Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
Orchestrates issue-driven agent loops that post cycle status to issue threads and incorporate human feedback in each cycle.
platforms
["all"]
requires
[{"issue":"issue number or URL to work on in the loop"},{"work":"description of the work to perform each cycle (the actual task — bug fix, feature, doc update)"}]
ensures
[{"cycle-status":"structured status comment posted to issue thread after every cycle"},{"thread-scan":"all human comments classified and incorporated into subsequent cycles"},"if resolved: completion summary comment posted with checklist confirmation"]
errors
[{"tracker-unavailable":"cannot read or post comments to issue tracker"},{"max-cycles-reached":"issue not resolved within cycle limit; escalation comment posted"}]
invariants
["every cycle posts a status comment before returning, even partial cycles","human feedback from any cycle is always acknowledged in the next status comment","completion is only declared when all issue acceptance criteria are met"]
Skill access pattern (post-kernel-pivot, 2026.5+)
Skill names referenced in this document are AIWG skills, not slash commands. Most are not kernel-listed and cannot be invoked as /skill-name by the platform. Reach them via:
aiwg discover "<capability>"
aiwg show skill <name>
Only kernel-listed skills (aiwg-doctor, aiwg-refresh, aiwg-status, aiwg-help, use, steward) are directly invokable as slash commands. See skill-discovery rule.
Issue-Driven Agent Loop Skill
Triggers
Alternate expressions and non-obvious activations (primary phrases are matched automatically from the skill description):
"work through the bugs" → issue-driven loop shorthand
"address the backlog" → iterative issue queue processing
"tackle issue [N]" → single-issue loop
Purpose
Transforms the issue tracker from a passive record into an active 2-way collaboration surface. Each Al cycle posts structured status to the issue thread, scans for human feedback, and responds substantively. The human can monitor and steer agent work asynchronously by commenting on the issue.
Behavior
When triggered, this skill invokes the address-issues command with appropriate parameters extracted from the user's natural language request.
Parameter Extraction
User Says
Extracted Parameters
"address issues 17 18 19"
17 18 19
"fix open bugs"
--filter "status:open label:bug"
"work on the bug backlog"
--filter "status:open label:bug"
"tackle issue 17"
17
"go through all open issues"
--all-open
"address the open issues interactively"
--all-open --interactive
"focus on security bugs"
--all-open --guidance "Focus on security bugs"
"fix bugs 17-19, they're all auth related"
17 18 19 --guidance "These are all related to auth"
Parameter Support
This skill supports the standard AIWG --interactive and --guidance parameters:
--interactive: When detected in user intent (e.g., "interactively", "walk me through"), adds --interactive to invoke discovery questions before starting and pause between issues for human go/no-go.
--guidance: When the user provides upfront direction (e.g., "focus on bugs", "skip feature requests", "security first"), extracts the guidance text and passes it via --guidance "..." to tailor prioritization and approach without interactive prompts.
Execution Steps
Parse intent — identify issue numbers, filters, or "all open" from user message
Invoke address-issues with extracted parameters
Monitor execution — the command handles the full cycle protocol
The 3-Step Cycle Protocol
Each cycle of the agent loop follows this protocol:
Step 1: Work — Read issue context, implement fix/feature, run tests
Step 2: Post Status — Post structured comment to the issue thread:
**AL CYCLE #N – [Progress|Blocked|Review Needed]**
- Actions taken this cycle
- Task checklist (updated)
- Blockers (if any)
- Next steps
Step 3: Scan & Respond — Read all new thread comments, classify them (feedback/question/approval/correction), and incorporate into the next cycle. Never ignore human input.
Thread Scanning Classification
Comment Type
Agent Response
Feedback
Incorporate into next cycle's work
Question
Answer in next status comment
Approval
Proceed to next phase or close issue
Correction
Adjust approach, acknowledge the change
Automated/bot
Ignore
Completion Criteria (per issue)
An issue is resolved when:
Implementation is complete
Tests pass
Documentation updated (if needed)
All thread feedback addressed
No unresolved blocker comments
Multi-Issue Strategy
Strategy
Trigger
Sequential
Default — one issue at a time
Batched
When user mentions "related" issues or same module
Parallel
When user says "in parallel" (respects context budget)
User: "go through all open issues, give each one up to 8 cycles"
→ /address-issues --all-open --max-cycles 8
With guidance
User: "fix the open bugs, focus on security issues first"
→ /address-issues --filter "status:open label:bug" --guidance "Security issues are top priority"
Guidance with batch context
User: "address issues 17, 18, 19 — they're all related to the auth refactor"
→ /address-issues 17 18 19 --guidance "These are all related to the auth refactor, address them as a batch"
