tasklist-agent

A local, always-on task dashboard at http://tasklist that reads your Slack, turns it into a deduplicated task list you can check off / comment on in the browser, lets you ask Claude for help per-task, and pushes a daily digest to Discord.

No paid API. All the "intelligence" runs through headless Claude Code (claude -p), billed to your existing Claude Code subscription — never the Anthropic API. Slack's Web API and Discord webhooks are free.

How it works

A Node + Express server (bound to 127.0.0.1, kept alive by macOS launchd) owns a SQLite store and serves one HTML page. It has no LLM logic itself — it spawns claude -p for three jobs, and those runs talk back over the localhost API:

• Ingest — every 30 min: reads your Slack (public/private channels, DMs, mentions) via your user token, extracts action items, upserts them as tasks (deduped by channel:message-ts).
• Reply — on demand: comment @claude … on a task → a claude -p run answers in that task's thread (and can search Slack / draft for you). The page polls and shows it.
• Digest — daily 08:00: summarizes tasks + Slack highlights and posts to Discord.

Prerequisites

• macOS, Node ≥18, and the claude CLI installed and logged into your subscription.
• Slack user token (xoxp-…): create a Slack app → User Token Scopes: channels:history, groups:history, im:history, mpim:history, channels:read, groups:read, im:read, mpim:read, users:read, search:read → install to your workspace → copy the token.
• Discord webhook URL: target channel → Integrations → Webhooks → New Webhook → copy URL.

Setup (one-time)

cd ~/projects/tasklist-agent
npm install
cp .env.example .env        # then edit: SLACK_USER_TOKEN=..., DISCORD_WEBHOOK_URL=...
./setup.sh                  # writes /etc/hosts + pf redirect (sudo), installs the launchd agent

setup.sh also pins an absolute CLAUDE_BIN and TASKLIST_API=http://tasklist into .env, and a working PATH into the launchd plist so the always-on service can find claude/node. Then open http://tasklist.

Use http://tasklist, not :8787. Once the pf redirect (80 → 8787) is active, macOS blackholes direct connections to 127.0.0.1:8787 — so http://tasklist:8787 and localhost curls hang. Everything (browser, and the agent's own API calls via TASKLIST_API) goes through http://tasklist. If you skip the pf step entirely, http://tasklist:8787 works as the direct fallback.

Updating (after editing code)

There's no build step and the launchd agent loads the code once at startup, so new code only takes effect on restart:

npm run update              # restarts the launchd agent, re-reading server.js from disk
npm run update -- --kill-strays   # also stop any manual instances left on other ports

No sudo needed. Use ./setup.sh instead only when the launchd plist, hosts entry, or pf redirect itself changed.

First-run checklist (important — tests pass with a stub, so verify the real path)

The automated suite exercises a stubbed claude; these only surface on a live run:

1. Tool permission. Runs are spawned with --allowedTools Bash so claude -p can curl. Sanity-check once by hand: SLACK_USER_TOKEN=… TASKLIST_API=http://127.0.0.1:8787 claude -p "$(node -e 'import("./src/prompts.js").then(m=>console.log(m.ingestPrompt({apiBase:"http://127.0.0.1:8787"})))')" --allowedTools Bash --output-format json and confirm it actually hits Slack + posts tasks.
2. Run under launchd, not just your shell — PATH differs. Check launchctl list | grep tasklist shows a PID, and watch logs/err.log.
3. Slack mentions query in src/prompts.js (search.messages?query=to:me) may need your real handle/user-id; adjust if mentions don't show up.
4. Watch the Open list over the first hour or two for duplicates (dedup is keyed on channel + message-ts).

Tunables

• Who you are / what Slack is relevant to you: context.md (gitignored; copy of context.example.md). The ingest agent reads it every run — no restart needed — to decide which messages (direct mentions, @channel/@here, and role-relevant threads even when you're not tagged) become tasks.
• Schedules: src/cron.js (*/30 * * * * ingest, 0 8 * * * digest).
• What counts as an action item / digest format: src/prompts.js.
• Run timeouts: INGEST_TIMEOUT_MS / DIGEST_TIMEOUT_MS in .env (default 10 min).
• Port: PORT in .env (default 8787).

Uninstall

./teardown.sh    # unloads launchd agent, removes hosts entry + pf redirect

Notes & known follow-ups

• Dedup key is channel:message-ts (not the task title). This deliberately deviates from the original spec's title-based fingerprint: titles are LLM-generated and change each ingest, which would flood duplicates. Trade-off: two action items in the same Slack message collapse to one task.
• The localhost API has no auth — fine for a single-user 127.0.0.1 tool. Don't rebind to 0.0.0.0 without adding auth (a @claude comment spawns a subscription-billed run).
• Minor deferred: on a catastrophic SQLite error (disk full/corruption) a lock acquire can throw unhandled in the cron/reply callbacks; it self-heals on the next launchd restart (boot reconcile clears stale locks). Wrap those callbacks in .catch() when convenient.

Full design + build history: docs/superpowers/specs/ and docs/superpowers/plans/.