feat: auto-loop + continuity contract (ScheduleWakeup integration)

[zelinewang]

Summary

Integrates Claude Code's built-in ScheduleWakeup tool (the /loop dynamic-mode primitive) into the /dev workflow so a single /dev invocation can run end-to-end — including waiting on CI, long builds, agent teams, or deploy health — without requiring the human to manually say "continue" between phases.

This PR bundles three logical change groups accumulated over ~5 weeks of daily use:

1. Month of local iteration (~72 net lines)

Small quality-of-life improvements that landed locally but never got pushed:

• Language/framework detection table in P1 (TypeScript/Python/Go/Rust/Java/Kotlin/C++/Flutter indicator files → language-specific ECC reviewer)
• Source Evaluation subsection in P1: mandatory credibility check on external info (web search / DeepWiki / LLM summaries) before a finding influences downstream decisions
• Various phase-adaptation polish (proactive MCP use notes, database-reviewer on SQL tasks, architect agent for DEEP P3, etc.)

2. LOOP INTEGRATION (~192 net lines) — new dedicated section

Authoritative reference for ScheduleWakeup usage inside /dev:

• Core Contract: tool schema, sentinel distinction (<<autonomous-loop-dynamic>> vs CronCreate's <<autonomous-loop>>)
• 4 Standard Patterns: CI Wait (P9 DEEP), Long Task (P7/P8), Agent Merge (P7 DEEP Agent Teams), Deploy Health (P9 CICD) — each with concrete signal source, delay, max iter, success/failure/timeout semantics, and reason-field template
• Cost Budget table: 300s prompt-cache TTL trap documented — 60–270s cache-hit regime, never 300s, 301–3600s cache-miss regime for idle waits
• Exit Conditions Matrix (mandatory): every loop MUST declare success / failure / timeout signals before first ScheduleWakeup call
• Anti-Patterns table (6 forbidden patterns): no exit conditions / 300s delay / no max_iter / cross-phase loop reuse / generic reason / retry without root-cause fix
• CHECKPOINT interaction: do NOT ScheduleWakeup during human-decision checkpoints

3. CONTINUITY CONTRACT (~64 net lines) — new top-level section

Formalizes the end-to-end execution discipline:

• /dev owns the task from P0 to P10 archive, pauses ONLY at CHECKPOINT 1, CHECKPOINT 2, or loop-escalation
• Auto-loop default on STANDARD/DEEP tiers for phases with external wait (CI, long build, agent teams, deploy health). User does NOT need --loop flag.
• --no-loop opts out (blocking mode). --loop is redundant/explicit-only.
• Includes end-to-end flow diagrams for DEEP-with-CI and STANDARD-without-CI.

Why bundle these together?

The three groups are intertwined:

• Auto-loop behavior (group 3) depends on the loop patterns defined in group 2
• Loop patterns (group 2) reference the phase adaptations refined in group 1
• Splitting into 3 PRs would force reviewers to hold partial state across reviews
• As a solo-maintained open-source plugin, bundling keeps git history cleaner

Risk assessment

This skill is prompt/markdown — no executable code path. There is no crash, memory, or security risk. The only meaningful risks are:

1. Cost drift — loop without exit condition → infinite wake-up cycles. Mitigated by mandatory Exit Conditions Matrix + max_iter cap on every pattern.
2. Cache miss — picking 300s delay → full context reread every wake. Mitigated by explicit 300s-is-WORST callout in Cost Budget table; default is 270s.
3. Semantic ambiguity — Claude picks a sub-optimal loop when user intended blocking mode. Mitigated by --no-loop escape hatch and per-wake status blocks giving user visibility.

Test plan

• Review skills/dev-orchestrator/SKILL.md diff — focus on LOOP INTEGRATION section and CONTINUITY CONTRACT consistency
• Verify all 4 loop patterns have concrete signals (not "check if done" handwaves)
• Check internal references: every mention of "Pattern 1/2/3/4" resolves; every "LOOP INTEGRATION"/"CONTINUITY CONTRACT" cross-ref resolves
• Verify --no-loop / --max-loop-iter / --loop-delay documented in OVERRIDE FLAGS table
• Dry-run mental model: imagine /dev fix X + push PR + wait CI — does the flow make sense end-to-end?
• No auto-merge — human review first

File stats

File	Before	After	Change
skills/dev-orchestrator/SKILL.md	311	639	+328 net
commands/dev.md	11	11	+4 chars (description tweak + new flags in argument-hint)

[zelinewang]

Pushed v3 refactor on feat/auto-loop-continuity (commit da57e4e).

Snapshot of pre-v3 state saved as tag v2.1-2checkpoints on commit a398a4b (current main HEAD).

Key changes in this update:

• Removed CHECKPOINT 1 (after P5) and CHECKPOINT 2 (after P9) — now fully auto-run
• Added Requirement Clarification Gate at P3 prefix (one-shot, only if needed)
• Fixed 10 codex-review findings: Pattern 1 routing, fix-cycle cap, unknown-signal exit, LOOP WAKE fields, language detection ambiguities

Not auto-merging — human review first.