feat: add role title as H1 header for visual differentiation

When reviewer-name is set, prepend "# Security Review" / "# Sonnet Review" etc. as a top-level header. Makes it immediately obvious which role each review represents in the Gitea UI, especially when multiple reviews come from the same bot account.
fix: symlink traversal + worst-wins pre-check + user scoping
2026-05-01 21:35:41 -07:00 · 2026-05-01 21:31:17 -07:00 · 2026-05-01 21:12:34 -07:00 · 2026-05-01 21:05:18 -07:00 · 2026-05-01 20:59:34 -07:00 · 2026-05-01 20:55:09 -07:00
12 changed files with 1045 additions and 87 deletions
@@ -34,6 +34,10 @@ inputs:
  llm-model:
    description: 'LLM model name'
    required: true
+  llm-provider:
+    description: 'LLM API provider: openai or anthropic (default openai)'
+    required: false
+    default: 'openai' 
  conventions-file:
    description: 'Path to conventions file in the repo (e.g. CLAUDE.md)'
    required: false
@@ -62,6 +66,14 @@ inputs:
    description: 'Print review to stdout instead of posting'
    required: false
    default: 'false'
+  update-existing:
+    description: 'Delete previous review from same bot after posting new one. Accepts: true/1/yes or false/0/no (default true)'
+    required: false
+    default: 'true'
+  system-prompt-file:
+    description: 'Local file with additional system prompt instructions (e.g. security review focus)'
+    required: false
+    default: ''

 runs:
  using: 'composite'
@@ -140,6 +152,9 @@ runs:
        PATTERNS_FILES: ${{ inputs.patterns-files }}
        LLM_TEMPERATURE: ${{ inputs.temperature }}
        LLM_TIMEOUT: ${{ inputs.timeout }}
+        LLM_PROVIDER: ${{ inputs.llm-provider }}
+        UPDATE_EXISTING: ${{ inputs.update-existing }}
+        SYSTEM_PROMPT_FILE: ${{ inputs.system-prompt-file }}
      run: |
        ARGS=""
        if [ "${{ inputs.dry-run }}" = "true" ]; then
@@ -32,6 +32,10 @@ jobs:
          - name: gpt
            token_secret: GPT_REVIEW_TOKEN
            model: gpt-4.1
+          - name: security
+            token_secret: SONNET_REVIEW_TOKEN
+            model: gpt-5
+            system_prompt_file: SECURITY_REVIEW.md
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-go@v5
@@ -44,6 +48,7 @@ jobs:
          GITEA_REPO: ${{ github.repository }}
          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_MODEL: ${{ matrix.model }}
@@ -51,4 +56,5 @@ jobs:
          PATTERNS_REPO: "rodin/go-patterns"
          PATTERNS_FILES: "README.md,patterns/"
          LLM_TIMEOUT: "600"
+          SYSTEM_PROMPT_FILE: ${{ matrix.system_prompt_file }}
        run: ./review-bot
@@ -1,17 +1,242 @@
 # review-bot

-Automated code review bot for Gitea. Fetches a pull request diff, sends it to an LLM for analysis, and posts a structured review back to the PR.
+AI-powered code review bot for Gitea pull requests. Fetches diff + context, sends to an LLM, and posts a structured review (APPROVE / REQUEST_CHANGES) back to the PR.

 ## Features

- Fetches PR metadata, diff, and CI status from Gitea API
- Sends context-rich prompts to any OpenAI-compatible LLM
- Parses structured JSON review responses
- Posts formatted reviews (APPROVE / REQUEST_CHANGES) back to Gitea
- Supports custom coding conventions via repo files
- Zero external dependencies — Go stdlib only
+- **Multi-provider**: OpenAI-compatible and Anthropic Messages API
+- **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

-## Usage
+## Quick Start: Composite Action
+
+The easiest way to use review-bot in your Gitea CI:
+
+```yaml
+# .gitea/workflows/review.yml
+name: Review
+on:
+  pull_request:
+    types: [opened, synchronize]
+
+jobs:
+  review:
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v4
+      - uses: https://gitea.weiker.me/rodin/review-bot/.gitea/actions/review@v0.1.0
+        with:
+          reviewer-token: ${{ secrets.REVIEW_TOKEN }}
+          reviewer-name: code-review
+          llm-base-url: ${{ secrets.LLM_BASE_URL }}
+          llm-api-key: ${{ secrets.LLM_API_KEY }}
+          llm-model: gpt-4.1
+```
+
+That's it. Every PR gets an automated review.
+
+## Examples
+
+### Single reviewer with conventions
+
+```yaml
+jobs:
+  review:
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v4
+      - uses: https://gitea.weiker.me/rodin/review-bot/.gitea/actions/review@v0.1.0
+        with:
+          reviewer-token: ${{ secrets.REVIEW_TOKEN }}
+          reviewer-name: reviewer
+          llm-base-url: ${{ secrets.LLM_BASE_URL }}
+          llm-api-key: ${{ secrets.LLM_API_KEY }}
+          llm-model: gpt-4.1
+          conventions-file: CONVENTIONS.md
+          timeout: '600'
+```
+
+### Two reviewers with different models (diversity of opinion)
+
+```yaml
+jobs:
+  review:
+    runs-on: ubuntu-24.04
+    strategy:
+      matrix:
+        include:
+          - name: gpt
+            model: gpt-4.1
+            token_secret: GPT_REVIEW_TOKEN
+          - name: claude
+            model: claude-sonnet-4-20250514
+            token_secret: CLAUDE_REVIEW_TOKEN
+            provider: anthropic
+    steps:
+      - uses: actions/checkout@v4
+      - uses: https://gitea.weiker.me/rodin/review-bot/.gitea/actions/review@v0.1.0
+        with:
+          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-model: ${{ matrix.model }}
+          llm-provider: ${{ matrix.provider }}
+          conventions-file: CONVENTIONS.md
+```
+
+Each reviewer posts independently and only cleans up its own stale reviews.
+
+### Multiple review types from a single bot account
+
+Use the same Gitea token but different `reviewer-name` values to run specialized reviews without needing multiple bot accounts:
+
+```yaml
+jobs:
+  review:
+    runs-on: ubuntu-24.04
+    strategy:
+      matrix:
+        include:
+          - name: code-quality
+            model: gpt-4.1
+          - name: security
+            model: gpt-4.1
+            system_prompt_file: .review/SECURITY.md
+          - name: performance
+            model: gpt-4.1
+            system_prompt_file: .review/PERFORMANCE.md
+    steps:
+      - uses: actions/checkout@v4
+      - uses: https://gitea.weiker.me/rodin/review-bot/.gitea/actions/review@v0.1.0
+        with:
+          reviewer-token: ${{ secrets.REVIEW_TOKEN }}
+          reviewer-name: ${{ matrix.name }}
+          llm-base-url: ${{ secrets.LLM_BASE_URL }}
+          llm-api-key: ${{ secrets.LLM_API_KEY }}
+          llm-model: ${{ matrix.model }}
+          system-prompt-file: ${{ matrix.system_prompt_file }}
+```
+
+The sentinel `<!-- review-bot:security -->` ensures the security review only replaces previous security reviews, never the code-quality or performance reviews.
+
+### With language patterns from another repo
+
+```yaml
+- uses: https://gitea.weiker.me/rodin/review-bot/.gitea/actions/review@v0.1.0
+  with:
+    reviewer-token: ${{ secrets.REVIEW_TOKEN }}
+    reviewer-name: reviewer
+    llm-base-url: ${{ secrets.LLM_BASE_URL }}
+    llm-api-key: ${{ secrets.LLM_API_KEY }}
+    llm-model: gpt-4.1
+    conventions-file: CLAUDE.md
+    patterns-repo: rodin/go-patterns,rodin/kubernetes-conventions
+    patterns-files: "README.md,patterns/"
+```
+
+Pattern repos are fetched at review time. The reviewer uses them as criteria for idiomatic code.
+
+### Dry run (test without posting)
+
+```yaml
+- uses: https://gitea.weiker.me/rodin/review-bot/.gitea/actions/review@v0.1.0
+  with:
+    reviewer-token: ${{ secrets.REVIEW_TOKEN }}
+    reviewer-name: test
+    llm-base-url: ${{ secrets.LLM_BASE_URL }}
+    llm-api-key: ${{ secrets.LLM_API_KEY }}
+    llm-model: gpt-4.1
+    dry-run: 'true'
+```
+
+Prints the review to CI logs without posting to the PR. Useful for testing prompt changes.
+
+### Using Anthropic directly
+
+```yaml
+- uses: https://gitea.weiker.me/rodin/review-bot/.gitea/actions/review@v0.1.0
+  with:
+    reviewer-token: ${{ secrets.REVIEW_TOKEN }}
+    reviewer-name: claude
+    llm-base-url: https://api.anthropic.com
+    llm-api-key: ${{ secrets.ANTHROPIC_API_KEY }}
+    llm-model: claude-sonnet-4-20250514
+    llm-provider: anthropic
+```
+
+## 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-model` | Yes | — | Model name |
+| `llm-provider` | No | `openai` | API provider: `openai` or `anthropic` |
+| `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 |
+| `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 |
+
+## How Review Cleanup Works
+
+When `reviewer-name` is set, the bot embeds a hidden sentinel in each review:
+
+```html
+<!-- review-bot:code-review -->
+```
+
+On the next run, it finds and deletes any review containing its own sentinel (except the one it just posted). This means:
+
+- **One review per bot per PR** — no clutter from repeated pushes
+- **Multiple bots coexist** — each only cleans up its own reviews
+- **Same token, different roles** — a single bot account can post "code-review" and "security" reviews without conflict
+- **No extra permissions** — identity comes from the sentinel, not the API
+
+If `reviewer-name` is empty, cleanup is skipped (reviews stack like before).
+
+### Shared Token: Worst-Wins Behavior
+
+When multiple review types share the same Gitea bot account (e.g. code-quality and security), Gitea determines the user's approval state from their **most recent review**. This creates a race condition: if security finds issues (REQUEST_CHANGES) but code-quality finishes last (APPROVE), the PR appears approved.
+
+review-bot handles this automatically with **worst-wins reconciliation**: before posting, each job checks whether any sibling review from the same user already has REQUEST_CHANGES. If so and this job would post APPROVE, it posts as REQUEST_CHANGES instead — maintaining the block. This ensures the PR stays blocked until all checks pass, regardless of execution order.
+
+**If you need independent approval/block per review type**, use separate Gitea bot accounts with their own tokens.
+
+## Custom Review Prompts
+
+Use `system-prompt-file` to specialize the review focus. The file contents are appended to the base system prompt as "Additional Review Instructions."
+
+Example `SECURITY_REVIEW.md`:
+
+```markdown
+You are performing a security-focused code review.
+
+Focus areas:
+- Injection attacks (SQL, command, path traversal, template)
+- Authentication/Authorization (missing checks, privilege escalation)
+- Secrets exposure (hardcoded credentials, tokens in logs)
+- Input validation (unsanitized input, unsafe deserialization)
+- Race conditions (TOCTOU, unsynchronized shared state)
+
+Rules:
+- Only report findings with security implications
+- Ignore style, naming, and general code quality
+- MAJOR = exploitable vulnerability, MINOR = hardening opportunity, NIT = theoretical risk
+- If no security-relevant changes exist, APPROVE with empty findings
+```
+
+## CLI Usage

 ```bash
 review-bot \
@@ -19,71 +244,74 @@ review-bot \
  --repo owner/name \
  --pr 42 \
  --reviewer-token "$GITEA_TOKEN" \
+  --reviewer-name "code-review" \
  --llm-base-url https://api.openai.com/v1 \
  --llm-api-key "$OPENAI_API_KEY" \
-  --llm-model gpt-4 \
-  --reviewer-name "Sonnet" \
-  --conventions-file CONVENTIONS.md \
-  --dry-run
+  --llm-model gpt-4.1 \
+  --conventions-file CONVENTIONS.md
 ```

 ## Environment Variables

-All flags can be set via environment variables:
+All flags have environment variable equivalents:

-| Flag | Env Var | Required | Description |
-|------|---------|----------|-------------|
-| `--gitea-url` | `GITEA_URL` | Yes | Gitea instance base URL |
-| `--repo` | `GITEA_REPO` | Yes | Repository in `owner/name` format |
-| `--pr` | `PR_NUMBER` | Yes | Pull request number |
-| `--reviewer-token` | `REVIEWER_TOKEN` | Yes | Gitea API token for posting reviews |
-| `--llm-base-url` | `LLM_BASE_URL` | Yes | OpenAI-compatible API base URL |
-| `--llm-api-key` | `LLM_API_KEY` | Yes | LLM API key |
-| `--llm-model` | `LLM_MODEL` | Yes | Model identifier |
-| `--reviewer-name` | `REVIEWER_NAME` | No | Display name in review footer |
-| `--conventions-file` | `CONVENTIONS_FILE` | No | Path to conventions file in repo |
-| `--dry-run` | — | No | Print review to stdout instead of posting |
+| Flag | Env Var |
+|------|---------|
+| `--gitea-url` | `GITEA_URL` |
+| `--repo` | `GITEA_REPO` |
+| `--pr` | `PR_NUMBER` |
+| `--reviewer-token` | `REVIEWER_TOKEN` |
+| `--reviewer-name` | `REVIEWER_NAME` |
+| `--llm-base-url` | `LLM_BASE_URL` |
+| `--llm-api-key` | `LLM_API_KEY` |
+| `--llm-model` | `LLM_MODEL` |
+| `--llm-provider` | `LLM_PROVIDER` |
+| `--conventions-file` | `CONVENTIONS_FILE` |
+| `--patterns-repo` | `PATTERNS_REPO` |
+| `--patterns-files` | `PATTERNS_FILES` |
+| `--system-prompt-file` | `SYSTEM_PROMPT_FILE` |
+| `--llm-temperature` | `LLM_TEMPERATURE` |
+| `--llm-timeout` | `LLM_TIMEOUT` |
+| `--update-existing` | `UPDATE_EXISTING` |

-## Adding to a Gitea Repository
+## Setup

-1. Build the binary or use the CI workflow approach (build in CI).
+1. **Create a Gitea bot account** (e.g. `review-bot`)
+2. **Generate a token** with scopes: `write:issue`, `write:repository`
+3. **Add secrets** to your Gitea repo (Settings → Actions → Secrets):
+   - `REVIEW_TOKEN` — the bot's Gitea token
+   - `LLM_BASE_URL` — your LLM endpoint
+   - `LLM_API_KEY` — your LLM key
+4. **Add the workflow** (see Quick Start above)

-2. Add secrets to your Gitea repo (Settings → Actions → Secrets):
-   - `SONNET_REVIEW_TOKEN` — Gitea token for the Sonnet reviewer account
-   - `GPT_REVIEW_TOKEN` — Gitea token for the GPT reviewer account
-   - `LLM_BASE_URL` — Your LLM API endpoint
-   - `LLM_API_KEY` — Your LLM API key
+### Token Scopes Required

-3. Copy `.gitea/workflows/ci.yml` to your repo (or adapt it).
+| Scope | Purpose |
+|-------|---------|
+| `write:issue` | Post and delete reviews |
+| `write:repository` | Read PR diffs, file content, commit statuses |

-4. On every PR, the bot will:
-   - Run tests and vet
-   - Build review-bot
-   - Post reviews from each configured LLM reviewer
+No `read:user` scope needed — the bot identifies itself from the review response.

 ## Development

 ```bash
-# Run tests
-go test ./...
-
-# Run vet
-go vet ./...
-
-# Build
+go test ./...        # Unit tests
+go vet ./...         # Static analysis
 go build -o review-bot ./cmd/review-bot

-# Integration tests (requires env vars)
+# Integration tests (requires env vars set)
 go test -tags=integration ./...
 ```

 ## Architecture

 ```
-cmd/review-bot/     CLI entrypoint
-gitea/              Gitea API client
-llm/                OpenAI-compatible LLM client
+cmd/review-bot/     CLI entrypoint + orchestration
+gitea/              Gitea API client (reviews, PRs, files)
+llm/                Multi-provider LLM client (OpenAI + Anthropic)
 review/             Prompt building, response parsing, formatting
+budget/             Token estimation + context trimming
 ```

 ## License
@@ -0,0 +1,18 @@
+You are performing a security-focused code review. Your primary concern is identifying vulnerabilities, not general code quality.
+
+Focus areas:
+- **Injection attacks**: SQL injection, command injection, path traversal, template injection
+- **Authentication/Authorization**: Missing auth checks, privilege escalation, IDOR
+- **Secrets exposure**: Hardcoded credentials, API keys in code, tokens in logs
+- **Input validation**: Untrusted input used without sanitization, unsafe deserialization
+- **Cryptography**: Weak algorithms, predictable randomness, improper key management
+- **Error handling**: Information leakage in error messages, stack traces exposed
+- **Dependencies**: Known vulnerable patterns, unsafe use of external libraries
+- **Race conditions**: TOCTOU bugs, unsynchronized shared state
+- **Resource exhaustion**: Unbounded allocations, missing timeouts, denial-of-service vectors
+
+Rules for this review:
+- Only report findings with actual security implications. Ignore style, naming, and general code quality.
+- Severity mapping: MAJOR = exploitable vulnerability or data exposure. MINOR = defense-in-depth improvement or hardening opportunity. NIT = theoretical concern with low practical risk.
+- If the code has no security-relevant changes, APPROVE with an empty findings list.
+- Do not duplicate findings that a standard code review would catch (logic bugs, missing error checks) unless they have a security dimension.
@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"log"
 	"os"
+	"path/filepath"
 	"strconv"
 	"strings"
 	"time"
@@ -30,11 +31,14 @@ func main() {
 	llmAPIKey := flag.String("llm-api-key", envOrDefault("LLM_API_KEY", ""), "LLM API key")
 	llmModel := flag.String("llm-model", envOrDefault("LLM_MODEL", ""), "LLM model name")
 	conventionsFile := flag.String("conventions-file", envOrDefault("CONVENTIONS_FILE", ""), "Conventions file path in repo (e.g. CLAUDE.md)")
+	systemPromptFile := flag.String("system-prompt-file", envOrDefault("SYSTEM_PROMPT_FILE", ""), "Local file with additional system prompt instructions")
 	patternsRepo := flag.String("patterns-repo", envOrDefault("PATTERNS_REPO", ""), "Repo with language patterns (e.g. rodin/elixir-patterns)")
 	patternsFiles := flag.String("patterns-files", envOrDefault("PATTERNS_FILES", "README.md"), "Comma-separated file paths to fetch from patterns repo")
 	dryRun := flag.Bool("dry-run", false, "Print review to stdout instead of posting")
+	updateExisting := flag.Bool("update-existing", envOrDefaultBool("UPDATE_EXISTING", true), "Delete previous review from same bot before posting (default true)")
 	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")

 	flag.Parse()

@@ -75,6 +79,12 @@ func main() {
 	if *llmTemp > 0 {
 		llmClient.WithTemperature(*llmTemp)
 	}
+	switch llm.Provider(*llmProvider) {
+	case llm.ProviderOpenAI, llm.ProviderAnthropic:
+		llmClient.WithProvider(llm.Provider(*llmProvider))
+	default:
+		log.Fatalf("Invalid --llm-provider %q, must be openai or anthropic", *llmProvider)
+	}
 	if *llmTimeout > 0 {
 		llmClient.WithTimeout(time.Duration(*llmTimeout) * time.Second)
 	}
@@ -142,9 +152,45 @@ func main() {
 		log.Printf("Loaded patterns from %s (%d bytes)", *patternsRepo, len(patterns))
 	}

+	// 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)
+		if err != nil {
+			log.Fatalf("Failed to resolve workspace path: %v", err)
+		}
+		promptPath := filepath.Join(absWorkspace, *systemPromptFile)
+		promptPath = filepath.Clean(promptPath)
+		if !strings.HasPrefix(promptPath, absWorkspace+string(filepath.Separator)) && promptPath != absWorkspace {
+			log.Fatalf("system-prompt-file resolves outside workspace (got %q, workspace %q)", promptPath, absWorkspace)
+		}
+		// Resolve symlinks and re-validate to prevent symlink traversal
+		resolvedPath, err := filepath.EvalSymlinks(promptPath)
+		if err != nil {
+			log.Fatalf("Failed to resolve system prompt file %q: %v", promptPath, err)
+		}
+		if !strings.HasPrefix(resolvedPath, absWorkspace+string(filepath.Separator)) && resolvedPath != absWorkspace {
+			log.Fatalf("system-prompt-file symlink resolves outside workspace (got %q, workspace %q)", resolvedPath, absWorkspace)
+		}
+		data, err := os.ReadFile(resolvedPath)
+		if err != nil {
+			log.Fatalf("Failed to read system prompt file %q: %v", promptPath, err)
+		}
+		additionalPrompt = string(data)
+		log.Printf("Loaded system prompt file: %s (%d bytes)", *systemPromptFile, len(additionalPrompt))
+	}
+
 	// Step 7: Budget-aware prompt assembly
+	systemBase := review.BuildSystemBase()
+	if additionalPrompt != "" {
+		systemBase += "\n\n## Additional Review Instructions\n\n" + additionalPrompt
+	}
 	sections := budget.Sections{
-		SystemBase:  review.BuildSystemBase(),
+		SystemBase:  systemBase,
 		Patterns:    patterns,
 		Conventions: conventions,
 		FileContext: fileContext,
@@ -188,11 +234,53 @@ func main() {
 		return
 	}

+	// Validate reviewer-name: only safe characters allowed in sentinel
+	if err := validateReviewerName(*reviewerName); err != nil {
+		log.Fatalf("%v", err)
+	}
+	sentinel := fmt.Sprintf("<!-- review-bot:%s -->", *reviewerName)
+
 	log.Printf("Posting review (event=%s)...", event)
-	if err := giteaClient.PostReview(ctx, owner, repoName, prNumber, event, reviewBody); err != nil {
+	posted, err := giteaClient.PostReview(ctx, owner, repoName, prNumber, event, reviewBody)
+	if err != nil {
 		log.Fatalf("Failed to post review: %v", err)
 	}
-	log.Printf("Review posted successfully!")
+	log.Printf("Review posted (id=%d, user=%s)", posted.ID, posted.User.Login)
+
+	// Delete stale reviews from this bot using sentinel matching
+	if *updateExisting && *reviewerName != "" {
+		reviews, err := giteaClient.ListReviews(ctx, owner, repoName, prNumber)
+		if err != nil {
+			log.Printf("Warning: could not list existing reviews: %v", err)
+		} else {
+			for _, r := range reviews {
+				if r.ID != posted.ID && r.User.Login == posted.User.Login && strings.Contains(r.Body, sentinel) {
+					if err := giteaClient.DeleteReview(ctx, owner, repoName, prNumber, r.ID); err != nil {
+						log.Printf("Warning: could not delete old review %d: %v", r.ID, err)
+					} else {
+						log.Printf("Deleted stale review %d", r.ID)
+					}
+				}
+			}
+
+			// Worst-wins: if we posted APPROVE but a sibling review from the
+			// same user (same token, different role) has REQUEST_CHANGES,
+			// delete ours and re-post as REQUEST_CHANGES to maintain the block.
+			if event == "APPROVED" && shouldEscalate(reviews, posted.ID, posted.User.Login, sentinel) {
+				log.Printf("Sibling review has REQUEST_CHANGES; escalating")
+				if err := giteaClient.DeleteReview(ctx, owner, repoName, prNumber, posted.ID); err != nil {
+					log.Printf("Warning: could not delete review for escalation: %v", err)
+				} else {
+					_, err := giteaClient.PostReview(ctx, owner, repoName, prNumber, "REQUEST_CHANGES", reviewBody)
+					if err != nil {
+						log.Printf("Warning: could not re-post as REQUEST_CHANGES: %v", err)
+					} else {
+						log.Printf("Review escalated to REQUEST_CHANGES")
+					}
+				}
+			}
+		}
+	}
 }

 // fetchFileContext fetches the full content of modified files from the PR branch.
@@ -326,3 +414,37 @@ func envOrDefaultInt(key string, defaultVal int) int {
 	}
 	return defaultVal
 }
+
+func envOrDefaultBool(key string, defaultVal bool) bool {
+	v := strings.TrimSpace(strings.ToLower(os.Getenv(key)))
+	if v == "" {
+		return defaultVal
+	}
+	return v == "true" || v == "1" || v == "yes"
+}
+
+// validateReviewerName checks that the name contains only safe characters
+// for embedding in an HTML comment sentinel ([a-zA-Z0-9_-]).
+func validateReviewerName(name string) error {
+	if name == "" {
+		return nil
+	}
+	for _, ch := range name {
+		if !((ch >= 'a' && ch <= 'z') || (ch >= 'A' && ch <= 'Z') || (ch >= '0' && ch <= '9') || ch == '-' || ch == '_') {
+			return fmt.Errorf("reviewer-name must contain only [a-zA-Z0-9_-] (got %q)", name)
+		}
+	}
+	return nil
+}
+
+// shouldEscalate checks if the current APPROVED review should be escalated
+// to REQUEST_CHANGES because a sibling bot review (same user, different role)
+// already has REQUEST_CHANGES.
+func shouldEscalate(reviews []gitea.Review, postedID int64, postedLogin, ownSentinel string) bool {
+	for _, r := range reviews {
+		if r.ID != postedID && !r.Stale && r.User.Login == postedLogin && r.State == "REQUEST_CHANGES" && strings.Contains(r.Body, "<!-- review-bot:") && !strings.Contains(r.Body, ownSentinel) {
+			return true
+		}
+	}
+	return false
+}
@@ -0,0 +1,150 @@
+package main
+
+import (
+	"testing"
+
+	"gitea.weiker.me/rodin/review-bot/gitea"
+)
+
+func TestValidateReviewerName(t *testing.T) {
+	tests := []struct {
+		name    string
+		input   string
+		wantErr bool
+	}{
+		{"valid simple", "sonnet", false},
+		{"valid with dash", "code-review", false},
+		{"valid with underscore", "my_bot", false},
+		{"valid alphanumeric", "bot123", false},
+		{"valid uppercase", "MyBot", false},
+		{"empty is valid", "", false},
+		{"invalid html close", "foo-->", true},
+		{"invalid space", "my bot", true},
+		{"invalid dot", "my.bot", true},
+		{"invalid slash", "my/bot", true},
+		{"invalid angle", "bot<script>", true},
+		{"invalid colon", "bot:name", true},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			err := validateReviewerName(tc.input)
+			if tc.wantErr && err == nil {
+				t.Errorf("expected error for %q, got nil", tc.input)
+			}
+			if !tc.wantErr && err != nil {
+				t.Errorf("expected no error for %q, got %v", tc.input, err)
+			}
+		})
+	}
+}
+
+func makeReview(id int64, login, state string, stale bool, body string) gitea.Review {
+	r := gitea.Review{
+		ID:    id,
+		Body:  body,
+		State: state,
+		Stale: stale,
+	}
+	r.User.Login = login
+	return r
+}
+
+func TestShouldEscalate(t *testing.T) {
+	tests := []struct {
+		name        string
+		reviews     []gitea.Review
+		postedID    int64
+		postedLogin string
+		ownSentinel string
+		want        bool
+	}{
+		{
+			name:        "no reviews",
+			reviews:     nil,
+			postedID:    100,
+			postedLogin: "bot",
+			ownSentinel: "<!-- review-bot:sonnet -->",
+			want:        false,
+		},
+		{
+			name: "sibling same user has REQUEST_CHANGES",
+			reviews: []gitea.Review{
+				makeReview(101, "bot", "REQUEST_CHANGES", false, "bad\n<!-- review-bot:security -->"),
+			},
+			postedID:    100,
+			postedLogin: "bot",
+			ownSentinel: "<!-- review-bot:sonnet -->",
+			want:        true,
+		},
+		{
+			name: "sibling different user has REQUEST_CHANGES (should NOT escalate)",
+			reviews: []gitea.Review{
+				makeReview(101, "other-bot", "REQUEST_CHANGES", false, "bad\n<!-- review-bot:gpt -->"),
+			},
+			postedID:    100,
+			postedLogin: "bot",
+			ownSentinel: "<!-- review-bot:sonnet -->",
+			want:        false,
+		},
+		{
+			name: "same user REQUEST_CHANGES but stale (should NOT escalate)",
+			reviews: []gitea.Review{
+				makeReview(101, "bot", "REQUEST_CHANGES", true, "old\n<!-- review-bot:security -->"),
+			},
+			postedID:    100,
+			postedLogin: "bot",
+			ownSentinel: "<!-- review-bot:sonnet -->",
+			want:        false,
+		},
+		{
+			name: "same user same sentinel (own stale review, should NOT escalate)",
+			reviews: []gitea.Review{
+				makeReview(101, "bot", "REQUEST_CHANGES", false, "old\n<!-- review-bot:sonnet -->"),
+			},
+			postedID:    100,
+			postedLogin: "bot",
+			ownSentinel: "<!-- review-bot:sonnet -->",
+			want:        false,
+		},
+		{
+			name: "same user APPROVED sibling (should NOT escalate)",
+			reviews: []gitea.Review{
+				makeReview(101, "bot", "APPROVED", false, "good\n<!-- review-bot:security -->"),
+			},
+			postedID:    100,
+			postedLogin: "bot",
+			ownSentinel: "<!-- review-bot:sonnet -->",
+			want:        false,
+		},
+		{
+			name: "human REQUEST_CHANGES no sentinel (should NOT escalate)",
+			reviews: []gitea.Review{
+				makeReview(101, "bot", "REQUEST_CHANGES", false, "please fix this"),
+			},
+			postedID:    100,
+			postedLogin: "bot",
+			ownSentinel: "<!-- review-bot:sonnet -->",
+			want:        false,
+		},
+		{
+			name: "skip own posted ID",
+			reviews: []gitea.Review{
+				makeReview(100, "bot", "REQUEST_CHANGES", false, "x\n<!-- review-bot:security -->"),
+			},
+			postedID:    100,
+			postedLogin: "bot",
+			ownSentinel: "<!-- review-bot:sonnet -->",
+			want:        false,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			got := shouldEscalate(tc.reviews, tc.postedID, tc.postedLogin, tc.ownSentinel)
+			if got != tc.want {
+				t.Errorf("shouldEscalate() = %v, want %v", got, tc.want)
+			}
+		})
+	}
+}
@@ -129,9 +129,9 @@ func (c *Client) GetFileContentRef(ctx context.Context, owner, repo, filepath, r
 	return string(body), nil
 }

-// PostReview submits a review to a PR.
+// PostReview submits a review to a PR and returns the created review.
 // event should be "APPROVED" or "REQUEST_CHANGES".
-func (c *Client) PostReview(ctx context.Context, owner, repo string, number int, event, body string) error {
+func (c *Client) PostReview(ctx context.Context, owner, repo string, number int, event, body string) (*Review, error) {
 	reqURL := fmt.Sprintf("%s/api/v1/repos/%s/%s/pulls/%d/reviews", c.baseURL, owner, repo, number)

 	payload := struct {
@@ -144,27 +144,36 @@ func (c *Client) PostReview(ctx context.Context, owner, repo string, number int,

 	data, err := json.Marshal(payload)
 	if err != nil {
-		return fmt.Errorf("marshal review payload: %w", err)
+		return nil, fmt.Errorf("marshal review payload: %w", err)
 	}

 	req, err := http.NewRequestWithContext(ctx, http.MethodPost, reqURL, bytes.NewReader(data))
 	if err != nil {
-		return fmt.Errorf("create review request: %w", err)
+		return nil, fmt.Errorf("create review request: %w", err)
 	}
 	req.Header.Set("Authorization", "token "+c.token)
 	req.Header.Set("Content-Type", "application/json")

 	resp, err := c.http.Do(req)
 	if err != nil {
-		return fmt.Errorf("post review: %w", err)
+		return nil, fmt.Errorf("post review: %w", err)
 	}
 	defer resp.Body.Close()

 	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
 		respBody, _ := io.ReadAll(resp.Body)
-		return fmt.Errorf("post review failed (status %d): %s", resp.StatusCode, string(respBody))
+		return nil, fmt.Errorf("post review failed (status %d): %s", resp.StatusCode, string(respBody))
 	}
-	return nil
+
+	respBody, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, fmt.Errorf("read review response: %w", err)
+	}
+	var review Review
+	if err := json.Unmarshal(respBody, &review); err != nil {
+		return nil, fmt.Errorf("parse review response: %w", err)
+	}
+	return &review, nil
 }

 func (c *Client) doGet(ctx context.Context, reqURL string) ([]byte, error) {
@@ -266,3 +275,71 @@ func (c *Client) GetAllFilesInPath(ctx context.Context, owner, repo, path string
 	}
 	return results, nil
 }
+
+// Review represents a pull request review from the Gitea API.
+type Review struct {
+	ID   int64  `json:"id"`
+	Body string `json:"body"`
+	User struct {
+		Login string `json:"login"`
+	} `json:"user"`
+	State string `json:"state"`
+	Stale bool   `json:"stale"`
+}
+
+// ListReviews returns all reviews on a pull request.
+// Paginates through all pages to ensure no reviews are missed.
+func (c *Client) ListReviews(ctx context.Context, owner, repo string, number int) ([]Review, error) {
+	const pageSize = 50
+	var all []Review
+	for page := 1; ; page++ {
+		reqURL := fmt.Sprintf("%s/api/v1/repos/%s/%s/pulls/%d/reviews?limit=%d&page=%d",
+			c.baseURL,
+			url.PathEscape(owner),
+			url.PathEscape(repo),
+			number,
+			pageSize,
+			page)
+		body, err := c.doGet(ctx, reqURL)
+		if err != nil {
+			return nil, fmt.Errorf("list reviews (page %d): %w", page, err)
+		}
+		var batch []Review
+		if err := json.Unmarshal(body, &batch); err != nil {
+			return nil, fmt.Errorf("parse reviews (page %d): %w", page, err)
+		}
+		all = append(all, batch...)
+		if len(batch) < pageSize {
+			break
+		}
+	}
+	return all, nil
+}
+
+// DeleteReview deletes a review by ID. The token must belong to the review author.
+func (c *Client) DeleteReview(ctx context.Context, owner, repo string, number int, reviewID int64) error {
+	reqURL := fmt.Sprintf("%s/api/v1/repos/%s/%s/pulls/%d/reviews/%d",
+		c.baseURL,
+		url.PathEscape(owner),
+		url.PathEscape(repo),
+		number,
+		reviewID)
+
+	req, err := http.NewRequestWithContext(ctx, http.MethodDelete, reqURL, nil)
+	if err != nil {
+		return fmt.Errorf("create delete request: %w", err)
+	}
+	req.Header.Set("Authorization", "token "+c.token)
+
+	resp, err := c.http.Do(req)
+	if err != nil {
+		return fmt.Errorf("delete review: %w", err)
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		respBody, _ := io.ReadAll(resp.Body)
+		return fmt.Errorf("delete review failed (status %d): %s", resp.StatusCode, string(respBody))
+	}
+	return nil
+}
@@ -123,15 +123,21 @@ func TestPostReview(t *testing.T) {
 		}

 		w.WriteHeader(http.StatusOK)
-		w.Write([]byte(`{}`))
+		w.Write([]byte(`{"id":100,"user":{"login":"review-bot"},"state":"APPROVED","stale":false}`))
 	}))
 	defer server.Close()

 	client := NewClient(server.URL, "test-token")
-	err := client.PostReview(context.Background(), "owner", "repo", 3, "APPROVED", "LGTM")
+	review, err := client.PostReview(context.Background(), "owner", "repo", 3, "APPROVED", "LGTM")
 	if err != nil {
 		t.Fatalf("unexpected error: %v", err)
 	}
+	if review.ID != 100 {
+		t.Errorf("expected review ID 100, got %d", review.ID)
+	}
+	if review.User.Login != "review-bot" {
+		t.Errorf("expected user login %q, got %q", "review-bot", review.User.Login)
+	}
 }

 func TestGetPullRequest_Non200(t *testing.T) {
@@ -169,7 +175,7 @@ func TestPostReview_Non200(t *testing.T) {
 	defer server.Close()

 	client := NewClient(server.URL, "test-token")
-	err := client.PostReview(context.Background(), "owner", "repo", 1, "APPROVED", "test")
+	_, err := client.PostReview(context.Background(), "owner", "repo", 1, "APPROVED", "test")
 	if err == nil {
 		t.Fatal("expected error for 403, got nil")
 	}
@@ -318,3 +324,105 @@ func TestEscapePath(t *testing.T) {
 		})
 	}
 }
+
+func TestListReviews(t *testing.T) {
+	pageCount := 0
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/api/v1/repos/owner/repo/pulls/5/reviews" {
+			t.Errorf("unexpected path: %s", r.URL.Path)
+		}
+		if r.URL.Query().Get("limit") != "50" {
+			t.Errorf("expected limit=50, got %s", r.URL.Query().Get("limit"))
+		}
+		pageCount++
+		w.Header().Set("Content-Type", "application/json")
+		// Return 2 results (less than page size) to signal end
+		w.Write([]byte(`[{"id":10,"user":{"login":"bot-a"},"state":"APPROVED","stale":false},{"id":11,"user":{"login":"bot-b"},"state":"REQUEST_CHANGES","stale":true}]`))
+	}))
+	defer server.Close()
+
+	client := NewClient(server.URL, "test-token")
+	reviews, err := client.ListReviews(context.Background(), "owner", "repo", 5)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(reviews) != 2 {
+		t.Fatalf("expected 2 reviews, got %d", len(reviews))
+	}
+	if reviews[0].User.Login != "bot-a" {
+		t.Errorf("expected bot-a, got %s", reviews[0].User.Login)
+	}
+	if pageCount != 1 {
+		t.Errorf("expected 1 page fetch (results < page size), got %d", pageCount)
+	}
+}
+
+func TestListReviews_Pagination(t *testing.T) {
+	pageCount := 0
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		pageCount++
+		page := r.URL.Query().Get("page")
+		w.Header().Set("Content-Type", "application/json")
+		if page == "1" {
+			// Return exactly 50 items to trigger next page fetch
+			items := "["
+			for i := 0; i < 50; i++ {
+				if i > 0 {
+					items += ","
+				}
+				items += fmt.Sprintf(`{"id":%d,"user":{"login":"bot"},"state":"APPROVED","stale":false}`, i+1)
+			}
+			items += "]"
+			w.Write([]byte(items))
+		} else {
+			// Page 2: return fewer than 50 to signal end
+			w.Write([]byte(`[{"id":51,"user":{"login":"bot"},"state":"APPROVED","stale":false}]`))
+		}
+	}))
+	defer server.Close()
+
+	client := NewClient(server.URL, "test-token")
+	reviews, err := client.ListReviews(context.Background(), "owner", "repo", 5)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if len(reviews) != 51 {
+		t.Fatalf("expected 51 reviews across 2 pages, got %d", len(reviews))
+	}
+	if pageCount != 2 {
+		t.Errorf("expected 2 page fetches, got %d", pageCount)
+	}
+}
+
+func TestDeleteReview(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/api/v1/repos/owner/repo/pulls/5/reviews/10" {
+			t.Errorf("unexpected path: %s", r.URL.Path)
+		}
+		if r.Method != "DELETE" {
+			t.Errorf("expected DELETE, got %s", r.Method)
+		}
+		w.WriteHeader(http.StatusNoContent)
+	}))
+	defer server.Close()
+
+	client := NewClient(server.URL, "test-token")
+	err := client.DeleteReview(context.Background(), "owner", "repo", 5, 10)
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+}
+
+func TestDeleteReview_Forbidden(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusForbidden)
+		w.Write([]byte(`{"message":"forbidden"}`))
+	}))
+	defer server.Close()
+
+	client := NewClient(server.URL, "test-token")
+	err := client.DeleteReview(context.Background(), "owner", "repo", 5, 10)
+	if err == nil {
+		t.Fatal("expected error for 403, got nil")
+	}
+}
@@ -1,4 +1,6 @@
-// Package llm provides a client for OpenAI-compatible chat completion APIs.
+// Package llm provides clients for LLM chat completion APIs.
+//
+// Supports OpenAI-compatible (default) and Anthropic Messages API providers.
 package llm

 import (
@@ -12,24 +14,37 @@ import (
 	"time"
 )

-// Client calls an OpenAI-compatible chat completion API.
+// Provider identifies which API format to use.
+type Provider string
+
+const (
+	// ProviderOpenAI uses the OpenAI-compatible chat/completions endpoint.
+	ProviderOpenAI Provider = "openai"
+	// ProviderAnthropic uses the Anthropic Messages API endpoint.
+	ProviderAnthropic Provider = "anthropic"
+)
+
+// Client calls an LLM chat completion API.
 // A Client is safe for concurrent use by multiple goroutines after construction.
-// WithTimeout and WithTemperature must be called during setup, before concurrent use.
+// WithTimeout, WithTemperature, and WithProvider must be called during setup,
+// before concurrent use.
 type Client struct {
 	baseURL     string
 	apiKey      string
 	model       string
 	temperature float64
+	provider    Provider
 	http        *http.Client
 }

-// NewClient creates a new LLM client.
+// NewClient creates a new LLM client. Default provider is OpenAI-compatible.
 func NewClient(baseURL, apiKey, model string) *Client {
 	return &Client{
-		baseURL: strings.TrimRight(baseURL, "/"),
-		apiKey:  apiKey,
-		model:   model,
-		http:    &http.Client{Timeout: 5 * time.Minute},
+		baseURL:  strings.TrimRight(baseURL, "/"),
+		apiKey:   apiKey,
+		model:    model,
+		provider: ProviderOpenAI,
+		http:     &http.Client{Timeout: 5 * time.Minute},
 	}
 }

@@ -45,20 +60,39 @@ func (c *Client) WithTemperature(t float64) *Client {
 	return c
 }

+// WithProvider sets the API provider format (openai or anthropic).
+func (c *Client) WithProvider(p Provider) *Client {
+	c.provider = p
+	return c
+}
+
 // Message represents a chat message.
 type Message struct {
 	Role    string `json:"role"`
 	Content string `json:"content"`
 }

-// ChatRequest is the request payload.
+// 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)
+	}
+}
+
+// --- OpenAI-compatible implementation ---
+
+// ChatRequest is the OpenAI request payload.
 type ChatRequest struct {
 	Model       string    `json:"model"`
 	Messages    []Message `json:"messages"`
 	Temperature float64   `json:"temperature,omitempty"`
 }

-// ChatResponse is the response from the API.
+// ChatResponse is the OpenAI response.
 type ChatResponse struct {
 	Choices []struct {
 		Message struct {
@@ -67,8 +101,7 @@ type ChatResponse struct {
 	} `json:"choices"`
 }

-// Complete sends a chat completion request and returns the assistant's response content.
-func (c *Client) Complete(ctx context.Context, messages []Message) (string, error) {
+func (c *Client) completeOpenAI(ctx context.Context, messages []Message) (string, error) {
 	reqBody := ChatRequest{
 		Model:       c.model,
 		Temperature: c.temperature,
@@ -81,37 +114,126 @@ func (c *Client) Complete(ctx context.Context, messages []Message) (string, erro
 	}

 	url := c.baseURL + "/chat/completions"
-	req, err := http.NewRequestWithContext(ctx, "POST", url, bytes.NewReader(data))
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(data))
 	if err != nil {
 		return "", fmt.Errorf("create request: %w", err)
 	}
 	req.Header.Set("Authorization", "Bearer "+c.apiKey)
 	req.Header.Set("Content-Type", "application/json")

+	return c.doRequest(req, func(body []byte) (string, error) {
+		var resp ChatResponse
+		if err := json.Unmarshal(body, &resp); err != nil {
+			return "", fmt.Errorf("parse response: %w", err)
+		}
+		if len(resp.Choices) == 0 {
+			return "", fmt.Errorf("no choices in LLM response")
+		}
+		return resp.Choices[0].Message.Content, nil
+	})
+}
+
+// --- 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"`
+}
+
+type anthropicMsg struct {
+	Role    string `json:"role"`
+	Content string `json:"content"`
+}
+
+type anthropicResponse struct {
+	Content []struct {
+		Type string `json:"type"`
+		Text string `json:"text"`
+	} `json:"content"`
+}
+
+func (c *Client) completeAnthropic(ctx context.Context, messages []Message) (string, error) {
+	// Extract system message (first message with role "system")
+	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{
+		Model:     c.model,
+		MaxTokens: 8192,
+		System:    system,
+		Messages:  userMessages,
+	}
+	if c.temperature > 0 {
+		reqBody.Temperature = c.temperature
+	}
+
+	data, err := json.Marshal(reqBody)
+	if err != nil {
+		return "", fmt.Errorf("marshal request: %w", err)
+	}
+
+	url := c.baseURL + "/messages"
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(data))
+	if err != nil {
+		return "", fmt.Errorf("create request: %w", err)
+	}
+	req.Header.Set("x-api-key", c.apiKey)
+	req.Header.Set("anthropic-version", "2023-06-01")
+	req.Header.Set("Content-Type", "application/json")
+
+	return c.doRequest(req, func(body []byte) (string, error) {
+		var resp anthropicResponse
+		if err := json.Unmarshal(body, &resp); err != nil {
+			return "", fmt.Errorf("parse response: %w", err)
+		}
+		if len(resp.Content) == 0 {
+			return "", fmt.Errorf("no content in Anthropic response")
+		}
+		// Concatenate all text blocks
+		var sb strings.Builder
+		for _, block := range resp.Content {
+			if block.Type == "text" {
+				sb.WriteString(block.Text)
+			}
+		}
+		result := sb.String()
+		if result == "" {
+			return "", fmt.Errorf("no text content in Anthropic response")
+		}
+		return result, nil
+	})
+}
+
+// --- Shared HTTP execution ---
+
+func (c *Client) doRequest(req *http.Request, parse func([]byte) (string, error)) (string, error) {
 	resp, err := c.http.Do(req)
 	if err != nil {
 		return "", fmt.Errorf("LLM request: %w", err)
 	}
 	defer resp.Body.Close()

-	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
-		body, _ := io.ReadAll(resp.Body)
-		return "", fmt.Errorf("LLM API error (status %d): %s", resp.StatusCode, string(body))
-	}
-
 	body, err := io.ReadAll(resp.Body)
 	if err != nil {
 		return "", fmt.Errorf("read response: %w", err)
 	}

-	var chatResp ChatResponse
-	if err := json.Unmarshal(body, &chatResp); err != nil {
-		return "", fmt.Errorf("parse response: %w", err)
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return "", fmt.Errorf("LLM API error (status %d): %s", resp.StatusCode, string(body))
 	}

-	if len(chatResp.Choices) == 0 {
-		return "", fmt.Errorf("no choices in LLM response")
-	}
-
-	return chatResp.Choices[0].Message.Content, nil
+	return parse(body)
 }
@@ -208,3 +208,90 @@ func TestWithTimeout(t *testing.T) {
 		t.Error("expected timeout error with 50ms timeout and 200ms server delay")
 	}
 }
+
+
+func TestComplete_Anthropic_Success(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/messages" {
+			t.Errorf("unexpected path: %s", r.URL.Path)
+		}
+		if r.Header.Get("x-api-key") != "test-key" {
+			t.Errorf("expected x-api-key header, got %q", r.Header.Get("x-api-key"))
+		}
+		if r.Header.Get("anthropic-version") != "2023-06-01" {
+			t.Errorf("expected anthropic-version header, got %q", r.Header.Get("anthropic-version"))
+		}
+
+		var req map[string]interface{}
+		json.NewDecoder(r.Body).Decode(&req)
+
+		if req["system"] != "You are helpful" {
+			t.Errorf("expected system prompt, got %v", req["system"])
+		}
+		msgs := req["messages"].([]interface{})
+		if len(msgs) != 1 {
+			t.Errorf("expected 1 user message, got %d", len(msgs))
+		}
+		if req["max_tokens"] != float64(8192) {
+			t.Errorf("expected max_tokens 8192, got %v", req["max_tokens"])
+		}
+
+		w.Header().Set("Content-Type", "application/json")
+		w.Write([]byte(`{"content":[{"type":"text","text":"Hello from Claude!"}]}`))
+	}))
+	defer server.Close()
+
+	client := NewClient(server.URL, "test-key", "claude-sonnet").WithProvider(ProviderAnthropic)
+	got, err := client.Complete(context.Background(), []Message{
+		{Role: "system", Content: "You are helpful"},
+		{Role: "user", Content: "Hi"},
+	})
+	if err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if got != "Hello from Claude!" {
+		t.Errorf("expected %q, got %q", "Hello from Claude!", got)
+	}
+}
+
+func TestComplete_Anthropic_NoContent(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.Write([]byte(`{"content":[]}`))
+	}))
+	defer server.Close()
+
+	client := NewClient(server.URL, "test-key", "claude-sonnet").WithProvider(ProviderAnthropic)
+	_, err := client.Complete(context.Background(), []Message{{Role: "user", Content: "Hi"}})
+	if err == nil {
+		t.Fatal("expected error for empty content, got nil")
+	}
+}
+
+func TestComplete_Anthropic_APIError(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusBadRequest)
+		w.Write([]byte(`{"error":{"message":"invalid request"}}`))
+	}))
+	defer server.Close()
+
+	client := NewClient(server.URL, "test-key", "claude-sonnet").WithProvider(ProviderAnthropic)
+	_, err := client.Complete(context.Background(), []Message{{Role: "user", Content: "Hi"}})
+	if err == nil {
+		t.Fatal("expected error for 400, got nil")
+	}
+}
+
+func TestWithProvider(t *testing.T) {
+	client := NewClient("http://example.com", "key", "model")
+	if client.provider != ProviderOpenAI {
+		t.Errorf("expected default provider openai, got %s", client.provider)
+	}
+	result := client.WithProvider(ProviderAnthropic)
+	if result != client {
+		t.Error("WithProvider should return the same client for chaining")
+	}
+	if client.provider != ProviderAnthropic {
+		t.Errorf("expected provider anthropic, got %s", client.provider)
+	}
+}
@@ -9,6 +9,11 @@ import (
 func FormatMarkdown(result *ReviewResult, reviewerName string) string {
 	var sb strings.Builder

+	if reviewerName != "" {
+		title := strings.ToUpper(reviewerName[:1]) + reviewerName[1:]
+		sb.WriteString(fmt.Sprintf("# %s Review\n\n", title))
+	}
+
 	sb.WriteString("## Summary\n\n")
 	sb.WriteString(result.Summary)
 	sb.WriteString("\n\n")
@@ -30,6 +35,8 @@ func FormatMarkdown(result *ReviewResult, reviewerName string) string {

 	if reviewerName != "" {
 		sb.WriteString(fmt.Sprintf("\n---\n*Review by %s*\n", reviewerName))
+		// Hidden sentinel for identifying this bot's reviews during cleanup
+		sb.WriteString(fmt.Sprintf("\n<!-- review-bot:%s -->\n", reviewerName))
 	}

 	return sb.String()
@@ -116,3 +116,21 @@ func TestGiteaEvent(t *testing.T) {
 		}
 	}
 }
+
+func TestFormatMarkdown_Sentinel(t *testing.T) {
+	result := &ReviewResult{
+		Verdict:        "APPROVE",
+		Summary:        "All good.",
+		Recommendation: "Merge it.",
+	}
+	output := FormatMarkdown(result, "security")
+	if !strings.Contains(output, "<!-- review-bot:security -->") {
+		t.Error("expected sentinel comment in output")
+	}
+
+	// Empty reviewer name should NOT have sentinel
+	output2 := FormatMarkdown(result, "")
+	if strings.Contains(output2, "<!-- review-bot") {
+		t.Error("should not contain sentinel when reviewer name is empty")
+	}
+}
Author	SHA1	Message	Date
Rodin	dd6c6c610f	feat: add role title as H1 header for visual differentiation CI / test (pull_request) Successful in 14s Details CI / review (gpt-4.1, gpt, GPT_REVIEW_TOKEN) (pull_request) Successful in 26s Details CI / review (gpt-5, sonnet, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m34s Details CI / review (gpt-5, security, SECURITY_REVIEW.md, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m48s Details When reviewer-name is set, prepend "# Security Review" / "# Sonnet Review" etc. as a top-level header. Makes it immediately obvious which role each review represents in the Gitea UI, especially when multiple reviews come from the same bot account.	2026-05-01 21:35:41 -07:00
Rodin	436e6a8824	fix: symlink traversal + worst-wins pre-check + user scoping CI / test (pull_request) Successful in 14s Details CI / review (gpt-4.1, gpt, GPT_REVIEW_TOKEN) (pull_request) Successful in 20s Details CI / review (gpt-5, security, SECURITY_REVIEW.md, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m11s Details CI / review (gpt-5, sonnet, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m16s Details Security (MAJOR): - Add filepath.EvalSymlinks after Clean for system-prompt-file - Re-validate resolved path is still within workspace - Prevents symlink → /etc/shadow exfiltration via malicious repo Worst-wins: - Check BEFORE posting (not after) — no delete+repost dance - Identify sibling bots by <!-- review-bot: prefix in body - Only escalates for bot reviews, not human REQUEST_CHANGES - If sibling bot has REQUEST_CHANGES and we would APPROVE → post REQUEST_CHANGES instead Addresses security review finding #1 (MAJOR) and sonnet finding #1.	2026-05-01 21:31:17 -07:00
Rodin	687005d982	feat: worst-wins reconciliation for shared-token review types CI / test (pull_request) Successful in 13s Details CI / review (gpt-4.1, gpt, GPT_REVIEW_TOKEN) (pull_request) Successful in 24s Details CI / review (gpt-5, security, SECURITY_REVIEW.md, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m9s Details CI / review (gpt-5, sonnet, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m18s Details When multiple review types share a Gitea bot account, Gitea uses the latest review to determine the user's approval state. This creates a race: if security finds issues but code-quality finishes last with APPROVE, the PR appears approved. Now before posting, each job checks if any sibling review from the same user has REQUEST_CHANGES. If so and we would post APPROVE, we downgrade to COMMENT instead — the review is still visible but won't override the blocking state. Documented in README under "Shared Token: Worst-Wins."	2026-05-01 21:12:34 -07:00
Rodin	6a3c813279	fix: address review findings (path restriction, login cross-check, README) CI / test (pull_request) Successful in 14s Details CI / review (gpt-4.1, gpt, GPT_REVIEW_TOKEN) (pull_request) Successful in 19s Details CI / review (gpt-5, sonnet, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m15s Details CI / review (gpt-5, security, SECURITY_REVIEW.md, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m20s Details - system-prompt-file: reject absolute paths and paths containing ".." Prevents reading arbitrary files outside the workspace on shared runners. - Cleanup: cross-check r.User.Login == posted.User.Login before deletion Defense-in-depth: only attempt to delete reviews from same author. Flagged by both sonnet and security reviewers. - README: fix wording (cleanup happens after posting, not before) Issues filed for deferred work: - #24: Consistent url.PathEscape across all client endpoints - #25: Binary signature verification for supply-chain hardening	2026-05-01 21:05:18 -07:00
Rodin	b8af8306a6	docs: comprehensive README with action usage, cleanup behavior, custom prompts CI / test (pull_request) Successful in 14s Details CI / review (gpt-4.1, gpt, GPT_REVIEW_TOKEN) (pull_request) Successful in 23s Details CI / review (gpt-5, security, SECURITY_REVIEW.md, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m5s Details CI / review (gpt-5, sonnet, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m37s Details - Quick start example with composite action + matrix strategy - Full action inputs table with descriptions - How sentinel-based cleanup works (explains the reviewer-name concept) - Custom prompt file usage with security review example - CLI usage with all flags - Environment variables table - Token scopes documentation - Setup guide for new repos	2026-05-01 20:59:34 -07:00
Rodin	69e0a459c3	feat: sentinel-based review cleanup + system prompt file + security review CI / test (pull_request) Successful in 14s Details CI / review (gpt-4.1, gpt, GPT_REVIEW_TOKEN) (pull_request) Successful in 23s Details CI / review (gpt-5, sonnet, SONNET_REVIEW_TOKEN) (pull_request) Successful in 58s Details CI / review (gpt-5, security, SECURITY_REVIEW.md, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m35s Details Sentinel-based cleanup: - Reviews embed <!-- review-bot:NAME --> in body (hidden HTML comment) - Cleanup matches by sentinel, not token identity - Each reviewer-name is a logical identity (sonnet, gpt, security) - Same token can run multiple review types without conflict - No extra API scopes needed System prompt file (--system-prompt-file / SYSTEM_PROMPT_FILE): - Loads a local file with additional review instructions - Appended to system base as "Additional Review Instructions" - Enables specialized reviews (security, performance, etc.) - Partially addresses #5 Security review: - SECURITY_REVIEW.md prompt focused on vulnerabilities - 3rd CI matrix entry using same token, different prompt - Focus: injection, auth, secrets, input validation, crypto, races CI changes: - REVIEWER_NAME passed from matrix.name - SYSTEM_PROMPT_FILE passed from matrix (empty for standard reviews) - 3 reviewers: sonnet (general), gpt (general), security (focused)	2026-05-01 20:55:09 -07:00
Rodin	41c670b44b	fix: post-then-cleanup flow, remove dead code, pagination CI / test (pull_request) Successful in 14s Details CI / review (gpt-4.1, gpt, GPT_REVIEW_TOKEN) (pull_request) Successful in 31s Details CI / review (gpt-5, sonnet, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m22s Details - PostReview now returns *Review (id + user login from response) - Delete flow: post first, then delete stale reviews by same user - No read:user scope needed (identity from POST response) - Removed GetAuthenticatedUser (requires scope we lack) - ListReviews: full pagination (loops until partial page) - envOrDefaultBool: case-insensitive, whitespace-trimmed - action.yml: document accepted boolean values - Tests updated for new PostReview signature	2026-05-01 20:38:21 -07:00
Rodin	0d417e068e	feat: delete previous review before posting new one (#6 ) CI / test (pull_request) Successful in 13s Details CI / review (gpt-4.1, gpt, GPT_REVIEW_TOKEN) (pull_request) Successful in 21s Details CI / review (gpt-5, sonnet, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m20s Details Before posting a review, the bot now: 1. Calls GET /api/v1/user to identify its own login 2. Lists all reviews on the PR 3. Deletes any existing reviews from itself 4. Posts the fresh review This keeps PR threads clean — one review per bot at any time. New Gitea client methods: - GetAuthenticatedUser() — token self-identification - ListReviews() — fetch reviews on a PR - DeleteReview() — delete a review by ID Flag: --update-existing / UPDATE_EXISTING (default true) Set to false to preserve old behavior (stack reviews). All delete failures are non-fatal (logged as warnings). Closes #6	2026-05-01 20:17:01 -07:00
rodin	aee903caa2	Merge pull request 'feat: add context budget system for LLM overflow (#19 )' (#20 ) from fix/19-context-overflow into main CI / test (push) Successful in 14s Details CI / review (gpt-4.1, gpt, GPT_REVIEW_TOKEN) (push) Has been skipped Details CI / review (gpt-5, sonnet, SONNET_REVIEW_TOKEN) (push) Has been skipped Details	2026-05-02 03:07:16 +00:00
rodin	2adb23b3d9	Merge pull request 'feat: add Anthropic Messages API support (#18 )' (#21 ) from feat/18-anthropic-api into main CI / test (push) Successful in 14s Details CI / review (gpt-5, sonnet, SONNET_REVIEW_TOKEN) (push) Has been skipped Details CI / review (gpt-5-mini, gpt, GPT_REVIEW_TOKEN) (push) Has been skipped Details	2026-05-02 01:57:49 +00:00
Rodin	14a0c2a946	feat: add Anthropic Messages API support (#18 ) CI / test (pull_request) Successful in 13s Details CI / review (gpt-5, sonnet, SONNET_REVIEW_TOKEN) (pull_request) Successful in 1m2s Details CI / review (gpt-5-mini, gpt, GPT_REVIEW_TOKEN) (pull_request) Successful in 1m43s Details Adds --llm-provider flag (openai\|anthropic) to switch between API formats. Anthropic implementation: - POST /messages endpoint - x-api-key + anthropic-version headers - System prompt as top-level field (not a message) - max_tokens: 8192 for response generation - Parses content blocks [{type: "text", text: "..."}] Changes: - llm/client.go: Provider type, completeAnthropic(), doRequest() shared helper - cmd/review-bot/main.go: --llm-provider / LLM_PROVIDER flag - .gitea/actions/review/action.yml: llm-provider input + env - llm/client_test.go: 4 new tests for Anthropic path Backwards compatible — default provider is still openai. Closes #18	2026-05-01 18:49:17 -07:00