				ch := runes[i]
+				if ch == '\\' && i+1 < len(runes) {
+					out.WriteRune(ch)
+					i++
+					out.WriteRune(runes[i])
+					i++
+					continue
+				}
+				if ch == '"' {
+					out.WriteRune('"')
+					i++
+					break
+				}
+				out.WriteRune(ch)
+				i++
+			}
+			continue
+		}
+
+		// Value string — find the correct close using last-valid-candidate heuristic
+		out.WriteRune('"')
+		i++
+
+		closeIdx := findClosingQuote(runes, i)
+
+		// Write everything between open and close, escaping interior quotes
+		for j := i; j < closeIdx; j++ {
+			ch := runes[j]
+			if ch == '\\' && j+1 < closeIdx {
+				// Already-escaped sequence — pass through
+				out.WriteRune(ch)
+				j++
+				out.WriteRune(runes[j])
+			} else if ch == '"' {
+				out.WriteRune('\\')
+				out.WriteRune('"')
+			} else {
+				out.WriteRune(ch)
+			}
+		}
+
+		// Write the closing quote
+		out.WriteRune('"')
+		i = closeIdx + 1
+	}
+
+	return out.String()
+}
+
+// isValuePosition determines if the quote at position i is opening a JSON value
+// string (as opposed to an object key). We only apply repair to values that
+// follow ':' since those are the free-text fields where LLMs produce unescaped
+// quotes. Array elements and keys are left alone (parsed normally).
+func isValuePosition(runes []rune, i int) bool {
+	// Look backward, skipping whitespace, for the preceding structural char
+	j := i - 1
+	for j >= 0 && (runes[j] == ' ' || runes[j] == '\t' || runes[j] == '\n' || runes[j] == '\r') {
+		j--
+	}
+	if j < 0 {
+		return false
+	}
+	// After ':' → definitely a value
+	return runes[j] == ':'
+}
+
+// findClosingQuote finds the index of the true closing quote for a JSON string
+// value starting at position start (the character after the opening quote).
+// It collects all unescaped quote candidates and returns the FIRST one that
+// produces valid JSON continuation (deeper lookahead verifies the next token).
+func findClosingQuote(runes []rune, start int) int {
+	// Collect all candidate positions for the closing quote.
+	var candidates []int
+	for j := start; j < len(runes); j++ {
+		if runes[j] == '\\' {
+			j++ // skip escaped character
+			continue
+		}
+		if runes[j] == '"' {
+			candidates = append(candidates, j)
+		}
+	}
+
+	if len(candidates) == 0 {
+		return len(runes)
+	}
+
+	if len(candidates) == 1 {
+		return candidates[0]
+	}
+
+	// Try candidates from FIRST to LAST. The correct closing quote is the
+	// earliest one that produces valid JSON structure after it (verified by
+	// deeper lookahead that checks the next token is a valid JSON start).
+	for _, idx := range candidates {
+		if isValidJSONAfterClose(runes, idx+1) {
+			return idx
+		}
+	}
+
+	// Fallback: return the last candidate
+	return candidates[len(candidates)-1]
+}
+
+// isValidJSONAfterClose checks whether the runes after a candidate closing quote
+// look like valid JSON continuation for a VALUE string. Since we only use this
+// for value positions, ':' is NOT a valid continuation (values are never keys).
+// Checks deeper structure to avoid being fooled by JSON-like content in strings.
+func isValidJSONAfterClose(runes []rune, pos int) bool {
+	j := pos
+	for j < len(runes) && (runes[j] == ' ' || runes[j] == '\t' || runes[j] == '\n' || runes[j] == '\r') {
+		j++
+	}
+
+	if j >= len(runes) {
+		return true
+	}
+
+	next := runes[j]
+	if next == '}' || next == ']' {
+		// Closing a container. Verify what follows the close is also valid:
+		// another structural char, comma, or EOF.
+		return isValidAfterContainerClose(runes, j+1)
+	}
+	if next == ',' {
+		// After comma, must be followed by a valid JSON token
+		j++
+		for j < len(runes) && (runes[j] == ' ' || runes[j] == '\t' || runes[j] == '\n' || runes[j] == '\r') {
+			j++
+		}
+		if j >= len(runes) {
+			return false // trailing comma with nothing after — invalid
+		}
+		return isJSONTokenStart(runes, j)
+	}
+	// ':' is NOT valid here — we're in a value position, not a key.
+	// Any other character is also invalid.
+	return false
+}
+
+// isValidAfterContainerClose checks that after a } or ], the continuation is
+// structurally valid: more closes, comma+token, or EOF.
+func isValidAfterContainerClose(runes []rune, pos int) bool {
+	j := pos
+	for j < len(runes) && (runes[j] == ' ' || runes[j] == '\t' || runes[j] == '\n' || runes[j] == '\r') {
+		j++
+	}
+	if j >= len(runes) {
+		return true
+	}
+	next := runes[j]
+	if next == '}' || next == ']' {
+		return isValidAfterContainerClose(runes, j+1)
+	}
+	if next == ',' {
+		j++
+		for j < len(runes) && (runes[j] == ' ' || runes[j] == '\t' || runes[j] == '\n' || runes[j] == '\r') {
+			j++
+		}
+		if j >= len(runes) {
+			return false
+		}
+		return isJSONTokenStart(runes, j)
+	}
+	return false
+}
+
+// isJSONTokenStart returns true if the rune could begin a JSON value or key.
+// For keywords (true/false/null), verifies the full keyword is present.
+func isJSONTokenStart(runes []rune, pos int) bool {
+	if pos >= len(runes) {
+		return false
+	}
+	r := runes[pos]
+	switch {
+	case r == '"': // string
+		return true
+	case r == '{' || r == '[': // object or array
+		return true
+	case r == 't': // true
+		return pos+4 <= len(runes) && string(runes[pos:pos+4]) == "true"
+	case r == 'f': // false
+		return pos+5 <= len(runes) && string(runes[pos:pos+5]) == "false"
+	case r == 'n': // null
+		return pos+4 <= len(runes) && string(runes[pos:pos+4]) == "null"
+	case r >= '0' && r <= '9': // number
+		return true
+	case r == '-': // negative number
+		return true
+	}
+	return false
+}

review/parser_test.go

@@ -1,6 +1,7 @@
 package review
 
 import (
+	"encoding/json"
 	"testing"
 )
 
@@ -112,3 +113,112 @@ func TestParseResponse_MarkdownFencesNoLang(t *testing.T) {
 		t.Errorf("expected APPROVE, got %q", result.Verdict)
 	}
 }
+
+func TestParseResponse_UnescapedQuotesInStrings(t *testing.T) {
+	// Real failure from CI: Sonnet puts unescaped quotes like (e.g. "28") in findings
+	input := `{"verdict": "APPROVE", "summary": "Clean PR", "findings": [{"severity": "NIT", "file": "ci/Dockerfile", "line": 14, "finding": "The comment says OTP_VERSION is the major version (e.g. \"28\") but it actually contains unescaped quotes like (e.g. "28") which breaks JSON"}], "recommendation": "Ship it"}`
+
+	result, err := ParseResponse(input)
+	if err != nil {
+		t.Fatalf("expected repair to handle unescaped quotes, got error: %v", err)
+	}
+	if result.Verdict != "APPROVE" {
+		t.Errorf("expected APPROVE, got %q", result.Verdict)
+	}
+	if len(result.Findings) != 1 {
+		t.Fatalf("expected 1 finding, got %d", len(result.Findings))
+	}
+}
+
+func TestRepairJSON_NoOpOnValid(t *testing.T) {
+	valid := `{"key": "value", "num": 42}`
+	result := repairJSON(valid)
+	if result != valid {
+		t.Errorf("repairJSON should not modify valid JSON\n  got:  %s\n  want: %s", result, valid)
+	}
+}
+
+func TestRepairJSON_FixesUnescapedQuotes(t *testing.T) {
+	// Interior quote followed by non-structural character
+	input := `{"msg": "use "foo" here"}`
+	result := repairJSON(input)
+
+	// Should be parseable now
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(result), &m); err != nil {
+		t.Fatalf("repaired JSON should parse, got: %v\nrepaired: %s", err, result)
+	}
+}
+
+func TestRepairJSON_InteriorQuoteBeforeComma(t *testing.T) {
+	// Bug reported by reviewer: interior quoted word immediately before a comma
+	input := `{"msg": "say "yes", and go"}`
+	result := repairJSON(input)
+
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(result), &m); err != nil {
+		t.Fatalf("repaired JSON should parse, got: %v\nrepaired: %s", err, result)
+	}
+	// The full string content should be preserved
+	msg, ok := m["msg"].(string)
+	if !ok {
+		t.Fatal("msg field missing or not a string")
+	}
+	if msg != `say "yes", and go` {
+		t.Errorf("unexpected msg content: %q", msg)
+	}
+}
+
+func TestRepairJSON_InteriorQuoteBeforeCloseBrace(t *testing.T) {
+	// Bug reported by reviewer: JSON-shaped syntax inside string values
+	input := `{"msg": "input map {"key": "val"} caused error"}`
+	result := repairJSON(input)
+
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(result), &m); err != nil {
+		t.Fatalf("repaired JSON should parse, got: %v\nrepaired: %s", err, result)
+	}
+}
+
+func TestRepairJSON_MultipleFields(t *testing.T) {
+	// Multiple string fields with unescaped quotes in different positions
+	input := `{"a": "hello "world"", "b": "foo"}`
+	result := repairJSON(input)
+
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(result), &m); err != nil {
+		t.Fatalf("repaired JSON should parse, got: %v\nrepaired: %s", err, result)
+	}
+	if _, ok := m["b"]; !ok {
+		t.Error("expected 'b' field to be preserved")
+	}
+}
+
+func TestRepairJSON_PreservesEscapedQuotes(t *testing.T) {
+	// Already-escaped quotes should not be double-escaped
+	input := `{"msg": "already \"escaped\" here"}`
+	result := repairJSON(input)
+
+	if result != input {
+		t.Errorf("repairJSON should not modify already-escaped quotes\n  got:  %s\n  want: %s", result, input)
+	}
+
+	var m map[string]interface{}
+	if err := json.Unmarshal([]byte(result), &m); err != nil {
+		t.Fatalf("repaired JSON should parse, got: %v\nrepaired: %s", err, result)
+	}
+}
+
+func TestRepairJSON_ComplexNestedContent(t *testing.T) {
+	// Combines both reviewer bugs: quoted words before commas AND JSON-like content
+	input := `{"verdict": "APPROVE", "findings": [{"finding": "The map {"key": "val"} and (e.g. "28") and say "yes", then stop"}]}`
+	result := repairJSON(input)
+
+	var parsed map[string]interface{}
+	if err := json.Unmarshal([]byte(result), &parsed); err != nil {
+		t.Fatalf("repaired JSON should parse, got: %v\nrepaired: %s", err, result)
+	}
+	if parsed["verdict"] != "APPROVE" {
+		t.Errorf("expected verdict APPROVE, got %v", parsed["verdict"])
+	}
+}

review/persona.go

@@ -0,0 +1,252 @@
+package review
+
+import (
+	"bytes"
+	"embed"
+	"encoding/json"
+	"fmt"
+	"os"
+	"sort"
+	"strings"
+	"unicode/utf8"
+
+	"gopkg.in/yaml.v3"
+)
+
+//go:embed personas/*.yaml
+var embeddedPersonas embed.FS
+
+// MaxPersonaFileSize is the maximum size for persona files (64 KB).
+// This prevents denial-of-service via excessively large files.
+const MaxPersonaFileSize = 64 * 1024
+
+// MaxYAMLDepth is the maximum nesting depth allowed in YAML persona files.
+// This prevents stack exhaustion from deeply nested structures.
+const MaxYAMLDepth = 20
+
+// MaxYAMLNodes is the maximum number of YAML nodes allowed in persona files.
+// This prevents DoS via wide-but-shallow structures that bypass depth limits.
+const MaxYAMLNodes = 1000
+
+// Persona defines a specialized review role with focused expertise.
+type Persona struct {
+	Name         string   `json:"name" yaml:"name"`
+	DisplayName  string   `json:"display_name" yaml:"display_name"`
+	ModelPref    string   `json:"model_preference,omitempty" yaml:"model_preference,omitempty"`
+	Identity     string   `json:"identity" yaml:"identity"`
+	Focus        []string `json:"focus" yaml:"focus"`
+	Ignore       []string `json:"ignore" yaml:"ignore"`
+	Severity     Severity `json:"severity" yaml:"severity"`
+	OutputFormat string   `json:"output_format,omitempty" yaml:"output_format,omitempty"`
+}
+
+// Severity defines what constitutes each severity level for this persona.
+// These are prompt guidance for the LLM, not output format changes.
+type Severity struct {
+	Major string `json:"major" yaml:"major"`
+	Minor string `json:"minor" yaml:"minor"`
+	Nit   string `json:"nit" yaml:"nit"`
+}
+
+// LoadPersona loads a persona from a JSON or YAML file path.
+// Format is detected by file extension: .yaml/.yml for YAML, .json or other for JSON.
+// Files larger than MaxPersonaFileSize are rejected.
+//
+// Symlinks are supported: os.Stat follows symlinks, so a symlink pointing to
+// a regular file will pass the IsRegular() check. Symlinks to non-regular files
+// (directories, FIFOs, devices) are still rejected.
+func LoadPersona(path string) (*Persona, error) {
+	// os.Stat follows symlinks, so symlinks to regular files are supported.
+	// The IsRegular() check operates on the target, not the symlink itself.
+	info, err := os.Stat(path)
+	if err != nil {
+		return nil, fmt.Errorf("read persona file %s: %w", path, err)
+	}
+	if !info.Mode().IsRegular() {
+		return nil, fmt.Errorf("persona file %s is not a regular file", path)
+	}
+	if info.Size() > MaxPersonaFileSize {
+		return nil, fmt.Errorf("persona file %s exceeds maximum size (%d bytes)", path, MaxPersonaFileSize)
+	}
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, fmt.Errorf("read persona file %s: %w", path, err)
+	}
+	// Re-check size after read to defend against TOCTOU races where file
+	// grows between stat and read (e.g., appending process, replaced file).
+	if len(data) > MaxPersonaFileSize {
+		return nil, fmt.Errorf("persona file %s exceeds maximum size (%d bytes)", path, MaxPersonaFileSize)
+	}
+	return parsePersona(data, path)
+}
+
+// LoadBuiltinPersona loads a built-in persona by name.
+// Returns an error if the persona doesn't exist.
+// Built-in personas are stored in YAML format only (see embed directive).
+func LoadBuiltinPersona(name string) (*Persona, error) {
+	yamlFile := name + ".yaml"
+	data, err := embeddedPersonas.ReadFile("personas/" + yamlFile)
+	if err != nil {
+		available := ListBuiltinPersonas()
+		return nil, fmt.Errorf("unknown built-in persona %q (available: %s)", name, strings.Join(available, ", "))
+	}
+	return parsePersona(data, "builtin:"+yamlFile)
+}
+
+// ListBuiltinPersonas returns the names of all built-in personas in sorted order.
+// Returns an empty slice if the embedded directory cannot be read.
+func ListBuiltinPersonas() []string {
+	entries, err := embeddedPersonas.ReadDir("personas")
+	if err != nil {
+		return []string{}
+	}
+	seen := make(map[string]bool)
+	for _, e := range entries {
+		if e.IsDir() {
+			continue
+		}
+		name := e.Name()
+		// Strip extension to get persona name
+		var personaName string
+		switch {
+		case strings.HasSuffix(name, ".yaml"):
+			personaName = strings.TrimSuffix(name, ".yaml")
+		case strings.HasSuffix(name, ".yml"):
+			personaName = strings.TrimSuffix(name, ".yml")
+		case strings.HasSuffix(name, ".json"):
+			personaName = strings.TrimSuffix(name, ".json")
+		default:
+			continue
+		}
+		if !seen[personaName] {
+			seen[personaName] = true
+		}
+	}
+	names := make([]string, 0, len(seen))
+	for name := range seen {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+	return names
+}
+
+// parsePersona parses persona data from JSON or YAML format.
+// Format is detected by the source file extension.
+func parsePersona(data []byte, source string) (*Persona, error) {
+	lowerSource := strings.ToLower(source)
+	isYAML := strings.HasSuffix(lowerSource, ".yaml") || strings.HasSuffix(lowerSource, ".yml")
+
+	var p Persona
+	var err error
+	if isYAML {
+		err = unmarshalYAMLWithDepthLimit(data, &p, MaxYAMLDepth)
+	} else {
+		// Use json.Decoder with DisallowUnknownFields for consistency with
+		// YAML's KnownFields(true) - both reject unknown fields to catch typos.
+		dec := json.NewDecoder(bytes.NewReader(data))
+		dec.DisallowUnknownFields()
+		err = dec.Decode(&p)
+	}
+	if err != nil {
+		return nil, fmt.Errorf("parse persona %s: %w", source, err)
+	}
+	if err := validatePersona(&p, source); err != nil {
+		return nil, err
+	}
+	return &p, nil
+}
+
+// unmarshalYAMLWithDepthLimit unmarshals YAML data with explicit depth limiting
+// and strict field checking. This protects against stack exhaustion from deeply
+// nested structures and catches typos in field names.
+// Multi-document YAML files are rejected to prevent silent data loss.
+func unmarshalYAMLWithDepthLimit(data []byte, out any, maxDepth int) error {
+	// First pass: decode into a yaml.Node to check depth limits and node counts.
+	// This prevents stack exhaustion before we attempt to decode into structs.
+	var node yaml.Node
+	dec := yaml.NewDecoder(bytes.NewReader(data))
+	if err := dec.Decode(&node); err != nil {
+		return err
+	}
+
+	// Reject multi-document YAML files - silently ignoring additional documents
+	// could lead to confusing behavior where users think their changes take effect.
+	var extra yaml.Node
+	if dec.Decode(&extra) == nil {
+		return fmt.Errorf("multi-document YAML is not supported; only single-document files are allowed")
+	}
+
+	nodeCount := 0
+	if err := checkYAMLDepth(&node, 0, maxDepth, MaxYAMLNodes, make(map[*yaml.Node]struct{}), &nodeCount); err != nil {
+		return err
+	}
+
+	// Second pass: decode with strict field checking enabled.
+	// KnownFields(true) rejects unknown keys, catching typos like "focuss" or "identiy".
+	// We must re-decode from the original data because yaml.Node.Decode() doesn't
+	// support the KnownFields option.
+	strictDec := yaml.NewDecoder(bytes.NewReader(data))
+	strictDec.KnownFields(true)
+	return strictDec.Decode(out)
+}
+
+// checkYAMLDepth recursively checks that YAML nodes don't exceed the depth limit
+// or the total node count limit. It also detects alias cycles to prevent infinite
+// recursion from crafted YAML with self-referential aliases.
+func checkYAMLDepth(node *yaml.Node, depth, maxDepth, maxNodes int, seen map[*yaml.Node]struct{}, nodeCount *int) error {
+	if depth > maxDepth {
+		return fmt.Errorf("YAML nesting depth exceeds maximum (%d)", maxDepth)
+	}
+
+	// Track total nodes visited as defense-in-depth against wide-but-shallow attacks.
+	*nodeCount++
+	if *nodeCount > maxNodes {
+		return fmt.Errorf("YAML node count exceeds maximum (%d)", maxNodes)
+	}
+
+	// Cycle detection: if we've seen this node before, we're in a cycle.
+	if _, ok := seen[node]; ok {
+		return nil // Already validated this subtree, skip to avoid infinite recursion.
+	}
+	seen[node] = struct{}{}
+
+	// Handle alias nodes: follow the alias to its anchor target.
+	// Increment depth when following aliases since they expand the effective structure.
+	if node.Kind == yaml.AliasNode && node.Alias != nil {
+		return checkYAMLDepth(node.Alias, depth+1, maxDepth, maxNodes, seen, nodeCount)
+	}
+
+	for _, child := range node.Content {
+		if err := checkYAMLDepth(child, depth+1, maxDepth, maxNodes, seen, nodeCount); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func validatePersona(p *Persona, source string) error {
+	if p.Name == "" {
+		return fmt.Errorf("persona %s: name is required", source)
+	}
+	if p.Identity == "" {
+		return fmt.Errorf("persona %s: identity is required", source)
+	}
+	// DisplayName defaults to Name if not set
+	if p.DisplayName == "" {
+		p.DisplayName = p.Name
+	}
+	return nil
+}
+
+// CapitalizeFirst capitalizes the first rune of a string in a Unicode-safe way.
+// Returns the original string if it's empty.
+func CapitalizeFirst(s string) string {
+	if s == "" {
+		return s
+	}
+	r, size := utf8.DecodeRuneInString(s)
+	if r == utf8.RuneError {
+		return s
+	}
+	return strings.ToUpper(string(r)) + s[size:]
+}

review/persona_prompt.go

@@ -0,0 +1,104 @@
+package review
+
+import (
+	"fmt"
+	"strings"
+)
+
+// BuildPersonaSystemPrompt constructs a system prompt from a persona definition.
+// This replaces BuildSystemBase when a persona is provided.
+func BuildPersonaSystemPrompt(p *Persona) string {
+	var sb strings.Builder
+
+	// Identity section
+	sb.WriteString(p.Identity)
+	sb.WriteString("\n\n")
+
+	// Focus section
+	if len(p.Focus) > 0 {
+		sb.WriteString("## Focus Areas\n\n")
+		sb.WriteString("Concentrate your review on:\n")
+		for _, f := range p.Focus {
+			sb.WriteString(fmt.Sprintf("- %s\n", f))
+		}
+		sb.WriteString("\n")
+	}
+
+	// Ignore section
+	if len(p.Ignore) > 0 {
+		sb.WriteString("## Explicitly Out of Scope\n\n")
+		sb.WriteString("Do NOT comment on:\n")
+		for _, i := range p.Ignore {
+			sb.WriteString(fmt.Sprintf("- %s\n", i))
+		}
+		sb.WriteString("\n")
+	}
+
+	// Severity calibration
+	if p.Severity.Major != "" || p.Severity.Minor != "" || p.Severity.Nit != "" {
+		sb.WriteString("## Severity Calibration\n\n")
+		sb.WriteString("Use these severity definitions for YOUR domain:\n")
+		if p.Severity.Major != "" {
+			sb.WriteString(fmt.Sprintf("- **MAJOR**: %s\n", p.Severity.Major))
+		}
+		if p.Severity.Minor != "" {
+			sb.WriteString(fmt.Sprintf("- **MINOR**: %s\n", p.Severity.Minor))
+		}
+		if p.Severity.Nit != "" {
+			sb.WriteString(fmt.Sprintf("- **NIT**: %s\n", p.Severity.Nit))
+		}
+		sb.WriteString("\n")
+	}
+
+	// Output format instructions (shared schema from prompt.go)
+	sb.WriteString("## Review Instructions\n\n")
+	sb.WriteString("CONTEXT:\n")
+	sb.WriteString("- You will receive the full content of modified files for reference, followed by the diff showing what changed.\n")
+	sb.WriteString("- The diff shows ONLY what was added/removed. The full file content provides complete context.\n")
+	sb.WriteString("- Focus your review on the CHANGES (the diff), using the full files for context.\n\n")
+	sb.WriteString("Your task:\n")
+	sb.WriteString("1. Review the diff for issues within YOUR focus areas only.\n")
+	sb.WriteString("2. Consider the CI status — if CI has failed, that is an automatic REQUEST_CHANGES regardless of code quality.\n")
+	sb.WriteString("3. Output your review as structured JSON (and ONLY JSON, no markdown fences or other text).\n\n")
+	sb.WriteString("Output format:\n")
+	sb.WriteString(outputSchemaJSON)
+	sb.WriteString("\n\n")
+	sb.WriteString(verdictRules)
+	sb.WriteString("\n- Only report findings within your focus areas. Ignore everything else.\n")
+	sb.WriteString("- Line numbers should reference the new file line numbers from the diff headers.\n")
+	sb.WriteString("- If the diff has no changes relevant to your focus areas, APPROVE with no findings.\n")
+
+	// Custom output format if provided
+	if p.OutputFormat != "" {
+		sb.WriteString("\n\n## Additional Output Guidelines\n\n")
+		sb.WriteString(p.OutputFormat)
+	}
+
+	return sb.String()
+}
+
+// BuildSystemPromptWithPersona constructs the full system prompt, using either
+// a persona or the default generic prompt. This is a convenience wrapper that
+// combines BuildPersonaSystemPrompt (or BuildSystemBase) with patterns and conventions.
+// It is exported for use by callers who want one-shot prompt assembly.
+func BuildSystemPromptWithPersona(persona *Persona, conventions, patterns string) string {
+	var base string
+	if persona != nil {
+		base = BuildPersonaSystemPrompt(persona)
+	} else {
+		base = BuildSystemBase()
+	}
+
+	var sb strings.Builder
+	sb.WriteString(base)
+
+	if patterns != "" {
+		sb.WriteString(fmt.Sprintf("\n\n## Language Patterns & Idioms\n\nUse the following patterns as review criteria. Code that violates these established patterns is a finding:\n\n%s\n", patterns))
+	}
+
+	if conventions != "" {
+		sb.WriteString(fmt.Sprintf("\n\n## Repository Conventions\n\nThe repository has the following coding conventions that must be respected:\n\n%s\n", conventions))
+	}
+
+	return sb.String()
+}

review/persona_prompt_test.go

@@ -0,0 +1,157 @@
+package review
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestBuildPersonaSystemPrompt(t *testing.T) {
+	p := &Persona{
+		Name:        "security",
+		DisplayName: "Security Specialist",
+		Identity:    "You are a security specialist.",
+		Focus:       []string{"injection attacks", "auth bypass"},
+		Ignore:      []string{"code style", "performance"},
+		Severity: Severity{
+			Major: "exploitable vulnerabilities",
+			Minor: "defense in depth",
+			Nit:   "theoretical risks",
+		},
+	}
+
+	prompt := BuildPersonaSystemPrompt(p)
+
+	// Check identity is included
+	if !strings.Contains(prompt, "You are a security specialist.") {
+		t.Error("prompt should contain identity")
+	}
+
+	// Check focus areas
+	if !strings.Contains(prompt, "Focus Areas") {
+		t.Error("prompt should contain Focus Areas section")
+	}
+	if !strings.Contains(prompt, "injection attacks") {
+		t.Error("prompt should contain focus item")
+	}
+
+	// Check ignore section
+	if !strings.Contains(prompt, "Out of Scope") {
+		t.Error("prompt should contain Out of Scope section")
+	}
+	if !strings.Contains(prompt, "code style") {
+		t.Error("prompt should contain ignore item")
+	}
+
+	// Check severity calibration
+	if !strings.Contains(prompt, "Severity Calibration") {
+		t.Error("prompt should contain Severity Calibration section")
+	}
+	if !strings.Contains(prompt, "exploitable vulnerabilities") {
+		t.Error("prompt should contain major severity definition")
+	}
+
+	// Check JSON output format is included
+	if !strings.Contains(prompt, `"verdict"`) {
+		t.Error("prompt should contain JSON output format")
+	}
+	if !strings.Contains(prompt, "APPROVE") {
+		t.Error("prompt should mention APPROVE verdict")
+	}
+}
+
+func TestBuildPersonaSystemPromptMinimal(t *testing.T) {
+	// Minimal persona with only required fields
+	p := &Persona{
+		Name:     "minimal",
+		Identity: "You are a minimal reviewer.",
+	}
+
+	prompt := BuildPersonaSystemPrompt(p)
+
+	// Should still work without optional fields
+	if !strings.Contains(prompt, "You are a minimal reviewer.") {
+		t.Error("prompt should contain identity")
+	}
+
+	// Should not have empty sections
+	if strings.Contains(prompt, "Focus Areas") && !strings.Contains(prompt, "Concentrate your review on:") {
+		t.Error("should not have Focus Areas header without content")
+	}
+}
+
+func TestBuildSystemPromptWithPersona(t *testing.T) {
+	t.Run("with persona", func(t *testing.T) {
+		p := &Persona{
+			Name:     "test",
+			Identity: "Test persona identity.",
+			Focus:    []string{"testing"},
+		}
+
+		prompt := BuildSystemPromptWithPersona(p, "test conventions", "test patterns")
+
+		if !strings.Contains(prompt, "Test persona identity.") {
+			t.Error("should contain persona identity")
+		}
+		if !strings.Contains(prompt, "test conventions") {
+			t.Error("should contain conventions")
+		}
+		if !strings.Contains(prompt, "test patterns") {
+			t.Error("should contain patterns")
+		}
+	})
+
+	t.Run("without persona", func(t *testing.T) {
+		prompt := BuildSystemPromptWithPersona(nil, "test conventions", "test patterns")
+
+		// Should use default system base
+		if !strings.Contains(prompt, "expert code reviewer") {
+			t.Error("should contain default system base when no persona")
+		}
+		if !strings.Contains(prompt, "test conventions") {
+			t.Error("should contain conventions")
+		}
+	})
+
+	t.Run("empty conventions and patterns", func(t *testing.T) {
+		p := &Persona{
+			Name:     "test",
+			Identity: "Test identity.",
+		}
+
+		prompt := BuildSystemPromptWithPersona(p, "", "")
+
+		if strings.Contains(prompt, "Language Patterns") {
+			t.Error("should not contain patterns section when empty")
+		}
+		if strings.Contains(prompt, "Repository Conventions") {
+			t.Error("should not contain conventions section when empty")
+		}
+	})
+}
+
+func TestPersonaPromptContainsOutputRules(t *testing.T) {
+	p := &Persona{
+		Name:     "test",
+		Identity: "Test.",
+	}
+
+	prompt := BuildPersonaSystemPrompt(p)
+
+	// Must contain the critical output rules
+	requiredStrings := []string{
+		"APPROVE",
+		"REQUEST_CHANGES",
+		"MAJOR",
+		"MINOR",
+		"NIT",
+		"verdict",
+		"findings",
+		"CI",
+	}
+
+	for _, s := range requiredStrings {
+		if !strings.Contains(prompt, s) {
+			t.Errorf("prompt should contain %q", s)
+		}
+	}
+}

review/persona_test.go

@@ -0,0 +1,778 @@
+package review
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gopkg.in/yaml.v3"
+)
+
+func TestLoadBuiltinPersona(t *testing.T) {
+	tests := []struct {
+		name        string
+		personaName string
+		wantErr     bool
+		wantDisplay string
+	}{
+		{
+			name:        "security persona",
+			personaName: "security",
+			wantErr:     false,
+			wantDisplay: "Security Specialist",
+		},
+		{
+			name:        "architect persona",
+			personaName: "architect",
+			wantErr:     false,
+			wantDisplay: "Software Architect",
+		},
+		{
+			name:        "docs persona",
+			personaName: "docs",
+			wantErr:     false,
+			wantDisplay: "Documentation Reviewer",
+		},
+		{
+			name:        "unknown persona",
+			personaName: "nonexistent",
+			wantErr:     true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			p, err := LoadBuiltinPersona(tt.personaName)
+			if tt.wantErr {
+				if err == nil {
+					t.Error("expected error, got nil")
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("unexpected error: %v", err)
+			}
+			if p.Name != tt.personaName {
+				t.Errorf("Name = %q, want %q", p.Name, tt.personaName)
+			}
+			if p.DisplayName != tt.wantDisplay {
+				t.Errorf("DisplayName = %q, want %q", p.DisplayName, tt.wantDisplay)
+			}
+			if p.Identity == "" {
+				t.Error("Identity should not be empty")
+			}
+			if len(p.Focus) == 0 {
+				t.Error("Focus should not be empty")
+			}
+		})
+	}
+}
+
+func TestListBuiltinPersonas(t *testing.T) {
+	names := ListBuiltinPersonas()
+	if len(names) == 0 {
+		t.Fatal("expected at least one built-in persona")
+	}
+
+	// Check for expected personas
+	expected := map[string]bool{"security": false, "architect": false, "docs": false}
+	for _, name := range names {
+		if _, ok := expected[name]; ok {
+			expected[name] = true
+		}
+	}
+	for name, found := range expected {
+		if !found {
+			t.Errorf("expected built-in persona %q not found", name)
+		}
+	}
+}
+
+func TestLoadPersonaFromYAMLFile(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "test.yaml")
+
+	content := `# Test persona
+name: test
+display_name: Test Persona
+identity: |
+  You are a test persona.
+  Multi-line identity works.
+focus:
+  - testing
+  - validation
+ignore:
+  - nothing
+severity:
+  major: Big problems
+  minor: Small problems
+  nit: Tiny problems
+`
+
+	if err := os.WriteFile(path, []byte(content), 0644); err != nil {
+		t.Fatalf("failed to write test file: %v", err)
+	}
+
+	p, err := LoadPersona(path)
+	if err != nil {
+		t.Fatalf("LoadPersona failed: %v", err)
+	}
+
+	if p.Name != "test" {
+		t.Errorf("Name = %q, want %q", p.Name, "test")
+	}
+	if p.DisplayName != "Test Persona" {
+		t.Errorf("DisplayName = %q, want %q", p.DisplayName, "Test Persona")
+	}
+	if len(p.Focus) != 2 {
+		t.Errorf("Focus len = %d, want 2", len(p.Focus))
+	}
+	if !strings.Contains(p.Identity, "Multi-line") {
+		t.Error("Identity should contain multi-line content")
+	}
+}
+
+func TestLoadPersonaFromYMLFile(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "test.yml")
+
+	content := `name: test
+display_name: Test YML
+identity: Test identity
+focus:
+  - testing
+ignore: []
+severity:
+  major: Big
+  minor: Small
+  nit: Tiny
+`
+
+	if err := os.WriteFile(path, []byte(content), 0644); err != nil {
+		t.Fatalf("failed to write test file: %v", err)
+	}
+
+	p, err := LoadPersona(path)
+	if err != nil {
+		t.Fatalf("LoadPersona failed: %v", err)
+	}
+
+	if p.Name != "test" {
+		t.Errorf("Name = %q, want %q", p.Name, "test")
+	}
+	if p.DisplayName != "Test YML" {
+		t.Errorf("DisplayName = %q, want %q", p.DisplayName, "Test YML")
+	}
+}
+
+func TestLoadPersonaFromJSONFile(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "test.json")
+
+	content := `{
+		"name": "test",
+		"display_name": "Test Persona",
+		"identity": "You are a test persona.\nMulti-line identity works.",
+		"focus": ["testing", "validation"],
+
+		"ignore": ["nothing"],
+		"severity": {
+			"major": "Big problems",
+			"minor": "Small problems",
+			"nit": "Tiny problems"
+		}
+	}`
+
+	if err := os.WriteFile(path, []byte(content), 0644); err != nil {
+		t.Fatalf("failed to write test file: %v", err)
+	}
+
+	p, err := LoadPersona(path)
+	if err != nil {
+		t.Fatalf("LoadPersona failed: %v", err)
+	}
+
+	if p.Name != "test" {
+		t.Errorf("Name = %q, want %q", p.Name, "test")
+	}
+	if p.DisplayName != "Test Persona" {
+		t.Errorf("DisplayName = %q, want %q", p.DisplayName, "Test Persona")
+	}
+	if len(p.Focus) != 2 {
+		t.Errorf("Focus len = %d, want 2", len(p.Focus))
+	}
+	if !strings.Contains(p.Identity, "Multi-line") {
+		t.Error("Identity should contain multi-line content")
+	}
+}
+
+func TestLoadPersonaValidation(t *testing.T) {
+	tests := []struct {
+		name    string
+		content string
+		ext     string
+		wantErr string
+	}{
+		{
+			name:    "missing name yaml",
+			content: "identity: test\n",
+			ext:     ".yaml",
+			wantErr: "name is required",
+		},
+		{
+			name:    "missing identity yaml",
+			content: "name: test\n",
+			ext:     ".yaml",
+			wantErr: "identity is required",
+		},
+		{
+			name:    "missing name json",
+			content: `{"identity": "test"}`,
+			ext:     ".json",
+			wantErr: "name is required",
+		},
+		{
+			name:    "missing identity json",
+			content: `{"name": "test"}`,
+			ext:     ".json",
+			wantErr: "identity is required",
+		},
+		{
+			name:    "display_name defaults to name",
+			content: "name: test\nidentity: test identity\n",
+			ext:     ".yaml",
+			// No error expected - should succeed
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dir := t.TempDir()
+			path := filepath.Join(dir, "test"+tt.ext)
+			if err := os.WriteFile(path, []byte(tt.content), 0644); err != nil {
+				t.Fatalf("failed to write test file: %v", err)
+			}
+
+			p, err := LoadPersona(path)
+			if tt.wantErr != "" {
+				if err == nil {
+					t.Errorf("expected error containing %q, got nil", tt.wantErr)
+					return
+				}
+				if !strings.Contains(err.Error(), tt.wantErr) {
+					t.Errorf("error = %q, want containing %q", err.Error(), tt.wantErr)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("unexpected error: %v", err)
+			}
+			// Check display_name defaulting
+			if p.DisplayName == "" {
+				t.Error("DisplayName should default to Name")
+			}
+			if p.DisplayName != p.Name {
+				t.Errorf("DisplayName should default to Name, got %q", p.DisplayName)
+			}
+		})
+	}
+}
+
+func TestLoadPersonaFileNotFound(t *testing.T) {
+	_, err := LoadPersona("/nonexistent/path/persona.yaml")
+	if err == nil {
+		t.Error("expected error for nonexistent file")
+	}
+}
+
+func TestLoadPersonaInvalidYAML(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "invalid.yaml")
+	if err := os.WriteFile(path, []byte("not valid yaml:\n  - [broken"), 0644); err != nil {
+		t.Fatalf("failed to write test file: %v", err)
+	}
+
+	_, err := LoadPersona(path)
+	if err == nil {
+		t.Error("expected error for invalid YAML")
+	}
+}
+
+func TestLoadPersonaInvalidJSON(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "invalid.json")
+	if err := os.WriteFile(path, []byte("not valid json {"), 0644); err != nil {
+		t.Fatalf("failed to write test file: %v", err)
+	}
+
+	_, err := LoadPersona(path)
+	if err == nil {
+		t.Error("expected error for invalid JSON")
+	}
+}
+
+func TestLoadPersonaCaseInsensitiveExtension(t *testing.T) {
+	tests := []struct {
+		name string
+		ext  string
+	}{
+		{"lowercase yaml", ".yaml"},
+		{"uppercase YAML", ".YAML"},
+		{"mixed case Yaml", ".Yaml"},
+		{"lowercase yml", ".yml"},
+		{"uppercase YML", ".YML"},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dir := t.TempDir()
+			path := filepath.Join(dir, "test"+tt.ext)
+			content := "name: test\nidentity: test identity\n"
+			if err := os.WriteFile(path, []byte(content), 0644); err != nil {
+				t.Fatalf("failed to write test file: %v", err)
+			}
+
+			p, err := LoadPersona(path)
+			if err != nil {
+				t.Fatalf("LoadPersona failed for extension %s: %v", tt.ext, err)
+			}
+			if p.Name != "test" {
+				t.Errorf("Name = %q, want %q", p.Name, "test")
+			}
+		})
+	}
+}
+
+func TestCapitalizeFirst(t *testing.T) {
+	tests := []struct {
+		input string
+		want  string
+	}{
+		{"hello", "Hello"},
+		{"Hello", "Hello"},
+		{"HELLO", "HELLO"},
+		{"a", "A"},
+		{"", ""},
+		{"日本語", "日本語"}, // Non-ASCII: Japanese doesn't have case
+		{"über", "Über"},   // German umlaut
+		{"élève", "Élève"}, // French accent
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.input, func(t *testing.T) {
+			got := CapitalizeFirst(tt.input)
+			if got != tt.want {
+				t.Errorf("CapitalizeFirst(%q) = %q, want %q", tt.input, got, tt.want)
+			}
+		})
+	}
+}
+
+func TestListBuiltinPersonasReturnsEmptySlice(t *testing.T) {
+	// ListBuiltinPersonas should return an empty slice (not nil) on error.
+	// We can't easily test the error case, but we can verify the success case
+	// returns a proper slice.
+	names := ListBuiltinPersonas()
+	if names == nil {
+		t.Error("ListBuiltinPersonas should return empty slice, not nil")
+	}
+}
+
+func TestYAMLMultilineStrings(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "multiline.yaml")
+
+	// Test literal block scalar (|) which preserves newlines
+	content := `name: multiline
+display_name: Multiline Test
+identity: |
+  First line.
+  Second line.
+  Third line.
+focus:
+  - item one
+ignore: []
+severity:
+  major: Major issue
+  minor: Minor issue
+  nit: Nit
+`
+
+	if err := os.WriteFile(path, []byte(content), 0644); err != nil {
+		t.Fatalf("failed to write test file: %v", err)
+	}
+
+	p, err := LoadPersona(path)
+	if err != nil {
+		t.Fatalf("LoadPersona failed: %v", err)
+	}
+
+	// Literal block scalar preserves newlines
+	if !strings.Contains(p.Identity, "\n") {
+		t.Error("Identity should contain newlines from literal block scalar")
+	}
+	if !strings.Contains(p.Identity, "Second line") {
+		t.Error("Identity should contain 'Second line'")
+	}
+}
+
+func TestYAMLComments(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "comments.yaml")
+
+	content := `# This is a comment
+name: commented  # inline comment
+display_name: Commented Persona
+# Another comment
+identity: Test identity
+focus:
+  - item  # comment after item
+ignore: []
+severity:
+  major: Major
+  minor: Minor
+  nit: Nit
+`
+
+	if err := os.WriteFile(path, []byte(content), 0644); err != nil {
+		t.Fatalf("failed to write test file: %v", err)
+	}
+
+	p, err := LoadPersona(path)
+	if err != nil {
+		t.Fatalf("LoadPersona failed: %v", err)
+	}
+
+	// Comments should be ignored
+	if p.Name != "commented" {
+		t.Errorf("Name = %q, want %q", p.Name, "commented")
+	}
+	if p.Focus[0] != "item" {
+		t.Errorf("Focus[0] = %q, want %q", p.Focus[0], "item")
+	}
+}
+
+func TestYAMLDeeplyNestedRejection(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "deeply-nested.yaml")
+
+	// Build a deeply nested YAML structure that exceeds MaxYAMLDepth (20).
+	// Each level adds 2 to the depth count (key + value mapping).
+	var sb strings.Builder
+	sb.WriteString("name: test\nidentity: test\nnested:\n")
+	indent := "  "
+	for i := 0; i < 25; i++ {
+		sb.WriteString(strings.Repeat(indent, i+1))
+		sb.WriteString(fmt.Sprintf("level%d:\n", i))
+	}
+	sb.WriteString(strings.Repeat(indent, 26))
+	sb.WriteString("value: too-deep\n")
+
+	if err := os.WriteFile(path, []byte(sb.String()), 0644); err != nil {
+		t.Fatalf("failed to write test file: %v", err)
+	}
+
+	_, err := LoadPersona(path)
+	if err == nil {
+		t.Error("expected error for deeply nested YAML, got nil")
+	}
+	if !strings.Contains(err.Error(), "nesting depth exceeds") {
+		t.Errorf("error = %q, want containing 'nesting depth exceeds'", err.Error())
+	}
+}
+
+func TestYAMLFileSizeLimit(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "huge.yaml")
+
+	// Create a file larger than MaxPersonaFileSize (64 KB)
+	content := "name: test\nidentity: " + strings.Repeat("x", MaxPersonaFileSize+1) + "\n"
+	if err := os.WriteFile(path, []byte(content), 0644); err != nil {
+		t.Fatalf("failed to write test file: %v", err)
+	}
+
+	_, err := LoadPersona(path)
+	if err == nil {
+		t.Error("expected error for oversized file, got nil")
+	}
+	if !strings.Contains(err.Error(), "exceeds maximum size") {
+		t.Errorf("error = %q, want containing 'exceeds maximum size'", err.Error())
+	}
+}
+
+func TestYAMLAliasCycleDetection(t *testing.T) {
+	// Test that our checkYAMLDepth function handles alias cycles gracefully
+	// by using the seen map to prevent infinite recursion.
+	// We test this directly because go-yaml's parser handles most cycles
+	// at parse time, but we need to ensure our checker is robust.
+
+	// Create a node structure where an alias points to a parent node,
+	// simulating what could happen with malicious input that bypasses
+	// go-yaml's cycle detection.
+	parent := &yaml.Node{
+		Kind: yaml.MappingNode,
+		Content: []*yaml.Node{
+			{Kind: yaml.ScalarNode, Value: "name"},
+			{Kind: yaml.ScalarNode, Value: "test"},
+			{Kind: yaml.ScalarNode, Value: "nested"},
+		},
+	}
+
+	// Create a child that aliases back to the parent (artificial cycle)
+	aliasToParent := &yaml.Node{
+		Kind:  yaml.AliasNode,
+		Alias: parent,
+	}
+	parent.Content = append(parent.Content, aliasToParent)
+
+	nodeCount := 0
+	seen := make(map[*yaml.Node]struct{})
+
+	// This should NOT hang or stack overflow - the seen map prevents infinite recursion
+	err := checkYAMLDepth(parent, 0, MaxYAMLDepth, MaxYAMLNodes, seen, &nodeCount)
+	if err != nil {
+		t.Errorf("unexpected error traversing cyclic structure: %v", err)
+	}
+
+	// Verify we tracked the parent in the seen map
+	if _, ok := seen[parent]; !ok {
+		t.Error("parent node not tracked in seen map")
+	}
+}
+
+func TestYAMLMultiDocumentRejection(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "multi.yaml")
+
+	// Multi-document YAML (documents separated by ---)
+	content := `name: first
+identity: first document
+---
+name: second
+identity: second document
+`
+	if err := os.WriteFile(path, []byte(content), 0644); err != nil {
+		t.Fatalf("failed to write test file: %v", err)
+	}
+
+	_, err := LoadPersona(path)
+	if err == nil {
+		t.Error("expected error for multi-document YAML, got nil")
+	}
+	if !strings.Contains(err.Error(), "multi-document") {
+		t.Errorf("error = %q, want containing 'multi-document'", err.Error())
+	}
+}
+
+func TestYAMLNodeCountLimit(t *testing.T) {
+	dir := t.TempDir()
+	path := filepath.Join(dir, "wide.yaml")
+
+	// Build a YAML structure that's shallow but wide - many keys at the same level
+	// to test the node count limit (should exceed MaxYAMLNodes = 1000)
+	var sb strings.Builder
+	sb.WriteString("name: test\nidentity: test\n")
+	for i := 0; i < 600; i++ {
+		sb.WriteString(fmt.Sprintf("key%d: value%d\n", i, i))
+	}
+
+	if err := os.WriteFile(path, []byte(sb.String()), 0644); err != nil {
+		t.Fatalf("failed to write test file: %v", err)
+	}
+
+	_, err := LoadPersona(path)
+	if err == nil {
+		t.Error("expected error for wide YAML exceeding node count, got nil")
+	}
+	if !strings.Contains(err.Error(), "node count exceeds") {
+		t.Errorf("error = %q, want containing 'node count exceeds'", err.Error())
+	}
+}
+
+func TestCheckYAMLDepthCycleDetectionDirect(t *testing.T) {
+	// Direct test of cycle detection in checkYAMLDepth by creating
+	// a node structure with an artificial cycle.
+	// This tests the seen map logic independent of go-yaml's parsing.
+	node := &yaml.Node{
+		Kind: yaml.MappingNode,
+		Content: []*yaml.Node{
+			{Kind: yaml.ScalarNode, Value: "key"},
+			{Kind: yaml.ScalarNode, Value: "value"},
+		},
+	}
+
+	// Create a cycle by making a child reference the parent
+	cycleChild := &yaml.Node{
+		Kind:  yaml.AliasNode,
+		Alias: node, // Points back to the parent
+	}
+	node.Content = append(node.Content,
+		&yaml.Node{Kind: yaml.ScalarNode, Value: "cyclic"},
+		cycleChild,
+	)
+
+	nodeCount := 0
+	seen := make(map[*yaml.Node]struct{})
+	err := checkYAMLDepth(node, 0, MaxYAMLDepth, MaxYAMLNodes, seen, &nodeCount)
+
+	// Should complete without infinite recursion due to cycle detection
+	if err != nil {
+		t.Errorf("unexpected error: %v", err)
+	}
+	// The seen map should contain multiple entries
+	if len(seen) < 2 {
+		t.Errorf("seen map has %d entries, expected at least 2", len(seen))
+	}
+}
+
+func TestListBuiltinPersonasSortedOrder(t *testing.T) {
+	names := ListBuiltinPersonas()
+	if len(names) < 2 {
+		t.Skip("need at least 2 personas to test ordering")
+	}
+
+	// Verify the list is sorted
+	for i := 1; i < len(names); i++ {
+		if names[i-1] > names[i] {
+			t.Errorf("ListBuiltinPersonas not sorted: %q > %q", names[i-1], names[i])
+		}
+	}
+}
+
+func TestYAMLUnknownFieldsRejected(t *testing.T) {
+	tests := []struct {
+		name    string
+		content string
+		wantErr string
+	}{
+		{
+			name: "unknown top-level field",
+			content: `name: test
+identity: test identity
+unknown_field: should fail
+`,
+			wantErr: "unknown_field",
+		},
+		{
+			name: "typo in field name",
+			content: `name: test
+identiy: typo should fail
+`,
+			wantErr: "identiy",
+		},
+		{
+			name: "unknown field in severity",
+			content: `name: test
+identity: test
+severity:
+  major: Major
+  minro: typo
+`,
+			wantErr: "minro",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dir := t.TempDir()
+			path := filepath.Join(dir, "unknown.yaml")
+			if err := os.WriteFile(path, []byte(tt.content), 0644); err != nil {
+				t.Fatalf("failed to write test file: %v", err)
+			}
+
+			_, err := LoadPersona(path)
+			if err == nil {
+				t.Errorf("expected error for unknown field %q, got nil", tt.wantErr)
+				return
+			}
+			if !strings.Contains(err.Error(), tt.wantErr) {
+				t.Errorf("error = %q, want containing %q", err.Error(), tt.wantErr)
+			}
+		})
+	}
+}
+
+func TestJSONUnknownFieldsRejected(t *testing.T) {
+	tests := []struct {
+		name    string
+		content string
+		wantErr string
+	}{
+		{
+			name: "unknown top-level field",
+			content: `{
+				"name": "test",
+				"identity": "test identity",
+				"unknown_field": "should fail"
+			}`,
+			wantErr: "unknown_field",
+		},
+		{
+			name: "typo in field name",
+			content: `{
+				"name": "test",
+				"identiy": "typo should fail"
+			}`,
+			wantErr: "identiy",
+		},
+		{
+			name: "unknown field in severity",
+			content: `{
+				"name": "test",
+				"identity": "test",
+				"severity": {
+					"major": "ok",
+					"miner": "typo"
+				}
+			}`,
+			wantErr: "miner",
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			dir := t.TempDir()
+			path := filepath.Join(dir, "test.json")
+			if err := os.WriteFile(path, []byte(tt.content), 0644); err != nil {
+				t.Fatalf("failed to write test file: %v", err)
+			}
+
+			_, err := LoadPersona(path)
+			if err == nil {
+				t.Fatal("expected error for unknown field, got nil")
+			}
+			if !strings.Contains(err.Error(), tt.wantErr) {
+				t.Errorf("error = %q, want to contain %q", err.Error(), tt.wantErr)
+			}
+		})
+	}
+}
+
+func TestLoadPersonaSymlink(t *testing.T) {
+	// Create a regular persona file
+	dir := t.TempDir()
+	realFile := filepath.Join(dir, "real.yaml")
+	content := `name: test
+identity: test identity
+`
+	if err := os.WriteFile(realFile, []byte(content), 0644); err != nil {
+		t.Fatalf("failed to write test file: %v", err)
+	}
+
+	// Create a symlink to it
+	symlink := filepath.Join(dir, "link.yaml")
+	if err := os.Symlink(realFile, symlink); err != nil {
+		t.Fatalf("failed to create symlink: %v", err)
+	}
+
+	// LoadPersona should work via symlink
+	p, err := LoadPersona(symlink)
+	if err != nil {
+		t.Fatalf("LoadPersona via symlink failed: %v", err)
+	}
+	if p.Name != "test" {
+		t.Errorf("Name = %q, want %q", p.Name, "test")
+	}
+}

review/personas/architect.yaml

@@ -0,0 +1,37 @@
+# Software Architect Persona
+# Focuses on design quality, patterns, and code organization
+
+name: architect
+display_name: Software Architect
+
+identity: |
+  You are a software architect reviewing code for design quality.
+
+  Your expertise:
+  - Design patterns and anti-patterns
+  - Code organization and module boundaries
+  - API design and contracts
+  - Testability and dependency injection
+  - Consistency with existing architecture
+  - Technical debt identification
+
+focus:
+  - Design pattern violations or misuse
+  - Module boundary violations (inappropriate coupling)
+  - API design issues (unclear contracts, leaky abstractions)
+  - Testability problems (hidden dependencies, god objects)
+  - Inconsistency with existing codebase patterns
+  - Unnecessary complexity or over-engineering
+  - Missing abstractions or premature abstraction
+
+ignore:
+  - Security vulnerabilities (security persona handles these)
+  - Performance micro-optimizations
+  - Code style and formatting
+  - Documentation typos
+  - Test implementation details
+
+severity:
+  major: "Architectural violations that will cause maintenance problems or make the codebase harder to evolve"
+  minor: "Design issues that reduce clarity or testability but don't block progress"
+  nit: "Minor pattern deviations or style preferences"

review/personas/docs.yaml

@@ -0,0 +1,36 @@
+# Documentation Reviewer Persona
+# Focuses on clarity, documentation quality, and self-documenting code
+
+name: docs
+display_name: Documentation Reviewer
+
+identity: |
+  You are a documentation specialist reviewing code for clarity and documentation quality.
+
+  Your expertise:
+  - API documentation and examples
+  - Code comments and their accuracy
+  - Error message clarity
+  - README and guide quality
+  - Naming clarity and self-documenting code
+
+focus:
+  - Missing or outdated documentation
+  - Unclear or misleading comments
+  - Poor error messages (cryptic, unhelpful, missing context)
+  - Confusing naming (functions, variables, types)
+  - Missing examples for complex APIs
+  - Inconsistent terminology
+  - Documentation that contradicts the code
+
+ignore:
+  - Security vulnerabilities
+  - Performance issues
+  - Design patterns
+  - Test coverage
+  - Code style (unless it affects readability)
+
+severity:
+  major: "Documentation that actively misleads or missing docs for critical functionality"
+  minor: "Unclear documentation or poor error messages that will confuse users"
+  nit: "Minor clarity improvements or typo fixes"

review/personas/security.yaml

@@ -0,0 +1,37 @@
+# Security Specialist Persona
+# Focuses on vulnerabilities, auth issues, and security best practices
+
+name: security
+display_name: Security Specialist
+
+identity: |
+  You are a security specialist reviewing code for vulnerabilities.
+
+  Your expertise:
+  - OWASP Top 10 vulnerabilities
+  - Injection attacks (SQL, command, path traversal, template)
+  - Authentication and authorization patterns
+  - Secrets management and exposure risks
+  - Race conditions with security implications
+  - Event sourcing attack vectors (replay attacks, event injection)
+
+focus:
+  - Injection attacks (SQL, command, path traversal, template injection)
+  - Authentication and authorization gaps or bypasses
+  - Secrets exposure (hardcoded credentials, tokens in logs, config leaks)
+  - Input validation failures (unsanitized input, unsafe deserialization)
+  - Race conditions that could be exploited
+  - Cryptographic weaknesses (weak algorithms, improper key handling)
+  - Information disclosure through error messages or logs
+
+ignore:
+  - Code style and naming conventions
+  - Performance optimizations (unless security-related)
+  - Documentation quality
+  - General code quality or readability
+  - Test coverage
+
+severity:
+  major: "Exploitable vulnerabilities: auth bypass, injection, data exfiltration, privilege escalation, RCE"
+  minor: "Defense-in-depth issues: missing rate limiting, verbose errors, weak input validation"
+  nit: "Theoretical risks with low exploitability or impact"

review/prompt.go

@@ -7,6 +7,28 @@ import (
 	"strings"
 )
 
+// outputSchemaJSON is the shared JSON output format specification used by both
+// the generic reviewer and persona-based reviewers.
+const outputSchemaJSON = `{
+  "verdict": "APPROVE" or "REQUEST_CHANGES",
+  "summary": "Brief overall assessment (1-3 sentences)",
+  "findings": [
+    {
+      "severity": "MAJOR" or "MINOR" or "NIT",
+      "file": "path/to/file",
+      "line": <line number from the diff>,
+      "finding": "Description of the issue"
+    }
+  ],
+  "recommendation": "Full recommendation text explaining your verdict"
+}`
+
+// verdictRules is the shared verdict determination rules.
+const verdictRules = `Rules:
+- If there are any MAJOR findings → verdict must be REQUEST_CHANGES
+- If there are no MAJOR findings → verdict should be APPROVE
+- If CI has failed → verdict must be REQUEST_CHANGES with a finding noting the CI failure`
+
 // BuildSystemBase returns the core system prompt instructions without
 // patterns or conventions. Used by the budget package to separate
 // trimmable from non-trimmable content.
@@ -23,24 +45,10 @@ func BuildSystemBase() string {
 	sb.WriteString("2. Consider the CI status — if CI has failed, that is an automatic REQUEST_CHANGES regardless of code quality.\n")
 	sb.WriteString("3. Output your review as structured JSON (and ONLY JSON, no markdown fences or other text).\n\n")
 	sb.WriteString("Output format:\n")
-	sb.WriteString("{\n")
-	sb.WriteString("  \"verdict\": \"APPROVE\" or \"REQUEST_CHANGES\",\n")
-	sb.WriteString("  \"summary\": \"Brief overall assessment (1-3 sentences)\",\n")
-	sb.WriteString("  \"findings\": [\n")
-	sb.WriteString("    {\n")
-	sb.WriteString("      \"severity\": \"MAJOR\" or \"MINOR\" or \"NIT\",\n")
-	sb.WriteString("      \"file\": \"path/to/file\",\n")
-	sb.WriteString("      \"line\": <line number from the diff>,\n")
-	sb.WriteString("      \"finding\": \"Description of the issue\"\n")
-	sb.WriteString("    }\n")
-	sb.WriteString("  ],\n")
-	sb.WriteString("  \"recommendation\": \"Full recommendation text explaining your verdict\"\n")
-	sb.WriteString("}\n\n")
-	sb.WriteString("Rules:\n")
-	sb.WriteString("- If there are any MAJOR findings → verdict must be REQUEST_CHANGES\n")
-	sb.WriteString("- If there are no MAJOR findings → verdict should be APPROVE\n")
-	sb.WriteString("- If CI has failed → verdict must be REQUEST_CHANGES with a finding noting the CI failure\n")
-	sb.WriteString("- Be thorough but fair. Don't nitpick style unless it impacts readability significantly.\n")
+	sb.WriteString(outputSchemaJSON)
+	sb.WriteString("\n\n")
+	sb.WriteString(verdictRules)
+	sb.WriteString("\n- Be thorough but fair. Don't nitpick style unless it impacts readability significantly.\n")
 	sb.WriteString("- Line numbers should reference the new file line numbers from the diff headers.\n")
 	sb.WriteString("- If the diff is empty or trivial (only formatting/whitespace), APPROVE with no findings.\n")
 

review/remote_persona.go

@@ -0,0 +1,171 @@
+package review
+
+import (
+	"context"
+	"fmt"
+	"log/slog"
+	"sort"
+	"strings"
+)
+
+// PersonaFetcher abstracts fetching files from a remote repository.
+// This allows persona loading to work with any Git host API.
+type PersonaFetcher interface {
+	// ListContents returns file/directory entries at a path.
+	// Returns an error if the path doesn't exist or isn't accessible.
+	ListContents(ctx context.Context, owner, repo, path string) ([]ContentEntry, error)
+
+	// GetFileContent returns the raw content of a file from the default branch.
+	GetFileContent(ctx context.Context, owner, repo, filepath string) (string, error)
+}
+
+// ContentEntry represents a file or directory entry.
+type ContentEntry struct {
+	Name string // filename or directory name
+	Path string // full path from repo root
+	Type string // "file" or "dir"
+}
+
+// DefaultPersonasPath is the conventional location for repo-specific personas.
+const DefaultPersonasPath = ".review-bot/personas"
+
+// LoadRemotePersonas fetches personas from a remote repository's .review-bot/personas/ directory.
+// Returns a map of persona name to Persona. If the directory doesn't exist or is empty,
+// returns an empty map with no error (graceful fallback to built-in personas).
+//
+// Files larger than MaxPersonaFileSize are logged and skipped.
+// Invalid YAML files are logged and skipped (partial success model).
+// Only .yaml and .yml files are processed; other files are ignored.
+func LoadRemotePersonas(ctx context.Context, fetcher PersonaFetcher, owner, repo string) (map[string]*Persona, error) {
+	return LoadRemotePersonasFromPath(ctx, fetcher, owner, repo, DefaultPersonasPath)
+}
+
+// LoadRemotePersonasFromPath loads personas from a custom path in a remote repository.
+// It behaves the same as LoadRemotePersonas but allows specifying a path other than
+// the default .review-bot/personas directory.
+func LoadRemotePersonasFromPath(ctx context.Context, fetcher PersonaFetcher, owner, repo, path string) (map[string]*Persona, error) {
+	entries, err := fetcher.ListContents(ctx, owner, repo, path)
+	if err != nil {
+		// 404 is expected when repo doesn't have personas — return empty, not error
+		if isNotFoundError(err) {
+			slog.Debug("no remote personas directory found", "repo", fmt.Sprintf("%s/%s", owner, repo), "path", path)
+			return map[string]*Persona{}, nil
+		}
+		return nil, fmt.Errorf("list remote personas: %w", err)
+	}
+
+	// Cap the number of files to process to prevent resource exhaustion
+	// from repos with thousands of small files.
+	const maxPersonaFiles = 50
+
+	result := make(map[string]*Persona)
+	processed := 0
+	for _, entry := range entries {
+		if processed >= maxPersonaFiles {
+			slog.Warn("persona file limit reached", "limit", maxPersonaFiles, "repo", fmt.Sprintf("%s/%s", owner, repo))
+			break
+		}
+		if ctx.Err() != nil {
+			return nil, ctx.Err()
+		}
+
+		// Skip directories and non-YAML files
+		if entry.Type != "file" {
+			continue
+		}
+		if !isYAMLFile(entry.Name) {
+			continue
+		}
+
+		content, err := fetcher.GetFileContent(ctx, owner, repo, entry.Path)
+		if err != nil {
+			slog.Warn("could not fetch remote persona file", "file", entry.Path, "error", err)
+			continue
+		}
+
+		// Check size before parsing (defense in depth)
+		if len(content) > MaxPersonaFileSize {
+			slog.Warn("remote persona file exceeds size limit", "file", entry.Path, "size", len(content), "limit", MaxPersonaFileSize)
+			continue
+		}
+
+		// YAML parsing uses parsePersona which has defenses against YAML DoS attacks:
+		// - MaxPersonaFileSize (above) caps raw input size before any parsing
+		// - maxPersonaFiles (above) limits the number of files processed per repo
+		// - unmarshalYAMLWithDepthLimit enforces MaxYAMLDepth to prevent stack exhaustion
+		// - checkYAMLDepth tracks node counts (MaxYAMLNodes) against "billion laughs" expansion
+		// - Alias cycles are detected and capped by seen-node tracking
+		// See persona.go for the implementation details.
+		persona, err := parsePersona([]byte(content), entry.Path)
+		if err != nil {
+			slog.Warn("could not parse remote persona file", "file", entry.Path, "error", err)
+			continue
+		}
+
+		result[persona.Name] = persona
+		processed++
+		slog.Debug("loaded remote persona", "name", persona.Name, "file", entry.Path)
+	}
+
+	return result, nil
+}
+
+// MergePersonas combines remote and built-in personas.
+// Remote personas take precedence on name collision.
+// Returns the merged map and a list of persona names in sorted order.
+func MergePersonas(remote, builtin map[string]*Persona) (map[string]*Persona, []string) {
+	merged := make(map[string]*Persona)
+
+	// Add built-in first
+	for name, p := range builtin {
+		merged[name] = p
+	}
+
+	// Remote overrides built-in on collision
+	for name, p := range remote {
+		if _, exists := merged[name]; exists {
+			slog.Debug("remote persona overrides built-in", "name", name)
+		}
+		merged[name] = p
+	}
+
+	// Collect sorted names
+	names := make([]string, 0, len(merged))
+	for name := range merged {
+		names = append(names, name)
+	}
+	sort.Strings(names)
+
+	return merged, names
+}
+
+// LoadAllBuiltinPersonas loads all built-in personas into a map.
+func LoadAllBuiltinPersonas() map[string]*Persona {
+	result := make(map[string]*Persona)
+	for _, name := range ListBuiltinPersonas() {
+		p, err := LoadBuiltinPersona(name)
+		if err != nil {
+			slog.Warn("could not load built-in persona", "name", name, "error", err)
+			continue
+		}
+		result[name] = p
+	}
+	return result
+}
+
+// isYAMLFile returns true if the filename has a YAML extension.
+func isYAMLFile(name string) bool {
+	lower := strings.ToLower(name)
+	return strings.HasSuffix(lower, ".yaml") || strings.HasSuffix(lower, ".yml")
+}
+
+// isNotFoundError checks if an error indicates a 404 response.
+// This is a simple string check to avoid importing the gitea package
+// (which would create a circular dependency).
+func isNotFoundError(err error) bool {
+	if err == nil {
+		return false
+	}
+	errStr := err.Error()
+	return strings.Contains(errStr, "HTTP 404")
+}

review/remote_persona_test.go

@@ -0,0 +1,394 @@
+package review
+
+import (
+	"context"
+	"errors"
+	"testing"
+)
+
+// mockFetcher implements PersonaFetcher for testing.
+type mockFetcher struct {
+	contents     map[string][]ContentEntry // path -> entries
+	files        map[string]string         // path -> content
+	listErr      error                     // error to return from ListContents
+	getFileErr   map[string]error          // path -> error for GetFileContent
+	listNotFound bool                      // return 404-style error
+}
+
+func newMockFetcher() *mockFetcher {
+	return &mockFetcher{
+		contents:   make(map[string][]ContentEntry),
+		files:      make(map[string]string),
+		getFileErr: make(map[string]error),
+	}
+}
+
+func (m *mockFetcher) ListContents(ctx context.Context, owner, repo, path string) ([]ContentEntry, error) {
+	if m.listNotFound {
+		return nil, errors.New("HTTP 404: not found")
+	}
+	if m.listErr != nil {
+		return nil, m.listErr
+	}
+	entries, ok := m.contents[path]
+	if !ok {
+		return nil, errors.New("HTTP 404: not found")
+	}
+	return entries, nil
+}
+
+func (m *mockFetcher) GetFileContent(ctx context.Context, owner, repo, filepath string) (string, error) {
+	if err, ok := m.getFileErr[filepath]; ok {
+		return "", err
+	}
+	content, ok := m.files[filepath]
+	if !ok {
+		return "", errors.New("HTTP 404: file not found")
+	}
+	return content, nil
+}
+
+func TestLoadRemotePersonas_NoDirectory(t *testing.T) {
+	fetcher := newMockFetcher()
+	fetcher.listNotFound = true
+
+	result, err := LoadRemotePersonas(context.Background(), fetcher, "owner", "repo")
+	if err != nil {
+		t.Fatalf("expected no error for missing directory, got: %v", err)
+	}
+	if len(result) != 0 {
+		t.Errorf("expected empty map, got %d personas", len(result))
+	}
+}
+
+func TestLoadRemotePersonas_EmptyDirectory(t *testing.T) {
+	fetcher := newMockFetcher()
+	fetcher.contents[DefaultPersonasPath] = []ContentEntry{}
+
+	result, err := LoadRemotePersonas(context.Background(), fetcher, "owner", "repo")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(result) != 0 {
+		t.Errorf("expected empty map, got %d personas", len(result))
+	}
+}
+
+func TestLoadRemotePersonas_SinglePersona(t *testing.T) {
+	fetcher := newMockFetcher()
+	fetcher.contents[DefaultPersonasPath] = []ContentEntry{
+		{Name: "trading.yaml", Path: ".review-bot/personas/trading.yaml", Type: "file"},
+	}
+	fetcher.files[".review-bot/personas/trading.yaml"] = `
+name: trading
+display_name: Trading Expert
+identity: You are a trading systems expert.
+focus:
+  - order execution
+  - market data
+`
+
+	result, err := LoadRemotePersonas(context.Background(), fetcher, "owner", "repo")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(result) != 1 {
+		t.Fatalf("expected 1 persona, got %d", len(result))
+	}
+	if result["trading"] == nil {
+		t.Fatal("expected 'trading' persona")
+	}
+	if result["trading"].DisplayName != "Trading Expert" {
+		t.Errorf("expected display name 'Trading Expert', got %q", result["trading"].DisplayName)
+	}
+}
+
+func TestLoadRemotePersonas_MultiplePersonas(t *testing.T) {
+	fetcher := newMockFetcher()
+	fetcher.contents[DefaultPersonasPath] = []ContentEntry{
+		{Name: "one.yaml", Path: ".review-bot/personas/one.yaml", Type: "file"},
+		{Name: "two.yml", Path: ".review-bot/personas/two.yml", Type: "file"},
+	}
+	fetcher.files[".review-bot/personas/one.yaml"] = `
+name: one
+identity: First persona.
+`
+	fetcher.files[".review-bot/personas/two.yml"] = `
+name: two
+identity: Second persona.
+`
+
+	result, err := LoadRemotePersonas(context.Background(), fetcher, "owner", "repo")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(result) != 2 {
+		t.Fatalf("expected 2 personas, got %d", len(result))
+	}
+	if result["one"] == nil || result["two"] == nil {
+		t.Error("expected both personas to be loaded")
+	}
+}
+
+func TestLoadRemotePersonas_SkipsNonYAML(t *testing.T) {
+	fetcher := newMockFetcher()
+	fetcher.contents[DefaultPersonasPath] = []ContentEntry{
+		{Name: "valid.yaml", Path: ".review-bot/personas/valid.yaml", Type: "file"},
+		{Name: "readme.md", Path: ".review-bot/personas/readme.md", Type: "file"},
+		{Name: "config.json", Path: ".review-bot/personas/config.json", Type: "file"},
+	}
+	fetcher.files[".review-bot/personas/valid.yaml"] = `
+name: valid
+identity: Valid persona.
+`
+
+	result, err := LoadRemotePersonas(context.Background(), fetcher, "owner", "repo")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(result) != 1 {
+		t.Fatalf("expected 1 persona (skipping non-YAML), got %d", len(result))
+	}
+}
+
+func TestLoadRemotePersonas_SkipsDirectories(t *testing.T) {
+	fetcher := newMockFetcher()
+	fetcher.contents[DefaultPersonasPath] = []ContentEntry{
+		{Name: "valid.yaml", Path: ".review-bot/personas/valid.yaml", Type: "file"},
+		{Name: "subdir", Path: ".review-bot/personas/subdir", Type: "dir"},
+	}
+	fetcher.files[".review-bot/personas/valid.yaml"] = `
+name: valid
+identity: Valid persona.
+`
+
+	result, err := LoadRemotePersonas(context.Background(), fetcher, "owner", "repo")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(result) != 1 {
+		t.Fatalf("expected 1 persona (skipping dir), got %d", len(result))
+	}
+}
+
+func TestLoadRemotePersonas_SkipsInvalidYAML(t *testing.T) {
+	fetcher := newMockFetcher()
+	fetcher.contents[DefaultPersonasPath] = []ContentEntry{
+		{Name: "valid.yaml", Path: ".review-bot/personas/valid.yaml", Type: "file"},
+		{Name: "invalid.yaml", Path: ".review-bot/personas/invalid.yaml", Type: "file"},
+	}
+	fetcher.files[".review-bot/personas/valid.yaml"] = `
+name: valid
+identity: Valid persona.
+`
+	fetcher.files[".review-bot/personas/invalid.yaml"] = `
+this is not valid yaml: [unclosed bracket
+`
+
+	result, err := LoadRemotePersonas(context.Background(), fetcher, "owner", "repo")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(result) != 1 {
+		t.Fatalf("expected 1 persona (skipping invalid), got %d", len(result))
+	}
+	if result["valid"] == nil {
+		t.Error("expected valid persona to be loaded")
+	}
+}
+
+func TestLoadRemotePersonas_SkipsOversizedFiles(t *testing.T) {
+	fetcher := newMockFetcher()
+	fetcher.contents[DefaultPersonasPath] = []ContentEntry{
+		{Name: "huge.yaml", Path: ".review-bot/personas/huge.yaml", Type: "file"},
+	}
+	// Create content larger than MaxPersonaFileSize (64KB)
+	fetcher.files[".review-bot/personas/huge.yaml"] = `
+name: huge
+identity: ` + string(make([]byte, MaxPersonaFileSize+1000))
+
+	result, err := LoadRemotePersonas(context.Background(), fetcher, "owner", "repo")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(result) != 0 {
+		t.Errorf("expected 0 personas (oversized file skipped), got %d", len(result))
+	}
+}
+
+func TestLoadRemotePersonas_SkipsFetchErrors(t *testing.T) {
+	fetcher := newMockFetcher()
+	fetcher.contents[DefaultPersonasPath] = []ContentEntry{
+		{Name: "valid.yaml", Path: ".review-bot/personas/valid.yaml", Type: "file"},
+		{Name: "error.yaml", Path: ".review-bot/personas/error.yaml", Type: "file"},
+	}
+	fetcher.files[".review-bot/personas/valid.yaml"] = `
+name: valid
+identity: Valid persona.
+`
+	fetcher.getFileErr[".review-bot/personas/error.yaml"] = errors.New("network error")
+
+	result, err := LoadRemotePersonas(context.Background(), fetcher, "owner", "repo")
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(result) != 1 {
+		t.Fatalf("expected 1 persona (skipping error), got %d", len(result))
+	}
+}
+
+func TestLoadRemotePersonas_ListContentsError(t *testing.T) {
+	fetcher := newMockFetcher()
+	fetcher.listErr = errors.New("server error")
+
+	_, err := LoadRemotePersonas(context.Background(), fetcher, "owner", "repo")
+	if err == nil {
+		t.Fatal("expected error for list contents failure")
+	}
+}
+
+func TestLoadRemotePersonas_ContextCancellation(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel() // Cancel immediately
+
+	fetcher := newMockFetcher()
+	fetcher.contents[DefaultPersonasPath] = []ContentEntry{
+		{Name: "one.yaml", Path: ".review-bot/personas/one.yaml", Type: "file"},
+	}
+	fetcher.files[".review-bot/personas/one.yaml"] = `
+name: one
+identity: One.
+`
+
+	_, err := LoadRemotePersonas(ctx, fetcher, "owner", "repo")
+	if err == nil {
+		t.Fatal("expected context cancellation error")
+	}
+}
+
+func TestMergePersonas_NoOverlap(t *testing.T) {
+	remote := map[string]*Persona{
+		"trading": {Name: "trading", Identity: "Trading expert."},
+	}
+	builtin := map[string]*Persona{
+		"security": {Name: "security", Identity: "Security expert."},
+	}
+
+	merged, names := MergePersonas(remote, builtin)
+
+	if len(merged) != 2 {
+		t.Fatalf("expected 2 personas, got %d", len(merged))
+	}
+	if len(names) != 2 {
+		t.Fatalf("expected 2 names, got %d", len(names))
+	}
+	// Names should be sorted
+	if names[0] != "security" || names[1] != "trading" {
+		t.Errorf("expected sorted names [security, trading], got %v", names)
+	}
+}
+
+func TestMergePersonas_RemoteOverridesBuiltin(t *testing.T) {
+	remote := map[string]*Persona{
+		"security": {Name: "security", Identity: "Custom security expert."},
+	}
+	builtin := map[string]*Persona{
+		"security": {Name: "security", Identity: "Default security expert."},
+	}
+
+	merged, _ := MergePersonas(remote, builtin)
+
+	if merged["security"].Identity != "Custom security expert." {
+		t.Errorf("expected remote to override builtin, got identity: %q", merged["security"].Identity)
+	}
+}
+
+func TestMergePersonas_EmptyRemote(t *testing.T) {
+	remote := map[string]*Persona{}
+	builtin := map[string]*Persona{
+		"security": {Name: "security", Identity: "Security."},
+	}
+
+	merged, names := MergePersonas(remote, builtin)
+
+	if len(merged) != 1 {
+		t.Fatalf("expected 1 persona, got %d", len(merged))
+	}
+	if names[0] != "security" {
+		t.Errorf("expected 'security', got %q", names[0])
+	}
+}
+
+func TestMergePersonas_EmptyBuiltin(t *testing.T) {
+	remote := map[string]*Persona{
+		"trading": {Name: "trading", Identity: "Trading."},
+	}
+	builtin := map[string]*Persona{}
+
+	merged, names := MergePersonas(remote, builtin)
+
+	if len(merged) != 1 {
+		t.Fatalf("expected 1 persona, got %d", len(merged))
+	}
+	if names[0] != "trading" {
+		t.Errorf("expected 'trading', got %q", names[0])
+	}
+}
+
+func TestLoadAllBuiltinPersonas(t *testing.T) {
+	personas := LoadAllBuiltinPersonas()
+
+	// Should load at least the known built-in personas
+	expected := []string{"architect", "docs", "security"}
+	for _, name := range expected {
+		if personas[name] == nil {
+			t.Errorf("expected built-in persona %q to be loaded", name)
+		}
+	}
+}
+
+func TestIsYAMLFile(t *testing.T) {
+	tests := []struct {
+		name     string
+		expected bool
+	}{
+		{"test.yaml", true},
+		{"test.yml", true},
+		{"test.YAML", true},
+		{"test.YML", true},
+		{"test.json", false},
+		{"test.md", false},
+		{"yaml", false},
+		{"", false},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := isYAMLFile(tc.name); got != tc.expected {
+				t.Errorf("isYAMLFile(%q) = %v, want %v", tc.name, got, tc.expected)
+			}
+		})
+	}
+}
+
+func TestIsNotFoundError(t *testing.T) {
+	tests := []struct {
+		name     string
+		err      error
+		expected bool
+	}{
+		{"nil error", nil, false},
+		{"HTTP 404", errors.New("HTTP 404: not found"), true},
+		{"not found text", errors.New("path not found"), false},
+		{"server error", errors.New("server error"), false},
+		{"HTTP 500", errors.New("HTTP 500: internal error"), false},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := isNotFoundError(tc.err); got != tc.expected {
+				t.Errorf("isNotFoundError(%v) = %v, want %v", tc.err, got, tc.expected)
+			}
+		})
+	}
+}

scripts/check-deps.sh

@@ -0,0 +1,127 @@
+#!/usr/bin/env bash
+# check-deps.sh - Enforces the strict dependency allowlist from CONVENTIONS.md
+# Exit 1 if any unapproved import is found.
+#
+# Requires: Bash 4+ (for associative arrays), Go toolchain
+# 
+# The allowlist is parsed from CONVENTIONS.md to maintain a single source of truth.
+# Enforces Scope column: "test only" packages cannot appear in non-test code.
+
+set -euo pipefail
+
+# Check bash version
+if ((BASH_VERSINFO[0] < 4)); then
+    echo "❌ Bash 4+ required (found ${BASH_VERSION})"
+    echo "   On macOS: brew install bash"
+    exit 1
+fi
+
+CONVENTIONS_FILE="${1:-CONVENTIONS.md}"
+
+if [ ! -f "$CONVENTIONS_FILE" ]; then
+    echo "❌ CONVENTIONS.md not found"
+    exit 1
+fi
+
+# Parse approved packages from CONVENTIONS.md table using awk (POSIX-compatible)
+# Format: | `package` | use case | scope |
+declare -A ALLOWED_PROD=()
+declare -A ALLOWED_TEST=()
+
+while IFS= read -r line; do
+    # Use awk to extract package and scope from table row
+    pkg=$(echo "$line" | awk -F'|' '{gsub(/^[[:space:]]*`|`[[:space:]]*$/, "", $2); print $2}')
+    scope=$(echo "$line" | awk -F'|' '{gsub(/^[[:space:]]+|[[:space:]]+$/, "", $4); print tolower($4)}')
+    
+    if [ -n "$pkg" ] && [ "$pkg" != "Package" ] && [[ "$pkg" =~ ^[a-zA-Z] ]]; then
+        if [[ "$scope" == *"test"* ]]; then
+            ALLOWED_TEST["$pkg"]=1
+        else
+            ALLOWED_PROD["$pkg"]=1
+        fi
+    fi
+done < <(grep '| `' "$CONVENTIONS_FILE" 2>/dev/null || true)
+
+ALL_ALLOWED=("${!ALLOWED_PROD[@]}" "${!ALLOWED_TEST[@]}")
+
+if [ ${#ALL_ALLOWED[@]} -eq 0 ]; then
+    echo "⚠️  No approved packages found in $CONVENTIONS_FILE"
+    echo "   (This is fine if you want stdlib-only)"
+fi
+
+# Helper: check if import matches any package in an associative array (literal prefix, no glob)
+matches_allowlist() {
+    local import="$1"
+    shift
+    local -n allowlist=$1
+    
+    for allowed in "${!allowlist[@]}"; do
+        # Exact match
+        if [ "$import" = "$allowed" ]; then
+            return 0
+        fi
+        # Literal prefix match for subpackages: must match "pkg/" exactly
+        if [ "${import#"$allowed/"}" != "$import" ]; then
+            return 0
+        fi
+    done
+    return 1
+}
+
+# Get direct module dependencies from go.mod
+DIRECT_IMPORTS=$(go list -m -f '{{if and (not .Indirect) (not .Main)}}{{.Path}}{{end}}' all 2>&1) || {
+    echo "❌ Failed to list dependencies: $DIRECT_IMPORTS"
+    exit 1
+}
+DIRECT_IMPORTS=$(echo "$DIRECT_IMPORTS" | grep -v '^$' || true)
+
+if [ -z "$DIRECT_IMPORTS" ]; then
+    echo "✅ No external dependencies"
+    exit 0
+fi
+
+# Check ALL direct dependencies are in some allowlist
+VIOLATIONS=""
+while IFS= read -r import; do
+    [ -z "$import" ] && continue
+    
+    if ! matches_allowlist "$import" ALLOWED_PROD && ! matches_allowlist "$import" ALLOWED_TEST; then
+        VIOLATIONS="${VIOLATIONS}  - ${import} (not in allowlist)"$'\n'
+    fi
+done <<< "$DIRECT_IMPORTS"
+
+if [ -n "$VIOLATIONS" ]; then
+    echo "❌ UNAPPROVED DEPENDENCIES DETECTED"
+    echo ""
+    echo "The following imports are not in the allowlist:"
+    printf "%s" "$VIOLATIONS"
+    echo ""
+    echo "To add a dependency, update CONVENTIONS.md (requires Aaron's approval)"
+    exit 1
+fi
+
+# Enforce Scope: test-only packages must not appear in non-test code
+# Get imports used by non-test code only (go list -deps without -test excludes test deps)
+PROD_IMPORTS=$(go list -deps -f '{{if not .Standard}}{{.ImportPath}}{{end}}' ./... 2>/dev/null || true)
+
+TEST_ONLY_IN_PROD=""
+for test_pkg in "${!ALLOWED_TEST[@]}"; do
+    # Use word-boundary matching: exact match or followed by /
+    if echo "$PROD_IMPORTS" | grep -qE "^${test_pkg}(/|\$|$)"; then
+        TEST_ONLY_IN_PROD="${TEST_ONLY_IN_PROD}  - ${test_pkg} (marked 'test only' but used in production code)"$'\n'
+    fi
+done
+
+if [ -n "$TEST_ONLY_IN_PROD" ]; then
+    echo "❌ TEST-ONLY DEPENDENCIES IN PRODUCTION CODE"
+    echo ""
+    printf "%s" "$TEST_ONLY_IN_PROD"
+    echo ""
+    echo "These packages are marked 'test only' in CONVENTIONS.md"
+    echo "and must only be imported from *_test.go files."
+    exit 1
+fi
+
+echo "✅ All dependencies are approved"
+echo "   Direct module deps: $(echo "$DIRECT_IMPORTS" | wc -l | tr -d ' ')"
+echo "   Production allowlist: ${#ALLOWED_PROD[@]}, Test-only allowlist: ${#ALLOWED_TEST[@]}"