Autonomous task-execution loop that pulls actionable issues from a task tracker (Linear, Jira, Plane, or GitHub Issues), branches from main, implements, tests, opens a GitHub PR with no AI attribution, and self-paces the next iteration. Use when the user runs `/taskloop init` (one-time per-project setup that picks a provider and generates the loop's memory/config files), `/taskloop start` (re-arms the Monitor + ScheduleWakeup so the loop resumes for the current project), `/taskloop stop`, `/taskloop sync` (hot-repairs drifted/corrupt/missing memory files and adopts new skill rules without stopping the loop), `/taskloop add-rule [slug]`, or `/taskloop status`. Also triggers when the user says "start" / "porneste" / "resume loop" in a working directory that already has a taskloop config under `~/.claude/projects/SLUG/memory/`.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Autonomous task-execution loop that pulls actionable issues from a task tracker (Linear, Jira, Plane, or GitHub Issues), branches from main, implements, tests, opens a GitHub PR with no AI attribution, and self-paces the next iteration. Use when the user runs `/taskloop init` (one-time per-project setup that picks a provider and generates the loop's memory/config files), `/taskloop start` (re-arms the Monitor + ScheduleWakeup so the loop resumes for the current project), `/taskloop stop`, `/taskloop sync` (hot-repairs drifted/corrupt/missing memory files and adopts new skill rules without stopping the loop), `/taskloop add-rule [slug]`, or `/taskloop status`. Also triggers when the user says "start" / "porneste" / "resume loop" in a working directory that already has a taskloop config under `~/.claude/projects/SLUG/memory/`.
Taskloop
Overview
Taskloop wires a task tracker (Linear, Jira, Plane, GitHub Issues) to a coding agent in a tight feedback loop. Each iteration pulls one "ready" issue, branches, implements the change with full test verification, opens a PR, and transitions the issue to review โ fully autonomous, no AI attribution on commits/PRs. The loop self-paces via a Monitor background poll plus a ScheduleWakeup safety heartbeat.
Configuration is per-project (lives under ~/.claude/projects/<slug>/memory/) so the same skill works for any repo. The provider (Linear / Jira / Plane / GitHub Issues) is chosen once at init time; the loop spec and feedback rules are provider-agnostic.
Subcommand routing
The skill is invoked with an args string. Parse the first whitespace-delimited token:
First token
Action
Workflow guide
init (or empty args)
One-time per-project setup
references/workflows/init.md
start
Re-arm Monitor + ScheduleWakeup for current project
references/workflows/start.md
stop
TaskStop monitor, no further wakeups
references/workflows/stop.md
sync (--check for dry-run)
Hot-repair drifted/corrupt/missing memory files + adopt new skill rules, without stopping the loop; then hot-reload in-flight agents
references/workflows/sync.md
add-rule <slug>
Add a project-specific feedback rule
references/workflows/add-rule.md
status
Print monitor task id, last actionable issue, last PR
references/workflows/status.md
anything else
Print usage and stop
inline below
If the args is empty, default to init only when the project's memory dir doesn't exist yet. Otherwise default to start.
The loop spec calls these capabilities by name; swapping providers is a one-config-file change.
Provider details live in references/providers/<name>.md. Read only the one matching the chosen provider โ never load all of them.
Where things live
Per-project, under ~/.claude/projects/<slugified-cwd>/memory/:
.env # API keys, project IDs, LOOP_STRATEGY, LOOP_AUTOMERGE, chmod 600
.render-manifest.json # every placeholder value used to render this dir; lets `sync` re-render deterministically, chmod 600
.monitor-state # ephemeral, last queue snapshot
.pr-feedback-state # PR registry (repo+number+issue+merge lane+gatedHead+ownActivity, written at gh pr create) + per-PR last-seen feedback timestamps; drives sweep/merge/WIP-cap; survives stop/start
.queue-plan # last LOOP-PLAN triage + queue signature (ids/titles/desc-hashes/human-comment counts โ never the loop's own plan comments)
monitor.sh # rendered from provider template
monitor-prs.sh # the loop's own eyes on its PRs (merge/comment/CI deltas)
reconcile-merged.sh # level-triggered repair of tickets the PR watcher's edge trigger missed (Plane, GitHub)
review-recall.sh # cron: periodic ping listing review-ready PRs still unmerged
hook-agent-routing.mjs # PreToolUse gate on Task|Agent: upgrades under-modeled [LOOP-AGENT] spawns + writes the spawn ledger
hook-loop-guards.mjs # PreToolUse gate on Bash: denies a backgrounded gate inside a subagent, and a commit carrying AI attribution
usage-rollup.mjs # sums finished subagent transcripts into telemetry/usage-*.jsonl
targeted-regate.mjs # Workflow script for fix-round re-gates (targeted-regate-on-fixes rule)
telemetry/spawns-YYYY-MM.jsonl # append-only: issue, tier, strategy, requested โ routed model, effort (hook-written)
telemetry/usage-YYYY-MM.jsonl # append-only: per-agent token totals joined to the spawn that caused them
telemetry/guards-YYYY-MM.jsonl # append-only: every tool call a guard denied, and why
MEMORY.md # index โ loaded into every conversation
agent-brief.md # condensed startup brief spawned agents read INSTEAD of the full memory set
autonomous-loop.md # loop spec (provider-agnostic)
restart-instructions.md # "start" recipe
model-usage-log.md # routing policy + measured gate costs (the per-task ledger lives in telemetry/)
feedback/<slug>.md # universal + project-specific rules
Telemetry is written by mechanism, never by convention. The hook cannot forget to record a spawn, and the roll-up cannot forget what an agent cost. Both files are append-only and idempotent. Nothing in the loop reads them back โ auto decides a rung from the ticket's LOOP-PLAN marker alone, so a missing or rotated log can never change what the loop does. The telemetry is for the owner, after the fact.
MEMORY.md is the only file guaranteed-loaded; everything else is linked from it and pulled in as needed.
First-pass workflow
Parse args. Pick subcommand per the table above.
Compute project memory dir: ~/.claude/projects/<slug>/memory/ where <slug> = absolute CWD with / โ - (matches the harness's --Users-foo-bar convention).
Read the relevant references/workflows/*.md for step-by-step instructions and execute.
Use AskUserQuestion for any decision the workflow doesn't already pin (provider choice, secrets, project name resolution).
Never commit secrets. Files containing LINEAR_API_KEY / JIRA_API_TOKEN are chmod 600 and never staged to git.
Universal feedback rules
39 rules ship as assets. init copies them into the project's memory/feedback/ and links each one from MEMORY.md, grouped by when it applies. They are provider-neutral.
Prose is the weakest form of enforcement. Every rule that was empirically ignored in a real run has been converted into a mechanism, and new rules should be too. What is mechanical today:
Mechanism
Enforces
hook-agent-routing.mjs (PreToolUse on Task|Agent)
the model/effort ladder; also writes the append-only spawn ledger
hook-loop-guards.mjs (PreToolUse on Bash|ScheduleWakeup)
gates-run-foreground, no-ai-attribution (commits AND PR bodies), pr-visual-evidence, dedupe-before-filing, never-self-halt-loop (a self-issued stop:true is denied), and the merge-safety floor (a draft, red-check, or unapproved-non-auto-lane gh pr merge is denied)
targeted-regate.mjs schema
a review finding without a verbatim quote cannot exist
monitor-prs.sh / reconcile-merged.sh (both armed as Monitors at start)
merge, comment and CI events arrive as signals, not as chat; the reconcile is level-triggered repair, actually scheduled
usage-rollup.mjs
the spend record; the loop cannot forget to write it
GitHub
a draft PR cannot be merged
The rules, by when they apply (see assets/feedback/):
Gating a PR:pr-gate-passes (the four lenses: /review, rigorous, security, QA agent), ui-verification-and-evidence, cost-conscious-gating, targeted-regate-on-fixes, ci-failure-triage
After the PR is open:watch-pr-comments, read-media-in-comments, merge-on-approval, automerge-risk-lanes, ticket-done-on-merge, pr-registry-race, reconcile-tracker-vs-github, telegram-notify-owner
The promotion bar for a new rule (continuous-learning): it is kept only if it is verified, names the failure pattern concretely, and records the alternative that was ruled out. The rule set is read every tick, so every weak rule is a permanent tax on every future iteration. Prefer strengthening an existing rule over adding a sibling; prefer a hook over a paragraph.
Adding a new provider
To support a new tracker (Asana, ClickUp, Shortcut, โฆ):
Implement the 6-verb contract documented in references/adding-a-provider.md.
Create references/providers/<name>.md with env, queries, and recipes.
Add a monitor-<name>.sh.tmpl to assets/.
Add the provider to the init AskUserQuestion options.
Optionally add a reconcile-merged-<name>.sh.tmpl (see the Plane one). Without it the reconcile-tracker-vs-github rule still applies โ the loop just runs the reconcile inline each sweep through the provider's transition recipe instead of as a background script.
The loop spec template is already provider-neutral; only the adapter changes.
Key constraints
Per-project memory โ never write to another project's memory dir from this one. Slug the CWD precisely.
Idempotency โ running init twice on the same project should detect existing config and ask whether to overwrite individual files, not silently clobber.
No AI attribution anywhere โ neither in skill output (commit messages, PR bodies) nor in scaffolded files. The bundled feedback rule enforces this on the loop, but the skill itself must also obey it when writing commits via git.
Reference index
Read these only when the workflow points to them:
references/workflows/init.md โ full init interview + file generation
