Skill: agentic-loop

Orchestrates the complete development workflow from a GitHub Issue to a merged PR. Invokes phase skills in sequence, routing through the appropriate profile based on task complexity.

Prerequisites

• gh CLI authenticated and repo on GitHub.
• git >= 2.5 (worktree support).
• Issue number available (passed as argument or read from context).
• Plugin installed and .claude/imat-agentic.local.md configured (run /imat-agentic:configure first if missing).

Worktree convention

Each issue always runs in its own git worktree — never in the main working tree.

• Worktree path: ../imat-worktrees/issue-<N>-<slug> (sibling of repo root)
• All subsequent steps execute with cwd = <worktree-path>
• Artifacts at .claude/agentic/issue-<N>-<slug>/ live inside the worktree

Phase sequence

Phase 0 — Recall (optional)

If recall_before_explore: true in .claude/imat-agentic.local.md, invoke @recall to search .claude/knowledge/ for related patterns, ADRs, and solutions before spending tokens on exploration.

Recall results feed directly into Phase 3 as prior knowledge.

Phase 0.5 — Flow Router

Invoke @flow-router to classify the task and select a phase profile:

Profile	Phases included
full	Recall → Explore → Clarify → Design → Gate → Implement → Review → Finalize → Capture
balanced	Explore → Design → Gate → Implement → Review → Finalize → brief Capture
quick	Recall → (skip Explore) → Design → Gate → Implement → Review → Finalize
trivial	Implement → harness → Finalize (no explore/design overhead)
review-only	Review existing changes → Finalize

The flow router outputs a profile and skipped_phases list. Use these to gate each subsequent phase.

Phase 1 — Intake (Orchestrator)

gh issue view <N> --json number,title,body,labels

Fetch base and create worktree + branch:

git fetch origin
git worktree add -b issue-<N>-<slug> ../imat-worktrees/issue-<N>-<slug> origin/main

If branch already exists, omit -b. Verify with git worktree list.

Initialize .claude/agentic/issue-<N>-<slug>/state.yaml from references/templates/state.yaml. Fill in task.id, task.title, task.branch, task.worktree, task.phase: explore, config_used.profile.

Phase 2 — Explore (Explorer)

Skip if profile is quick, trivial, or review-only.

Invoke @codebase-exploration. Produces exploration.md.

Phase 3 — Clarification (Orchestrator)

After exploration, review exploration.md for blocking open questions. If any questions would block design or implementation, surface them to the user via AskUserQuestion before proceeding.

Persist answers to .claude/agentic/issue-<N>-<slug>/clarifications.md (use references/templates/clarifications.md).

Skip if no blocking questions.

Phase 4 — Architecture Design + Gate (Architect)

Skip if profile is trivial or review-only.

Invoke @design-gate. Produces plan.md with a gate verdict.

• PASS → continue to Phase 5
• DISCUSS → pause and present specific questions to user. Wait for explicit OK.
• RETHINK → return to Phase 2 with restrictions noted in plan.md. Maximum 2 rethink cycles.

Phase 5 — Implementation (Implementer)

Skip if profile is review-only.

Invoke @implementation. Executes each step in plan.md, running the harness after each. If PostToolUse hook is active, harness feedback arrives automatically.

Phase 6 — Code Review (Reviewer)

Invoke @code-review. Confidence threshold from config (default: 80).

• approved → Phase 7
• changes_requested → return to Phase 5. Maximum 2 retry cycles.

Phase 7 — Finalize (Orchestrator)

From worktree (cwd = ../imat-worktrees/issue-<N>-<slug>):

git add -A
git commit -m "tipo(scope): descripción en español"
git push -u origin issue-<N>-<slug>
gh pr create --title "<title>" --body "Closes #<N>" --draft

Update state.yaml: phase: done, status: completed, pr.url.

Phase 8 — Knowledge Capture (Knowledge)

Skip if profile is trivial.

Invoke @knowledge-capture. Produces ADRs/patterns if new learning emerged.

Config defaults (.claude/imat-agentic.local.md)

Key	Default	Description
quality_preset	sonnet	Model tier
parallel_explorers	2	Exploration instances
confidence_threshold	80	Review filter (70/80/90)
max_review_retries	2	Review loop cap
harness_profile	full	full / light / off
recall_before_explore	true	Run recall before exploration
auto_advance	false	Skip confirmations (ship-it only)

Additional Resources

• references/phases.md — detailed per-phase instructions and edge cases
• references/templates/state.yaml — state artifact template
• references/templates/clarifications.md — clarification artifact template