$ cat ~/map/pi-goal-list-loop-audit.md
pi-goal-list-loop-audit
Mission-control extension that separates pi implementation from verification — drafted goal contracts, detached auditor workers, and a regression shield that re-checks the verifier's own evidence.
extensions✓ Oct 16
Facts
- GitHub Stars
- ★ 27
- Forks
- 15
- license
- AGPL-3.0
- npm / mo
- —
- last push
- Oct 15, 2026
install
pi install npm:pi-goal-list-loop-audit## TL;DR
GLLA is mission control for long-running autonomous work in pi — /goal, /list, and /loop surfaces with durable on-disk state, bounded recovery, and a detached auditor that must accept evidence before a completion claim can close a goal.
## Key points
- $Three work surfaces: /goal drives one outcome against a saved Done when: contract, /list maintains a durable queue of independently auditable items, and /loop runs metric, metricless-specification, or project-audit improvement styles.
- $Verification is a gate: complete_goal runs the contract's mechanical checks, writes an identity-bound claim, and starts a detached fresh pi RPC worker that must produce raw evidence for each contract item before the goal archives.
- $The auditor runs by default without extensions, skills, prompt templates, themes, or context files — independent verification, not an OS sandbox, since its bash tool can still change files.
- $Durable state (goals, queue items, progress, pauses, retries, audit claims) is inspectable on disk under the working directory's .pi-glla/ by default, with an opt-in sessionDir state root that fails closed if unresolved.
- $Recovery is bounded and visible: configurable Zero-stream retries (default 3, range 0-10), explicit /goal resume, /list resume, and /loop resume paths, and frozen-subagent escalation that never aborts the parent goal.
- $Recommended companion: @juicesharp/rpiv-ask-user-question for structured questions, multi-select, previews, and Confirm dialogs (a plain-text fallback works); optional @tintinweb/pi-subagents adds parallel workers.
- $Coexistence rules: do not run a second agent_end driver, a second task or queue extension, or overlapping compaction, retry, or watchdog supervisors in the same session — and never simultaneously with a ralph-style loop.
## FAQ
Q When should I use GLLA instead of plain pi?
For work that benefits from an autonomous operator: features spanning several files or subsystems, migrations, security reviews, repository audits, research followed by implementation, documentation or test-quality overhauls, and backlogs of independently verifiable changes. Keep plain pi for one-line edits and tightly supervised work.
Q What exactly does the auditor check?
When the agent calls complete_goal, GLLA runs any specified mechanical checks, saves a claim, then has a detached fresh pi worker inspect the repository and produce raw evidence for every Done when: item; missing evidence keeps the goal open and is shown as not recorded, never guessed.
Q Do I have to install the question companion?
No — GLLA works without @juicesharp/rpiv-ask-user-question through plain-text fallbacks, but structured questions and Confirm dialogs are the intended interview experience.
## References
## Related entries
$ pi-tmux-orchestrator
Pi extension + dependency-free Python CLI for coordinating implementer, reviewer, and optional probe agents in a monitorable tmux grid with a per-run Unix-socket broker.
$ big-pi
Model-driven compression that summarizes completed phases into tiered blocks while Pi's JSONL transcript stays authoritative and untouched.
$ dsh-pi-agent
DeepSeek Harness (DSH) AgentFactory that swaps DSH's native agent loop for a headless pi --mode rpc child, so DSH's web UI, sessions, hooks, and presets keep working with pi's minimal prompt and four tools.
Last updated: Aug 31, 2026