mempalace-triage

Maintainer triage tooling for MemPalace/mempalace.

This repo is intentionally separate from MemPalace/mempalace so the two can be cloned side-by-side without name collision in CI/remote-agent environments, and so triage tooling evolves independently of the project itself.

What's here

Path	Purpose
server/src/triage/	Heuristic classifier module (TypeScript/Bun). Fetches issues/PRs from the upstream repo, tags severity by keyword, flags noise candidates and suspicious PRs, cross-references mempalace modules. Exposed as both a library (consumed by the poller) and a CLI (writes ISSUES.md).
server/	Bun + Elysia + Drizzle + SQLite backend. Polls the triage module every 15 min, persists state, serves REST + WebSocket for the dashboard.
dashboard/	Angular 21 + Tailwind 4 frontend that talks to the server.
.claude/skills/triage-issues/SKILL.md	Deep-triage skill — drives the part heuristics can't do (semantic dedup, real severity, PR malicious review, next-action ranking). Writes TRIAGE.md.

Quickstart (local)

# Requires gh CLI, authenticated to read MemPalace/mempalace, and bun.
cd server
bun install
bun run src/triage/cli.ts              # regenerate ISSUES.md at repo root
bun run src/triage/cli.ts --audit-prs  # terminal report on flagged PRs
bun run src/triage/cli.ts --no-cache   # force fresh fetch

Then (if you have Claude Code) invoke the triage-issues skill for the deep analysis layer on top of ISSUES.md. It writes TRIAGE.md.

Quickstart (scheduled remote agent)

This repo is also cloned by a scheduled Claude Code remote agent that runs weekday mornings. The agent runs the triage CLI, reads the skill, does the deep triage, and prints the report to the trigger run logs. Manage at https://claude.ai/code/scheduled

Running the full stack (server + dashboard)

Long-running processes are managed by PM2 via ecosystem files at the repo root. See CLAUDE.md for the PM2 commands. TL;DR: pm2 start ecosystem.dev.config.cjs for dev, pm2 start ecosystem.config.cjs for prod.

Output conventions

• ISSUES.md — heuristic output from the triage CLI (regenerated every run)
• TRIAGE.md — curated deep-triage report from the skill (regenerated)
• .cache/ — fetched issue/PR/diff data, 6h TTL, gitignored

Both outputs are meant to be read and discarded — they're point-in-time snapshots, not historical records.

Tuning heuristics

All keyword banks and regex patterns live in server/src/triage/constants.ts:

• CRITICAL_KEYWORDS / HIGH_KEYWORDS — severity signals
• FEATURE_TITLE_PREFIX / BUG_TITLE_PREFIX — issue type markers
• NOISE_TITLE_PATTERNS / NOISE_BODY_PATTERNS — junk filters
• SENSITIVE_PATHS — PR path patterns that trigger red flags
• DIFF_RED_FLAGS — dangerous diff content patterns
• MEMPALACE_MODULES — module names for cross-reference

Tweak and re-run; results are deterministic given the same cache.