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

1. Run bash scripts/run-skill.sh next-step analyze.js to collect deterministic findings
2. Parse the JSON output — findings, gates, diff summary, phase, feature_context, next_actions, backlog
3. If P0/P1 findings exist → format top 3 as actionable suggestions (Findings Mode)
4. If mid-pipeline (gates not all passed, no P0/P1) → use progression tables to suggest next step
5. If phase = post_precommit → suggest doc sync + request update (use next_actions)
6. If phase = feature_complete → output feature complete summary with backlog
7. 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 18 deterministic heuristics against git state, review state, and feature context:

The script exits with code 0 (no P0/P1), 1 (has P1), or 2 (has P0).

Output Fields (v2)

Script Failure Fallback

If the script fails or is unavailable, fall back to manual signal collection:

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 AND confidence >= 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_actions are 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:

Non-Command Suggestions

When the best next step is NOT a slash command:

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