Next Step Advisor
When NOT to Use
- Auto-loop already handling review/precommit flow (hooks track and assist)
- User gave a clear, specific instruction (just do it)
- Mid-execution of a skill workflow (follow that skill's progression)
Procedure
- Run
bash scripts/run-skill.sh next-step analyze.jsto collect deterministic findings - Parse the JSON output — findings, gates, diff summary, phase, feature_context, next_actions, backlog
- If P0/P1 findings exist → format top 3 as actionable suggestions (Findings Mode)
- If mid-pipeline (gates not all passed, no P0/P1) → use progression tables to suggest next step
- If phase = post_precommit → suggest doc sync + request update (use next_actions)
- If phase = feature_complete → output feature complete summary with backlog
- If all gates pass and no P0/P1 → output session summary with commit seed (Summary Mode, P2/P3 shown as oversights)
Script Integration
The analyze script runs 17 deterministic heuristics against git state, review state, and feature context:
| Priority | Heuristics | Meaning |
|---|---|---|
| P0 | Gate missing (code review, doc review, precommit), state drift | Required steps not completed |
| P1 | Test gap, security hotspot, migration risk, doc-sync-needed, request-stale | Important oversights |
| P2 | README missing, skill-lint needed, locale drift, mixed concerns, ac-incomplete | Quality improvements |
| P3 | Main branch warning, requirements-advisory, feature-complete | Informational |
The script exits with code 0 (no P0/P1), 1 (has P1), or 2 (has P0).
Output Fields (v2)
| Field | Type | Description |
|---|---|---|
| phase | string | mid_development / post_precommit / ready_to_commit / feature_complete / clean |
| feature_context | object | { key, source, confidence, docs_path, has_tech_spec, has_requirements, has_requests } |
| next_actions | array | [{ id, command, args, reason, confidence }] sorted by confidence desc |
| backlog | object/null | { total_features, incomplete_features } — only when phase = feature_complete |
Script Failure Fallback
If the script fails or is unavailable, fall back to manual signal collection:
| # | Signal | How |
|---|---|---|
| 1 | Git branch | git branch --show-current |
| 2 | Git status | git status -sb |
| 3 | Changed files | git diff --name-only HEAD + git status --porcelain (for untracked) |
| 4 | Review state | .claude_review_state.json |
Then use the Progression Tables below.
Dispatch Mode (--go flag)
When --go is provided, auto-execute the top next_action IF:
next_actions[0]exists ANDconfidence >= 0.8- No P0 findings
- Command is not null
Output: "Auto-dispatching: [command] [args]" then invoke the Skill tool and report the result.
Safety constraints:
- NEVER dispatch on P0. Fall back to advisory mode if confidence < 0.8.
- Only skill slash-commands from
next_actionsare dispatched (via Skill tool). No arbitrary shell execution. - Arguments are limited to file paths and flags extracted by
buildNextActions— no user-supplied strings.
Progression Tables (mid-pipeline fallback)
See references/progression-tables.md for full work-type detection and step-by-step progression tables (Feature Dev, Bug Fix, Docs, Refactoring, Investigation).
Output Format — Findings Mode
When the script detects actionable findings, format the top 3:
📍 [work type] | [branch] | [current phase]
⚠️ Findings ([count]):
1. **[P0/P1/P2] [finding id]** — [message] → [suggestion]
2. **[P0/P1/P2] [finding id]** — [message] → [suggestion]
3. **[P0/P1/P2] [finding id]** — [message] → [suggestion]
[+N more if suppressed]
➡️ Next: [primary action based on highest priority finding]
Output Format — Session Summary (all gates pass, no P0/P1)
When all gates pass and no actionable findings remain, summarize the session instead of suggesting "commit/push":
📍 [work type] | [branch] | all gates passed
✅ Session Summary:
- [what was accomplished, derived from diff summary and conversation]
- [key changes: N files changed, types of changes]
⚠️ Oversights (if any P2/P3 remain):
- [P2/P3 findings as optional improvements]
💬 Commit seed:
[type]: [concise description of all changes]
The commit seed is a suggestion — the user decides the final message.
Output Format — Feature Complete (phase = feature_complete, backlog exists)
📍 [work type] | [branch] | feature complete
✅ Feature [key] Complete
- All gates passed, no sync issues remaining
📋 Backlog ([N] incomplete):
- [feature-key] — status: [status], unchecked AC: [count]
- ...
➡️ Next: Pick a feature from the backlog or start new work
Document Completeness Check
When a feature directory exists in docs/features/, cross-check document pairs:
| State | Suggest |
|---|---|
| tech-spec exists, no request | /create-request |
| request exists, no tech-spec | /tech-spec |
| Both exist, code changed | /update-docs + /create-request --update |
Non-Command Suggestions
When the best next step is NOT a slash command:
| Signal | Suggest |
|---|---|
| Requirements unclear | "Confirm requirements with PM/user before proceeding" |
| Affects other services | "Confirm upstream/downstream API contracts" |
| Security-sensitive change | /codex-security before proceeding |
| No clear task | "Clarify what to do — don't start without clear requirements" |
| Session just started, dirty worktree | "Handle uncommitted changes first" |
Verification
- Script ran successfully (or fallback used)
- Findings formatted with priority and actionable suggestions
- Session summary includes commit seed when all clear
- Non-command suggestions used when appropriate