fix(tests): board→in_review evidence gate test expects 200 bypass (DLD-2583) - #4
Open
nydamon wants to merge 951 commits into
Open
fix(tests): board→in_review evidence gate test expects 200 bypass (DLD-2583)#4nydamon wants to merge 951 commits into
nydamon wants to merge 951 commits into
Conversation
…lish fix(ui): polish issue and agent surfaces
Co-authored-by: lockfile-bot <lockfile-bot@users.noreply.github.com>
…-runtime fix(runtime): improve agent recovery and heartbeat operations
…ny-portability feat(portability): improve company import and export flow
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Co-Authored-By: Paperclip <noreply@paperclip.ing>
…ted-agent-frontmatter Fix imported agent bundle frontmatter leakage
…-instructions-tab-reset Fix instructions tab state on agent switch
…on-aware-link-node Extract mention-aware link node helper and add tests
chore: add GitHub PR template
The plugin framework landed without updating the Dockerfile. The server now imports @paperclipai/plugin-sdk, so the deps stage needs its package.json for install and the build stage needs to compile it before building the server. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…n-sdk fix(docker): add plugin-sdk to Dockerfile build
…ping Addresses Greptile review on PR paperclipai#1363: numeric entities decode via code points; named entities use a small allowlist (amp, nbsp, etc.) so M&M resolves correctly; unknown named entities are preserved. Adds mid-token tests for & in agent names. Made-with: Cursor
- findMentionedAgents: keep normalizeAgentMentionToken + extractAgentMentionIds - decode @mention tokens with entities.decodeHTMLStrict (full HTML entities) - Add entities dependency; expand unit tests for Greptile follow-ups Made-with: Cursor
- Drop entities package (CI blocks pnpm-lock.yaml on PRs; reset lockfile to master) - Restore numeric + allowlisted named entity decoding in issues.ts - Split Greptile mid-token & case into its own test with review comment Made-with: Cursor
Made-with: Cursor
Made-with: Cursor
paperclipai#69) * fix: convert all remaining Date objects to ISO strings for postgres.js parameter binding Drizzle ORM's buildUpdateSet passes values directly to postgres.js without calling mapToDriverValue. postgres.js v3.4.8 cannot serialize Date objects as query parameters, causing TypeError crashes in .set() and .values() calls. Previous fixes only covered heartbeat.ts and execution-workspaces.ts. This commit fixes all remaining 23 server files that pass Date objects to Drizzle .set(), .values(), or .onConflictDoUpdate() operations. Critical path: costs.ts, workspace-runtime.ts, workspace-operations.ts, issues.ts, and budgets.ts are all called during heartbeat run execution and were causing agent runs to fail, leaving agents frozen. Co-Authored-By: Paperclip <noreply@paperclip.ing> https://claude.ai/code/session_01F1WkS82kZiayD8aT4zYEpr * ci: add temporary ai-review override workflow for PR paperclipai#69 Co-Authored-By: Paperclip <noreply@paperclip.ing> --------- Co-authored-by: Claude <noreply@anthropic.com> Co-authored-by: Paperclip <noreply@paperclip.ing>
…n agents Drizzle ORM's PgTimestamp.mapToDriverValue already calls value.toISOString() on Date objects before passing them to postgres.js. The prior PRs (paperclipai#62-paperclipai#69) converted `new Date()` to `new Date().toISOString()` (a string), which then failed when Drizzle called .toISOString() on the string — strings don't have that method. This caused "value.toISOString is not a function" errors during heartbeat recovery, freezing agents. Reverts all 25 affected files to the original upstream Date object pattern. https://claude.ai/code/session_01F1WkS82kZiayD8aT4zYEpr
…ents-88mhf revert: remove incorrect .toISOString() conversions causing frozen agents
…ry-timestamp fix: postgres.js Date parameter binding in process_lost retry WHERE clause
…binding in process_lost retry WHERE clause The postgres.js driver requires timestamp parameters to be ISO-formatted strings, not JavaScript Date objects. The WHERE clause was passing a Date object which caused ERR_INVALID_ARG_TYPE errors during parameter binding. Fixed by converting 'now' to 'nowStr' using toISOString() before injecting into the SQL template. This fixes the periodic 'failed to enqueue process_lost retry' errors visible in server logs.
The detectDirectDbClosures sweeper was re-flagging the same issues every 30-second sweep cycle because it checked a 60-minute window without filtering already-flagged issues. Now skips issues that already have an issue.db_bypass_detected activity log entry. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…te error messages Root cause of 80% gate failures: agents push code to GitHub but never register work products via POST /api/issues/:id/work-products. The gate blocks them, the error message doesn't tell them what API to call, and the documentation doesn't mention the endpoint at all. Changes: - Gate error messages now include the exact API endpoint and required fields - AGENTS.md Code Delivery Protocol includes full curl examples for registering commits and PRs as work products - Paperclip skill endpoint table now lists work product CRUD routes - API reference now has a full Work Products section with examples Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…te error messages (paperclipai#210) Root cause of 80% gate failures: agents push code to GitHub but never register work products via POST /api/issues/:id/work-products. The gate blocks them, the error message doesn't tell them what API to call, and the documentation doesn't mention the endpoint at all. Changes: - Gate error messages now include the exact API endpoint and required fields - AGENTS.md Code Delivery Protocol includes full curl examples for registering commits and PRs as work products - Paperclip skill endpoint table now lists work product CRUD routes - API reference now has a full Work Products section with examples Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The entrypoint drops from root to the node user via gosu, but the node user was not a member of the Docker socket's group (GID 999). This made hermes-bridge.sh unable to docker exec into the hermes-agent sidecar, causing adapter_failed errors on every Hermes heartbeat run. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…fails. Instead of silently swallowing th
…en screenshots exist (paperclipai#212) Two changes to reduce false evidence gate blocks: 1. Widened BROWSE_EVIDENCE_PATTERN to match actual agent output: goto, snapshot, gstack, playwright, VERIFIED, health score, browser verification/testing, console errors, screenshot attached. 2. Changed both engineer and QA evidence gates to accept screenshot attachments as primary evidence even without matching browse text. The text pattern is a v1 stopgap; the real proof of interactive testing is the screenshot itself. Text-only (no image) is still rejected. Audit showed done_requires_qa_browse_evidence was the #1 remaining false block — agents uploaded screenshots but their comments used goto/snapshot/gstack/playwright commands that didn't match the regex. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…perclipai#213) When a board user assigns an issue to themselves, agents (including control-plane roles like CEO/CTO) could still reassign or change status on the issue. This caused a cycle where the Monitor/CEO would escalate board-parked issues through the management chain, burning tokens and creating noise (observed on DLD-1100 with 39 comments). Board-assigned issues are at the top of the SLA chain. The new gate blocks agents from changing assignee or status when assigneeUserId is set and assigneeAgentId is null. Agents can still comment. Only the board can modify assignment or status on board-owned issues. Gate: board_assigned_protected Activity log: issue.board_assigned_protected Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…en screenshots exist (paperclipai#214) Two changes to reduce false evidence gate blocks: 1. Widened BROWSE_EVIDENCE_PATTERN to match actual agent output: goto, snapshot, gstack, playwright, VERIFIED, health score, browser verification/testing, console errors, screenshot attached. 2. Changed both engineer and QA evidence gates to accept screenshot attachments as primary evidence even without matching browse text. The text pattern is a v1 stopgap; the real proof of interactive testing is the screenshot itself. Text-only (no image) is still rejected. Audit showed done_requires_qa_browse_evidence was the #1 remaining false block — agents uploaded screenshots but their comments used goto/snapshot/gstack/playwright commands that didn't match the regex. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…rclipai#215) Every failed gate attempt bumps issue.updatedAt, which pushes the evidence time window past uploaded screenshots. DLD-2162 had 15 screenshots but was SQL-bypassed because 10+ failed attempts drifted updatedAt 30 minutes past the evidence. QA evidence gate now checks for screenshots from the QA reviewer without a time window. The QA PASS comment anchors the review cycle; screenshots from that reviewer on the issue are sufficient proof. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…aperclipai#216) Issues in "blocked" status were being marked "done" via direct SQL without the blocker being resolved. This obscures real problems — blocked work appears complete when nothing was done. Changes to detectDirectDbClosures(): - Auto-revert: when a bypass is detected and the last API-set status was "blocked", revert the issue to "blocked" and log issue.blocked_bypass_reverted - Improved dedup: checks if the most recent detection entry was created after the issue's current updatedAt, allowing re-detection after reverts (previously any historical detection caused permanent skip) - Non-blocked bypasses continue to be flagged only (may be legitimate ops closures) Triggered by DLD-2299 (poly-weather Phase 1A) which was blocked on SSH credentials, then SQL-closed as done without performing the VPS audit. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…avior (paperclipai#217) Agents using pi_local adapter resume conversations from session files. When these sessions are days old, the LLM echoes stale conclusions ("Fleet clean") instead of re-checking its inbox. The CEO had 321 stale task sessions and a 2-day-old runtime session, causing it to ignore all reopened/unblocked tasks. New sweeper expireStaleAgentSessions() runs every ~30s: - Clears runtime session pointers (agent_runtime_state.session_id) older than 4 hours, forcing fresh conversations - Prunes agent_task_sessions older than 48 hours to prevent unbounded table growth Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…aperclipai#219) * fix(sweeper): bump unpicked assignment retrigger limit from 1 to 3 MAX_RETRIGGERS=1 meant any issue that wasn't picked up on the first retrigger became permanently stranded. With agents handling multiple concurrent tasks, a single retry is insufficient — the agent may be busy on the first attempt but available 8 minutes later. Bumping to 3 gives the watchdog 24 minutes of retries (3 × 8min SLA windows) before giving up, which covers most transient busy periods. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix(test): update retrigger test to match new MAX_RETRIGGERS=3 Test seeded activationRetriggerCount=1 expecting skip, but with the limit bumped to 3, count 1 is now below threshold. Updated to 3. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…rements (paperclipai#218) * feat(governance): add route guard standards, multi-role QA, and auth PR requirements Engineering hardening across three layers — prevention, detection, enforcement — to catch auth/routing bugs before they ship. Prevention (thinking): - H1: Route access matrix standard — every web project must maintain a living docs/route-access-matrix.md mapping routes to roles/guards/sidebar - H5: Route guard rules — default-open policy, sidebar must match guards, core product pages never admin-gated Detection (testing): - H2: Role-matrix Playwright smoke test requirement for all web apps - H3: Onboarding/critical-path smoke test requirement - Definition of Done items 8-9 for auth/routing changes Enforcement (process): - H4: PR template section requiring route access matrix update and role test table for any PR touching auth/routing/sidebar - H6: QA must test as BOTH admin AND member — single-role test is automatic FAIL, with hard rejection rules and evidence format Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: resolve merge conflicts in hardening standards Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: revert accidental retrigger limit change from stash conflict Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…ipai#221) Board user comments are explicit instructions that should always reach the agent. The gate_block_backoff was unconditionally skipping wakeups after 3+ consecutive gate blocks, silently dropping board user comments. Now requestedByActorType=user bypasses the backoff and resets the gateBlockCount so the agent gets a fresh slate. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…paperclipai#220) Backend/CLI/API-only code issues have no browser surface to screenshot, causing the browse evidence gates to permanently block them. Agents were escalating every backend task to the board for manual close. Both evidence gates (in_review and done) now accept "no browser surface" as an exemption phrase in the actor's comment. The agent must include a justification (e.g. "Python backend module", "CLI tool"). The gate error messages now mention this exemption path. AGENTS.md updated with the full protocol: examples, rules, and when NOT to use the exemption. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…pai#222) The 15-minute comment retrigger cooldown was suppressing @mention wakes. When SPE @mentioned QA Agent to request a reworded QA PASS, the wakeup was skipped because QA had a recent run on the same issue. @mentions are deliberate requests for action and should always wake the target agent. Changed: cooldown only applies to issue_comment_retrigger, not issue_comment_mentioned. Updated test to expect mentions to bypass. Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add gmail.js, ga4.js, and gws-token.js to tools/google-workspace/ so they survive container rebuilds and volume wipes. The entrypoint symlinks them into /paperclip/bin/ at startup. All three are zero-dependency Node.js CLIs using service account DWD auth. - gmail.js: search, read, send, drafts, labels via Gmail API - ga4.js: GA4 reports, realtime, conversions, Measurement Protocol - gws-token.js: JWT token generator for any Google API scope - SKILL.md files for gmail and ga4-analytics skills Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Loops lifecycle and transactional email skill for agents operating the Loops platform. Includes API endpoint patterns, security model, evidence standards, and docs-first lookup protocol. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Senior Platform Engineer <spe@paperclip.local>
…ocket-permission fix(hermes): grant node user Docker socket access after privilege drop
…ce-tools feat: bake Google Workspace tools into Docker image
Fixes 404 API route not found errors affecting multiple agents (CEO, QA, Hermes) who call /api/agents/me/issues instead of /api/agents/me/inbox-lite. Added as a backward-compatible alias that accepts status and ids query params. Co-Authored-By: Paperclip <noreply@paperclip.ing>
paperclipai#227) Co-authored-by: Paperclip <noreply@paperclip.ing>
…-me-issues-rebased feat(DLD-2396): add /api/agents/me/issues alias route
… (DLD-2400) (paperclipai#230) Co-authored-by: Paperclip <noreply@paperclip.ing>
* feat(skill): add posthog analytics skill with REST API patterns Direct PostHog REST API skill — HogQL queries, events, insights, feature flags, persons, sessions, dashboards. No MCP or OAuth needed, uses personal API key auth via POSTHOG_API_KEY env var. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: wrap rg with memory safety limits to prevent OOM crashes ripgrep consumed 6.2GB RSS searching unbounded files, hitting the 12GB container mem_limit and triggering the kernel OOM killer. The wrapper enforces --max-filesize 10M, --max-columns 10000, and --max-count 500. Same relocation pattern as gh and docker wrappers. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: make Google Workspace CLIs ESM-compatible * test: align gate specs with current issue flow --------- Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
DLD-2560: Port the 13-skill pack (7 in skills/ + 6 in .agents/skills/) into /app/skills/ so the pi-local adapter can discover all skills at container startup. Changes: - Dockerfile: COPY 6 skills from .agents/skills/ into /app/skills/ (company-creator, create-agent-adapter, doc-maintenance, pr-report, release, release-changelog) - Dockerfile: COPY and chmod docker-entrypoint.sh + verify-skills.sh - Dockerfile: Use ENTRYPOINT to call docker-entrypoint.sh - scripts/verify-skills.sh: NEW — verifies all 13 skills have SKILL.md - scripts/docker-entrypoint.sh: call verify-skills.sh before server start The pi-local adapter resolves skills via PAPERCLIP_SKILL_ROOT_RELATIVE_CANDIDATES = ["../../skills", "../../../../../skills"] which maps to /app/skills/ inside the container. The explicit COPY ensures all 13 skills are present regardless of .dockerignore settings.
…bypass) DLD-2583: PR Verify has been failing 3x in 24h due to this test expecting 422 when assertEngineerBrowseEvidence now bypasses for non-agent actors (req.actor.type !== 'agent'). The board-user bypass was intentionally restored in f422c4f (DLD-2400). Board users (req.actor.type === 'board') should be able to move code issues to in_review without providing browse evidence — the board is a governance tool, not a code-delivery actor. The gate correctly fires for agents. Test update: expect 200 (gate bypassed) instead of 422.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d738cb67c3
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Comment on lines
+229
to
+233
| const { useState, useEffect, useCallback, useMemo, useRef, useContext, | ||
| createContext, createElement, Fragment, Component, forwardRef, | ||
| memo, lazy, Suspense, StrictMode, cloneElement, Children, | ||
| isValidElement, createRef } = R; | ||
| export { useState, useEffect, useCallback, useMemo, useRef, useContext, |
There was a problem hiding this comment.
The generated react shim only exposes a hardcoded subset of React exports, so any plugin UI bundle that imports other standard names (for example useLayoutEffect, useReducer, or startTransition) will fail module linking with a missing export error and never render its slots. This is a runtime break for otherwise valid React plugin code; the shim needs to forward the complete React export surface instead of a curated list.
Useful? React with 👍 / 👎.
Comment on lines
+305
to
+307
| result = result.replaceAll(` from ${from}`, ` from ${to}`); | ||
| // Also handle `import "..."` (side-effect imports) | ||
| result = result.replaceAll(`import ${from}`, `import ${to}`); |
There was a problem hiding this comment.
The bare-specifier rewrite only matches exact substrings with spaces ( from ... and import ...), but minified ESM commonly emits from"react"/import"react" without whitespace. In that case the bare imports are left unchanged, and the browser fails to resolve them when loading plugin modules. This makes plugin loading brittle for minified bundles; rewrite logic should tolerate optional whitespace (or parse imports structurally).
Useful? React with 👍 / 👎.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
7 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Root Cause
DLD-2583: PR Verify failing 3x in 24h on Viraforge/paperclip
Failing test:
board → in_review, code issue, no evidence → 422Location:
server/src/__tests__/browse-evidence-gate.test.ts:272Root cause: Commit
f422c4fe("restore board-user bypass in assertEngineerBrowseEvidence") addedreq.actor.type !== 'agent'to the evidence gate, intentionally allowing board users to bypass the browse evidence gate. The test was never updated to reflect this.What the test expected: 422 (gate fires for board, blocking without evidence)
What the code does: 200 (gate bypasses for non-agent actors incl. board users)
Fix
Update the test to expect 200 and rename it to
board → in_review, code issue, no evidence → 200 (bypass).Verification
npx vitest run server/src/__tests__/browse-evidence-gate.test.ts→ 17/17 tests passDLD-2583