reasoning-orchestrator（推理协调器）

Reasoning Orchestrator

CRITICAL: This skill does NOT perform any reasoning itself. It plans and delegates. All reasoning is done by reading and applying individual skill files, or by spawning subagents. Do not shortcut by reasoning inline.

How to Execute This Skill

Follow these steps exactly, in order. Do not skip steps. Do not collapse steps.

STEP 1 — Triage (do this yourself, no skill file needed)

Classify the situation across these dimensions and write the classification out loud before proceeding:

Temporal direction: Forward-looking / Present-state / Backward-looking

Problem maturity:

• Unframed — the problem itself may not be correctly defined
• Framed, unsolved — problem understood, solution not found
• Solved, needs validation — solution exists, needs stress-testing
• Decided, needs learning — something happened, extract signal

Primary obstacle:

• Don't understand the system → diagnostic track
• Understand it, stuck on solutions → generative track
• Have solutions, need to choose → convergence track
• Something went wrong → retrospective track
• About to commit to something → adversarial track
• Don't know what we don't know → epistemic track
• Incentives are misaligned / need to design rules → strategic track

Domain complexity (if unclear, assign unknown and include cynefin-framework in Step 2):

• Clear / Complicated / Complex / Chaotic / Unknown

Write the triage output in this format before moving to Step 2:

TRIAGE
Temporal: [forward / present / backward]
Maturity: [label]
Obstacle: [label] → [track]
Domain: [label]

STEP 2 — Build the Execution Plan (do this yourself, no skill file needed)

Using the triage output and the routing table below, produce a numbered execution plan. Label each step as SEQUENTIAL or PARALLEL. Do not execute anything yet.

Routing table — entry points by obstacle:

Write the plan in this format:

EXECUTION PLAN
Step 1 — PARALLEL: [skill-a] ∥ [skill-b] ∥ [skill-c]
Step 2 — SEQUENTIAL: [skill-d]
Step 3 — PARALLEL: [skill-e] ∥ [skill-f]
Step 4 — SEQUENTIAL: [skill-g]

STEP 3 — Execute the Plan (step by step, never all at once)

Work through the execution plan one step at a time. For each step:

For a SEQUENTIAL step:

1. Read the skill file: view: skills/[skill-name]/SKILL.md
2. Apply the skill to the current problem context, following the methodology in that file exactly
3. Write the output in the skill's output format
4. Write the routing decision: STEP [N] COMPLETE — SEQUENTIAL: [skill-name] Key finding: [1–2 sentences] Routing to Step [N+1]: [reason based on finding]
5. Proceed to the next step

For a PARALLEL step:

1. Spawn one subagent per skill using the Task tool. Each subagent receives:

- The full problem description - Any findings from prior sequential steps - This instruction: *"Read skills/[skill-name]/SKILL.md then apply that skill's full methodology to the problem. Output your findings in that skill's output format. Do not perform other reasoning."*

1. Wait for all subagents to complete
2. Synthesize their outputs: SYNTHESIS — Step [N] PARALLEL: [skill-a] ∥ [skill-b] ∥ [skill-c] Convergent findings (2+ skills agree — higher confidence): - [finding] Divergent findings (1 skill only — worth noting): - [finding] (from [skill]) Contradictions (skills disagree — resolve before proceeding): - [skill-a] says [X], [skill-b] says [Y] → resolution: [...] Key inputs for Step [N+1]: [what the next step needs]
3. Proceed to the next step

STEP 4 — Terminate

Stop when any of these conditions are met:

• decision-synthesis has completed and produced a decision with acceptable confidence
• The problem is understood well enough to act without further analysis
• If 4+ steps have run without converging: stop, flag the problem as likely unframed, restart from Step 1 with epistemic-mapping as the only entry point

Write the termination output:

CHAIN COMPLETE
Steps executed: [list with SEQUENTIAL/PARALLEL labels]
Key finding: [1–2 sentences]
Recommended action: [what to do now]

Parallelization Reference

These five clusters are the canonical parallel groups. Skills within a cluster apply independent lenses and never depend on each other's output.

Post-Step Routing Reference

After completing any step, use this table to adjust the plan if findings warrant it:

Skill Registry