continue

name

continue

description

Resume fully autonomous work on the `main` branch and drive the per-task TDD loop. Trigger when the user says 続けて, "resume", "pick up where we left off", "/continue", "次", "go", or starts a fresh session expecting prior context. `/continue` may also be re-invoked automatically (cron / wake hook); treat every invocation as the same go signal. The full step-by-step loop spec (Step 0 → 7) lives in CLAUDE.md § Autonomous Workflow + § The only stop (single condition: user explicit stop) and is loaded into every turn's system prompt; this skill is the invocation trigger and carries the resume procedure, the Phase-boundary review chain, the subagent delegation cheatsheet, and the model-selection guidance.

continue

The /continue slash command's job is to start running the autonomous TDD loop. The loop's full step-by-step spec lives in CLAUDE.md § Autonomous Workflow (Step 0 → 7) + § The only stop (single condition: user explicit stop), loaded into every turn's system prompt.

The loop never stops itself. Region / cluster / task / commit / Phase boundaries all roll into the next unit of work. Bad-Smell triggers are interrupts (see CLAUDE.md § Smell triggers are interrupts, not stops) — the surgery lands at the right depth, then the loop continues. Test failures are diagnosed and fixed in-flight, not handed off. /continue may also be re-invoked automatically (cron / wake hook); each invocation is the same go signal.

This file carries the procedural context that is only needed at invocation time: the resume procedure, the Phase-boundary review chain, subagent delegation, and model selection. The per-task loop spec is not duplicated here — CLAUDE.md is the single source.

Resume procedure (on every session pickup)

1. Read .dev/handover.md (SessionStart hook already prints it; re-read). Pay attention to any "Guardrail refresh" section — it points at recent principle.md / CLAUDE.md spirit edits the loop must honour.
2. Read CLAUDE.md § Project spirit (top section, governs all other rules) and .dev/principle.md (Bad Smell catalogue + Structural imagination phase). These two are the meta layer the per-task loop is checked against. 2a. Read .dev/project_facts.md (user-declared invariants — F-001 … F-007 at 2026-05-24) and .dev/structure_plan.md (anticipated directory tree Phase 5-20). When a task touches any topic these files cover, treat them as fact above ROADMAP / ADR text.
3. Read .dev/ROADMAP.md §9.0 (the gap-area model — the phase-queue/§9.-placeholder model is RETIRED per ADR-0142): the three gap areas' BUILT status + named gaps + draining D-NNN rows. Next-unit selection is handover.md's "First task on resume" if concrete, else the .dev/debt.yaml active: list EASIEST-FIRST (CLAUDE.md § When the active work unit completes).
4. (Dormant cadence per ADR-0025: skip the docs/ja/learn_clojurewasm/ chapter check. private/notes/ per-task notes continue.)
5. git log --oneline -10 — read the recent commit chain so the next task starts with the as-pushed reality, not the handover's narration of it.
6. Confirm HEAD is green without re-running the full gate — the pushed HEAD was already gated when it landed. Check the .dev/.gate_pass / .dev/.smoke_pass fingerprint (or run a quick bash test/run_all.sh --smoke); only run the full gate at resume if the fingerprint is stale or the tree is dirty (ADR-0107 — full e2e is heavy, so resume does not pay for it by default). If a check's output is large (>200 lines), run via subagent and ask only for pass/fail + the first failure.
7. Summarise to the user in 5-10 lines:
   • Phase (number + name)
   • Last commit
   • Test status
   • Active task (number + name + exit criterion)
   • Any Entry debts owed at this Phase (D-NNN list) so the loop's Step 1 picks them up before touching structure
8. Immediately proceed into the TDD loop per CLAUDE.md § Autonomous Workflow. Do not wait for "go" — /continue itself is the go signal.

Boundary review chain (auto-runs when a gap-area unit arc closes)

The phase-queue is RETIRED (ADR-0142); the boundary is now a major work-arc close (a gap-area unit family drained, a release cut, or ~a full gate-ceiling cycle of related units). Run this chain in parallel where possible, continue into the next self-selected unit without asking:

1. Run audit_scaffolding. Findings of any severity feed the loop's next interrupt (block-severity ones become an immediate surgery before the next Phase opens); the loop never halts on them.
2. In parallel (multi-agent fan-out, single message with multiple Agent tool calls):
   • Subagent A: built-in simplify on git diff <phase-start>..HEAD -- src/ — apply behaviour-preserving suggestions, queue larger ones.
   • Subagent B: built-in security-review on unpushed commits.
   • Subagent C: write any outstanding chapter(s) for the closed phase, pulling from private/notes/ task-notes.
3. Synthesise in main: 1 line per check + severity counts. No "shall I proceed?" question — proceed.
4. (Bench sweep retired 2026-06-11 — the bench/quick.sh / bench/quick_baseline.txt auto-baseline was removed from the gate, so no dangling samples accumulate. Perf is measured on demand via bench/compare_langs.sh / bench/run_bench.sh.)
5. Update handover.md to point at the next self-selected unit (debt.yaml easiest-first).
6. Proceed to that unit's Step 0. Auto-compaction handles context size transparently — no agent action needed.

Subagent delegation cheatsheet

Trigger	Action
Survey ≥ 1 reference codebase / OSS	Step 0 — general-purpose subagent (writes the survey note itself; Explore is read-only)
Test output > 200 lines	Step 5 — Bash subagent (run_in_background if long)
Search across > 5 files	Explore subagent (read-only — search only)
Phase boundary audit / simplify / review	Multi-agent fan-out (parallel)
Outstanding chapters at phase close	general-purpose subagent
Single-file edit, < 200 lines context	Stay in main

Default rule: subagent fork on context isolation, not on importance.

What NOT to invoke during the loop

• simplify per source commit — overkill; queue for Phase boundary.
• review (PR-style) per commit — overkill; reserve for pre-push or pre-tag.
• audit_scaffolding per task — runs at Phase boundary only (or every ~10 chapters).

Model selection (dual-model)

• Per-task TDD loop (Step 0 → 7): current session's model — Opus 4.7 is fine.
• Phase boundary chain (multi-agent fan-out): prefer Opus 4.6 for the long-context audit / simplify / chapter-write subagents — Opus 4.7's MRCR v2 retrieval is known to degrade above ~100k tokens versus 4.6. Sonnet 4.6 is a viable cost-efficient alternative.

When unsure, default to subagent inheriting the parent model; flip to Opus 4.6 only if a long-context task underperforms.