pathfinding（pathfinding 搜索）

Pathfinding

Adaptive Q&A → unclear requirements → clear path.

<when_to_use>

• Ambiguous/incomplete requirements
• Complex features needing exploration
• Greenfield projects with open questions
• Collaborative brainstorming or problem solving

NOT for: time-critical bugs, well-defined tasks, obvious questions

</when_to_use>

Start honest. Clear request → level 4–5. Vague → level 0–2.

At level 4: "Can proceed, but 1–2 more questions would reach full confidence. Continue or deliver now?"

Below level 5: include △ Caveats section.

Load the maintain-tasks skill for stage tracking. Stages advance only, never regress.

Task format — each stage gets context-specific title:

- Prep { domain } requirements
- Explore { approach } options
- Clarify { key unknowns, 3-4 words }
- Deliver { artifact type }

Situational (insert before Deliver when triggered):

• Resolve Conflicts → ◆ Caution or ◆◆ Hazard pushback
• Validate Assumptions → high-risk assumptions before delivery

Workflow:

• Start: Create stage matching initial confidence in_progress
• Transition: Mark current completed, add next in_progress
• High start (4+): Skip directly to Clarify or Deliver
• Early delivery: Skip to Deliver + △ Caveats

Calibrate first — user may have already provided context (docs, prior conversation, pointed you at files). If enough context exists, skip to level 3–4. Don't re-ask what's already clear.

If gaps remain, explore focus areas (pick what's relevant):

• Purpose: What problem? Why now?
• Constraints: Time, tech, team, dependencies
• Success: How will we know it works?
• Scope: What's in, what's out?

When multiple approaches exist:

• Propose 2–3 options with trade-offs
• Lead with recommendation ★ and reasoning
• Let user pick, combine, or redirect

Principles:

• YAGNI — cut what's not needed
• DRY — don't duplicate effort or logic
• Simplest thing — prefer boring solutions

Use EnterPlanMode for each question — enables keyboard navigation of options.

Structure:

• Prose above tool: context, reasoning, ★ recommendation if clear lean
• Inside tool: options only (concise, scannable)

At level 0 — start with session intent:

• Quick pulse check vs deep dive?
• Exploring possibilities or solving a specific problem?
• What does "done" look like?

Levels 1–4 — focus on substance:

• 2–4 options per question + "5. Something else"
• Inline [★] on recommended option + *italicized rationale*
• User replies: number, modifications, or combos

Loop: Answer → Restate → Update Confidence → Next action

After each answer emit:

• Confidence: {BAR} {NAME}
• Assumptions: {if material}
• Unknowns: {what we can clarify; note unknowables when relevant}
• Decisions: {what's locked in}
• Concerns: {what feels off + why}

Next action by level:

• 0–2: Ask clarifying questions
• 3: Summarize (3 bullets max), fork toward 5
• 4: Offer choice: refine or proceed
• 5: Deliver

When answer reveals a concern mid-stream:

• Pause before next question
• Surface with △ + brief description
• Ask: clarify now, note for later, or proceed with assumption?

Example: "△ This assumes the API supports batch operations — clarify now, note for later, or proceed?"

If user proceeds despite significant gap → escalate to pushback protocol.

Escalate when choice conflicts with goals/constraints/best practices:

• ◇ Alternative: Minor misalignment. Present option + reasoning.
• ◆ Caution: Clear conflict. Recommend alternative, explain risks, ask to proceed. Triggers Resolve Conflicts.
• ◆◆ Hazard: High failure risk. Require mitigation or explicit override. Triggers Resolve Conflicts.

Override: Accept "Proceed anyway: {REASON}" → log in next reflection → mark Resolve Conflicts complete.

Integrate skeptic agent for complexity sanity checks:

Recommend (offer choice):

• Level 5 reached with △ Caveats > 2
• Red flag language in decisions: "might need later", "more flexible", "best practice"

Before finalizing — you have {N} caveats. Want to run skeptic for a sanity check?
[AskUserQuestion]
1. Yes, quick check [★] — I'll challenge complexity interactively
2. Yes, deep analysis — launch skeptic agent in background
3. No, proceed — deliver as-is

Auto-invoke (no choice):

• Level 4+ with 3+ unknowns persisting across 2+ question cycles
• ◆◆ Hazard escalation triggered during session

When auto-invoking:

[Auto-invoking skeptic — {REASON}]

Launch with Task tool:

• subagent_type: "outfitter:skeptic"
• prompt: Include current decisions, unknowns, and caveats
• run_in_background: false (wait for findings before delivery)

After skeptic returns:

• Present findings to user
• If verdict is block → add Resolve Conflicts stage
• If verdict is caution → offer choice to address or acknowledge
• If verdict is proceed → continue to delivery

Level 5: Produce artifact immediately (doc, plan, code, outline). If none specified, suggest one.

After delivering, ask where to persist (if applicable):

[EnterPlanMode]
1. { discovered path } [★] — { source: `CLAUDE.md` preference | existing directory | convention }
2. Create issue — { Linear/GitHub/Beads based on project context }
3. ADR — { if architectural decision }
4. Don't persist — keep in conversation only
5. Something else — different location or format

Discovery order for option 1:

1. CLAUDE.md or project instructions with explicit plan storage preference
2. Existing .agents/plans/ directory
3. Existing docs/plans/ directory
4. Fall back to .agents/plans/ if nothing found

Always suggest filename based on topic. Match existing conventions if present.

Mark Deliver completed after artifact is delivered (persistence is optional follow-up).

Below 5: Append △ Caveats:

• Open questions + context
• Assumed decisions + defaults
• Known concerns + impact
• Deferred items + revisit timing

ALWAYS:

• Task stage matching initial confidence at start
• EnterPlanMode for each question (keyboard nav)
• Prose above tool for context + ★ recommendation
• One question at a time, wait for response
• Restate + update confidence before next move
• Update todos at level 4, level 5 thresholds
• Apply pushback protocol on conflicts
• Check skeptic triggers at level 4+ (unknowns, caveats, red flags)

NEVER:

• Proceed from 0–3 without clarifying questions
• Hide uncertainty below level 5
• Stack questions or bury decisions in paragraphs
• Put recommendation inside plan tool (keep in prose)
• Skip reflection after answer
• Regress stages
• Ignore skeptic auto-invoke triggers

• confidence.md — confidence deep dive
• questions.md — question crafting
• examples/ — session examples
• skeptic agent (outfitter:skeptic) — complexity sanity checks