@sentry/warden 0.28.0 → 0.29.0

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@sentry/warden",
-  "version": "0.28.0",
-  "description": "Event-driven agent that reacts to GitHub events and executes skills via Claude Code SDK",
+  "version": "0.29.0",
+  "description": "Event-driven agent that reacts to GitHub events and executes code-review skills",
   "type": "module",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -37,6 +37,8 @@
   "dependencies": {
     "@anthropic-ai/claude-agent-sdk": "^0.2.22",
     "@anthropic-ai/sdk": "^0.72.1",
+    "@earendil-works/pi-ai": "^0.74.0",
+    "@earendil-works/pi-coding-agent": "^0.74.0",
     "@inquirer/select": "^5.0.4",
     "@octokit/rest": "^22.0.1",
     "@sentry/dotagents-lib": "^1.16.1",
@@ -73,7 +75,7 @@
     "cli": "tsx src/cli/index.ts",
     "clean:dist": "rm -rf dist",
     "build": "pnpm run clean:dist && tsc -p tsconfig.build.json",
-    "build:action": "ncc build src/action/main.ts -o dist/action --no-source-map-register --license licenses.txt",
+    "build:action": "rm -rf dist/action && ncc build src/action/main.ts -o dist/action --no-source-map-register --license licenses.txt",
     "dev": "tsc --watch",
     "lint": "oxlint src",
     "lint:fix": "oxlint --fix src",

package/skills/warden/SKILL.md

@@ -39,7 +39,7 @@ warden --fix
 warden --fail-on high
 ```
 
-Set `WARDEN_ANTHROPIC_API_KEY` or log in via `claude login` before running.
+Set `WARDEN_MODEL` and the WARDEN-prefixed provider API key for that model before running.
 
 ## Pre-Commit Workflow
 

package/skills/warden/references/config-schema.md

@@ -25,7 +25,8 @@ version = 1                    # Required, must be 1
 
 ```toml
 [defaults]
-model = "claude-sonnet-4-5"           # Legacy default analysis model
+runtime = "pi"                        # Default runtime
+model = "openai/gpt-5.5"              # Legacy default analysis model
 maxTurns = 50                         # Legacy default analysis turns
 defaultBranch = "main"                # Base branch for comparisons
 failOn = "high"                       # Exit 1 if findings >= this severity
@@ -35,15 +36,15 @@ reportOnSuccess = false               # Post report even with no findings
 ignorePaths = ["*.test.ts"]           # Exclude matching files
 
 [defaults.agent]
-model = "claude-sonnet-4-5"           # Default repo-aware analysis model
+model = "openai/gpt-5.5"              # Default repo-aware analysis model
 maxTurns = 50                         # Max agentic turns per hunk
 
 [defaults.auxiliary]
-model = "claude-haiku-4-5"            # Helper model for extraction and fix gates
+model = "anthropic/claude-haiku-4-5"  # Helper model for extraction and fix gates
 maxRetries = 5                        # Retries for auxiliary structured calls
 
 [defaults.synthesis]
-model = "claude-opus-4-5"             # Consolidation and generated-skill build model
+model = "anthropic/claude-opus-4-5"   # Consolidation and generated-skill build model
 
 [defaults.chunking]
 enabled = true                 # Enable hunk-based chunking
@@ -70,7 +71,7 @@ paths = ["src/**"]             # Include only matching files
 ignorePaths = ["**/*.test.ts"] # Exclude matching files
 
 # Optional overrides (inherit from defaults if not set)
-model = "claude-opus-4-5"
+model = "anthropic/claude-opus-4-5"
 maxTurns = 100
 failOn = "high"
 reportOn = "medium"
@@ -121,8 +122,9 @@ Always skipped (cannot be overridden):
 
 | Variable | Purpose |
 |----------|---------|
-| `WARDEN_ANTHROPIC_API_KEY` | Claude API key (required) |
 | `WARDEN_MODEL` | Default model (lowest priority) |
+| `WARDEN_OPENAI_API_KEY` | OpenAI API key for OpenAI Pi models |
+| `WARDEN_ANTHROPIC_API_KEY` | Anthropic API key for Anthropic Pi models or Claude runtime |
 | `WARDEN_STATE_DIR` | Override cache location (default: `~/.local/warden`) |
 | `WARDEN_SKILL_CACHE_TTL` | Cache TTL in seconds for unpinned remotes (default: 86400) |
 

package/skills/warden/references/configuration.md

@@ -20,7 +20,7 @@ The `name` field references a skill you've created (via `warden add`) or defined
 version = 1
 
 [defaults.agent]
-model = "claude-sonnet-4-5"
+model = "openai/gpt-5.5"
 
 [[skills]]
 name = "my-skill"           # matches .agents/skills/my-skill/SKILL.md
@@ -58,7 +58,7 @@ actions = ["opened", "synchronize"]
 ```toml
 [[skills]]
 name = "my-skill"
-model = "claude-opus-4-5"
+model = "anthropic/claude-opus-4-5"
 maxTurns = 100
 paths = ["src/auth/**", "src/payments/**"]
 failOn = "high"
@@ -108,8 +108,9 @@ From highest to lowest priority:
 
 | Variable | Purpose |
 |----------|---------|
-| `WARDEN_ANTHROPIC_API_KEY` | Claude API key (required unless using Claude Code subscription) |
 | `WARDEN_MODEL` | Default model (lowest priority) |
+| `WARDEN_OPENAI_API_KEY` | OpenAI API key for OpenAI Pi models |
+| `WARDEN_ANTHROPIC_API_KEY` | Anthropic API key for Anthropic Pi models or Claude runtime |
 | `WARDEN_STATE_DIR` | Override cache location (default: `~/.local/warden`) |
 | `WARDEN_SKILL_CACHE_TTL` | Cache TTL in seconds for unpinned remotes (default: 86400) |
 

package/src/builtin-skills/code-review/SKILL.md

@@ -51,6 +51,7 @@ No proof, no finding. Suspicion is not a result.
 | Boundaries and edge cases | Empty, first, last, duplicate, pagination, sorting, timezone, locale, precision, overflow, migration, or compatibility cases produce wrong behavior. |
 | Persistence and migrations | Writes are non-atomic, migrations lose data, backfills skip rows, query filters update the wrong records, or rollback paths leave inconsistent state. |
 | API and dependency behavior | Published interfaces, CLI flags, config options, webhooks, service calls, or third-party dependency changes break documented or existing caller behavior. |
+| Public metadata and routing config | Robots rules, sitemaps, manifests, redirects, cache headers, or route config make documented public entry points unreachable, stale, or undiscoverable. |
 | UI correctness | The UI displays stale, wrong, duplicate, missing, or unsaved data because of the changed code, not because of style or preference. |
 | Build, test, and workflow breakage | Changed code, packaging, imports, exports, generated artifacts, CI, or release workflows fail deterministically or report false success. |
 
@@ -58,11 +59,12 @@ No proof, no finding. Suspicion is not a result.
 
 | Level | Use For |
 |-------|---------|
-| high | Data loss or corruption, critical-path crashes, broken production deploy or release, incorrect billing or permissions state, published interface breakage for normal callers, deadlock or hang in core flow, or false success after a failed destructive operation. |
+| high | Data loss or corruption, critical-path crashes, broken production deploy or release, incorrect billing or permissions state, published interface breakage for normal callers, public metadata/config that blocks normal discovery or reachability of shipped endpoints, deadlock or hang in core flow, or false success after a failed destructive operation. |
 | medium | Reproducible wrong results, recoverable crashes, duplicate or missed side effects, broken non-critical workflow, meaningful edge case in a shipped path, or compatibility break with a clear affected caller. |
-| low | Narrow but real bug with limited blast radius, confusing state that can cause user-visible mistakes, or a test/tooling bug that masks the intended behavior. |
+| low | Narrow but real bug with limited blast radius, confusing state that can cause user-visible mistakes, or a test/tooling bug that masks only a narrow non-shipped behavior. |
 
 - Use the lower severity when impact depends on unproven preconditions.
+- Score test and golden-file findings by the shipped behavior they authorize or hide, not by the file type. A test that locks in a high-impact production breakage is high severity.
 - Do not inflate severity for cleverness. The bug earns its level through impact.
 
 ## What Not To Report

package/src/builtin-skills/code-review/SOURCES.md

@@ -8,6 +8,7 @@
 | Local `architecture-review` and `code-simplifier` skills | Local prior art | Medium | Use boundaries to avoid overlap with style and architecture feedback. | Make exclusions explicit so code review reports only bugs. |
 | Common production correctness review failure modes | Engineering practice | Medium | Keep examples generic, transformed, and evidence-gated. | Cover logic, state, async, error handling, contracts, persistence, UI, build, and workflow bugs. |
 | Warden package skill loader and packaging tests | Repository behavior | High | Use for validation scope only. | Add tests that the new built-in skill resolves and is included in the npm package. |
+| GitHub issue #317 severity feedback | Maintainer issue feedback | High | Use only the generalized severity lesson; do not encode repo-specific private context. | Score tests, golden files, and config bugs by production blast radius when they hide or authorize shipped breakage. |
 
 ## Coverage Matrix
 
@@ -16,7 +17,7 @@
 | Bug class definitions and prerequisites | Complete | `SKILL.md` defines reportable categories and requires a concrete trigger, violated contract, and impact. |
 | Reachability and reproducibility evidence | Complete | Findings must prove the changed path is reachable and the failure is triggered by a specific input, state, ordering, or configuration. |
 | False-positive controls | Complete | Exclusions block style, architecture, maintainability, security, performance-only, test-coverage-only, and pattern-only findings. |
-| Severity and confidence calibration | Complete | Severity is tied to user-visible or operational impact, with lower severity for unproven preconditions. |
+| Severity and confidence calibration | Complete | Severity is tied to user-visible or operational impact, including high-impact public discovery/config breakage, with lower severity for unproven preconditions. |
 | Remediation expectations | Complete | Findings may include `suggestedFix` only when the fix is complete for the analyzed path. |
 | Language and workflow caveats | Complete for initial scope | References cover JS/TS, Python, and GitHub Actions workflow correctness. Other languages use the core contract until repeated examples justify references. |
 
@@ -34,6 +35,9 @@
 4. Build and workflow breakage are correctness bugs.
    - Reason: Warden reviews code changes that can break published packages, CI, and releases.
    - Decision: include deterministic build, packaging, test, and workflow failures when the changed code proves the failure.
+5. Tests and golden files inherit severity from the shipped behavior they protect.
+   - Reason: a changed test can lock in broken output for a public route or config surface while looking like a low-impact test-only edit.
+   - Decision: do not automatically classify test/tooling findings as low when they mask or authorize high-impact production breakage.
 
 ## Open Gaps
 

package/src/builtin-skills/code-review/SPEC.md

@@ -10,7 +10,7 @@ It should catch real regressions and production defects while preferring no find
 
 In scope:
 
-- Logic, state, async, data contract, persistence, edge-case, API, UI, build, test, and workflow correctness bugs.
+- Logic, state, async, data contract, persistence, edge-case, API, public metadata/config, UI, build, test, and workflow correctness bugs.
 - Changed production code, user entry points, public APIs, shipped workflows, and tests that can mask shipped regressions.
 - General bug-finding guidance that applies across application code.
 - Focused notes for JavaScript/TypeScript, Python, and GitHub Actions workflow correctness.
@@ -89,6 +89,7 @@ Data that must not be stored:
 - Holdout examples:
   - A style-only diff with no behavior change should produce no finding.
   - A security-only diff should be routed to `security-review`, not reported by `code-review`.
+  - A test or golden fixture that locks in broken shipped behavior should be scored by the shipped behavior's blast radius, not by the file type.
   - A type-safe or schema-validated path should not be reported as a runtime bug.
 - Acceptance gates:
   - Findings require trigger, expected behavior, actual behavior, and impact.