Loop Engineering in Codex

## Agent loop policy
 
For any multi-pass task:
 
1. Create or update docs/agent-runs/<task-slug>.md before editing.
2. Record objective, scope, verifier, pass log, risk, and escalation conditions.
3. Make one bounded pass at a time.
4. Run the verifier after each pass.
5. Stop instead of widening scope silently.
 
Required evidence:
 
- changed files
- commands run
- verifier result
- fresh review result
- remaining risk
 
Forbidden shortcuts:
 
- Do not delete, skip, or weaken tests to make a loop pass.
- Do not mute logs or alerts to reduce noise.
- Do not broaden catch blocks without root-cause evidence.
- Do not touch auth, payments, secrets, migrations, data deletion, or deployment scripts without an explicit owner approval note.
 
Review guidelines:
 
- P0: data loss, auth bypass, secret exposure, payment risk, destructive migration.
- P1: production regression, missing regression test, broken critical path.
- P2: maintainability risk that should be fixed before merge.
- P3: style, clarity, or small cleanup issue.

[profiles.loop_readonly]
sandbox_mode = "read-only"
approval_policy = "never"
 
[profiles.loop_workspace]
sandbox_mode = "workspace-write"
approval_policy = "on-request"
 
[profiles.loop_unattended]
sandbox_mode = "workspace-write"
approval_policy = "never"
 
# Use loop_unattended only for reversible work with deterministic checks,
# no production secrets in scope, and an explicit state file.

/goal Ship the private-beta invite request slice.
 
Goal:
A user can request a private-beta invite, see the saved pending state, and see a duplicate-request error for an existing email.
 
Read first:
- AGENTS.md
- docs/product/private-beta.md
- app/invite/**
- lib/db/**
- current diff
 
State:
- Maintain docs/agent-runs/private-beta-invite-loop.md.
 
Scope:
- In scope: invite UI, API handler, persistence, duplicate-request state, tests.
- Out of scope: billing, team roles, email provider, admin dashboard.
 
Verifier:
- npm test
- npm run lint
- npm run build
- browser check for happy path and duplicate-request error
 
Done when:
- The user path works end to end.
- The verifier passes.
- /review or a fresh reviewer reports no P0/P1 issue.
- The state file lists changed files, commands, screenshots, and remaining risk.
 
Escalate when:
- Auth/session behavior changes.
- A migration, external provider, or product decision is needed.
- The same failure repeats twice.

# Private-beta invite loop
 
## Objective
A user can request a private-beta invite, see the saved pending state, and see duplicate-request feedback.
 
## Scope
- In scope: invite UI, API handler, persistence, duplicate-request state, tests.
- Out of scope: billing, roles, email provider, admin dashboard.
 
## Verifier
`npm test && npm run lint && npm run build`
 
Browser checks:
- /invite accepts a new request.
- /invite shows duplicate-request feedback for an existing email.
 
## Pass log
| pass | hypothesis | files changed | verifier result | evidence | decision |
| --- | --- | --- | --- | --- | --- |
| 1 | | | | | |
 
## Review log
| reviewer | scope | result | blocking findings |
| --- | --- | --- | --- |
 
## Risk register
| risk | owner | trigger | next check |
| --- | --- | --- | --- |
 
## Final evidence
- Changed files:
- Commands:
- Screenshots:
- Remaining risks:

---
name: backlog-loop
description: Use when triaging and repairing a scoped batch of backlog or production-error issues with a state file, verifier, and review gate.
---
 
# Backlog loop skill
 
Use this procedure only for a scoped issue batch.
 
## Required inputs
 
- issue query or issue list
- repo instructions
- state file path under docs/agent-runs/
- verifier command
- review gate
 
## Procedure
 
1. Read the state file and nearest repo instructions.
2. Sort issues by impact, age, and reproducibility.
3. Pick one issue.
4. Reproduce it with a failing test, script, log query, or explicit non-repro note.
5. Patch root cause only.
6. Add or update a regression check.
7. Run the verifier.
8. Update the issue with evidence.
9. Ask for fresh review before closing.
 
## Stop conditions
 
- verifier passes and fresh review has no P0/P1 finding
- same failure repeats twice
- owner decision is needed
- scope expands beyond the batch

Name: Daily production-error sweep
Cadence: Weekdays at 09:00
Worktree: Fresh automation worktree
Default profile: loop_readonly
 
Prompt:
1. Read AGENTS.md and docs/agent-runs/production-error-loop.md.
2. Pull the top recurring error family from the configured issue or log source.
3. Identify count, impacted path, owner, suspected root cause, and recent related commits.
4. If the fix is low-risk and scoped, create a plan and wait for approval before editing.
5. If the fix is not clear, update the run log and open or update the issue with evidence.
 
Stop when:
- One prioritized error family has evidence and owner.
- Duplicate issues are linked.
- No code was changed without approval.
 
Escalate when:
- Customer data, auth, payments, secrets, deletion, migration, or deployment risk is in scope.
- The issue source cannot be queried.
- The same error was triaged twice without an owner decision.

Run the manager operating loop.
 
Read:
- AGENTS.md
- docs/agent-runs/manager-operating-loop.md
- open PRs older than two business days
- failed CI checks
- unresolved incident follow-up items
- changed code paths without docs updates
 
Classify each item:
- healthy
- blocked
- owner missing
- verifier missing
- stale decision
 
Output:
| item | owner | status | latest evidence | next observable check | escalation date |
| --- | --- | --- | --- | --- | --- |
 
Rules:
- Do not edit code.
- Create docs-only draft changes only when the missing update is obvious.
- Every escalation must name the owner or the owner gap.

/review Focus on:
 
- P0/P1 regressions
- missing regression tests
- auth, payments, secrets, migration, deletion, and deployment risk
- verifier weakening
- scope expansion beyond docs/agent-runs/private-beta-invite-loop.md
- mismatch between the final summary and the actual diff