Conducty Execute

Conducty Execute — Tracer-First Subagent Execution

Execute prompts from the active plan note (in the Obsidian vault — see [[conducty-obsidian]]) by dispatching fresh Claude Code subagents (via the Task tool) with context curation, time budgets, and review rigor that scales with risk. The first prompt in each group runs as a tracer to validate plan assumptions before the rest execute.

Core Concepts

Tracer-first: The prompt marked ★ TRACER in each group runs alone. If it succeeds, the plan's assumptions hold and the remaining prompts execute. If it fails, stop — the plan needs revision, not just a fix prompt. Tracers catch bad assumptions before they multiply across an entire group.

Time-budgeted: Each prompt has a time budget from the plan. If a subagent exceeds it, that's a signal — either the prompt is underscoped, the complexity was misjudged, or the agent is stuck. Treat it as a circuit breaker, not a hard kill.

Calibrated review: Not every prompt needs the same oversight. The plan assigns a review level based on complexity:

Context curation: The orchestrator (you) constructs exactly what each subagent needs. Not your session history, not the entire project — precisely the prompt text, relevant context files, and scene-setting for where this work fits. This is the surgical team model: the surgeon thinks, the team executes with curated information.

The Process

Phase 1: Tracer Execution

1. Read the plan note from the vault, extract the tracer prompt for the target group
2. Dispatch tracer as a single implementer subagent (Task tool)
3. Handle the result:
   - DONE → Run verification via [[conducty-verify]] → Pass? Proceed to Phase 2
   - DONE → Verification fails → STOP. This is a plan problem, not a prompt problem.
     Analyze: is the verification wrong, or is the plan's approach wrong?
   - BLOCKED/NEEDS_CONTEXT → STOP. The plan's assumptions are wrong.
     Don't execute the remaining prompts. Revise the plan.
4. Report tracer result to user before proceeding

The tracer is your early warning system. Respect its signal.

Phase 2: Group Execution

For each remaining prompt in the group:
  1. Dispatch implementer subagent (see [[implementer-prompt]])
  2. Handle implementer status (see below)
  3. Apply review level from the plan:
     - verify-only → Run verification. Pass? Mark complete.
     - spec-review → Run verification + dispatch spec reviewer ([[spec-reviewer-prompt]])
     - full-review → Run verification + spec reviewer + quality reviewer ([[quality-reviewer-prompt]])
  4. If review fails → implementer fixes → re-review (same level)
  5. Mark prompt complete with evidence
After all prompts → run [[conducty-checkpoint]] for the group
Gate: user confirms before next group

Resolve before dispatch (skill-clusters wiring)

Integration patch (skill-clusters — the rest of this skill is upstream conducty, MIT). Stock conducty dispatches every prompt to a bare general-purpose subagent. Here each prompt is first resolved to a skill-cluster, so the subagent runs loaded with the right capability hub — the one organ conducty lacks. See conductor-core for the full contract.

Once per group, resolve the plan's tasks to clusters:

node ~/.agents/skill-clusters/scripts/resolve-task.mjs <tasks.md> <plan.md> --json

Returns, per task: {cluster, dispatch: <cluster>-orchestrator, tier, activate?, spokes[], confidence}. Apply per prompt:

1. Classifier proposes, resolver validates. You (the conductor) are the classifier: from each prompt's intent, propose its cluster — and optionally a specific skill. Feed your proposals back to the resolver, which validates them against skill-index.json:

   # proposals.json: { "T002": "rust", "T010": {"cluster":"supabase","skill":"supabase-postgres-best-practices"} }
   node ~/.agents/skill-clusters/scripts/resolve-task.mjs <tasks.md> <plan.md> --propose proposals.json --json
   # or inline:  --propose "T002=rust,T010=supabase"
   

   The resolver returns source: "classifier" when it accepts your proposal, ◇!/overrode when your pick beat the keyword guess, and — crucially — rejects phantoms: a proposed cluster or skill not in the index is dropped (phantoms/badSkills) and the keyword result is kept. This is PAI's Capability-Name audit applied to dispatch. confidence < 0.34, cluster: null, or a rejected phantom with no keyword fallback → escalate to the human, do not guess-dispatch.

2. Activate deferred clusters the wave touches, once: node ~/.agents/skill-clusters/scripts/tier.mjs --activate <cluster> --apply.

3. Load the resolved hub into the subagent — add ~/.agents/skill-clusters/skills/<cluster>-orchestrator/SKILL.md to the curated context (checklist item 1.5 below). The hub routes to its spoke on demand; you never enumerate spokes — the subagent pulls the one it needs from the pointer.

4. Gate. The PAI SkillClusterResolver hook denies any non-enumerated skill with resolution guidance, so a mis-resolved dispatch fails closed rather than running the wrong capability.

subagent_type stays general-purpose (or Explore for read-only review) — the capability comes from the loaded <cluster>-orchestrator, not the subagent type.

Dispatching Subagents

Use Claude Code's Task tool. The default subagent_type is general-purpose (always available) — but in this repo the subagent runs loaded with the resolved <cluster>-orchestrator (see Resolve before dispatch above): the subagent type stays general-purpose, the capability comes from the loaded hub. If the harness exposes specialized subagents (e.g. Explore for read-only codebase analysis, Plan for architecture work), use them where their description fits — they don't replace the implementer model selection below.

Choose model by the prompt's complexity:

For read-only review subagents (spec reviewer, quality reviewer, code-review lens subagents), use the Explore subagent type if available; otherwise general-purpose. Run them at haiku or sonnet — review is pattern matching, not generation.

Context curation checklist — provide exactly:

1. The full prompt text (pasted, not a file reference — subagents don't read plan files) 1.5. The resolved <cluster>-orchestrator SKILL.md — the capability hub for this task (see Resolve before dispatch); the subagent runs loaded with it and pulls spokes on demand
2. Scene-setting: which project, where this fits in the broader plan, what came before
3. Relevant architecture from the project context file
4. Dependencies on prior prompt outputs (if any)
5. The no-go zones from the prompt
6. The time budget

Do not provide: Your session history, unrelated plan context, the entire project tree.

Handling Implementer Status

DONE: Apply the prompt's review level.

DONE_WITH_CONCERNS: Read the concerns. Correctness/scope concerns → address before review. Observational concerns (file getting large, questionable existing patterns) → note for future planning, proceed to review.

NEEDS_CONTEXT: The prompt was underspecified. Provide the missing context and re-dispatch. Check the project context file first. If the context doesn't exist, this is a prompt quality issue — note it for [[conducty-improve]].

BLOCKED: Assess the blocker:

1. Context problem → provide more context, re-dispatch
2. Task exceeds agent capability → re-dispatch with a more capable model (opus)
3. Task too large → split into sub-prompts, dispatch sequentially
4. Plan assumption is wrong → stop and revise the plan for this prompt

Never force-retry without changing something. If the agent said it's stuck, something about the prompt, the context, or the model needs to change.

Time Budget Enforcement

When a subagent has been running significantly longer than the prompt's time budget:

1. Check if it's making progress (producing output, committing) or stuck (silent, looping)
2. If stuck: interrupt and treat as BLOCKED
3. If progressing but slow: let it finish, but flag the time overrun in checkpoint notes
4. Either way: this means the complexity estimate was wrong. Note it for [[conducty-improve]] so future plans calibrate better.

Orchestrator Context Management

Long plans burn the orchestrator's context window. Each subagent dispatch returns a result back into your session, plan-note reads accumulate, and verification output stacks up. Without management you run out of context mid-plan and lose conceptual integrity.

The orchestrator is a stateful coordinator, but its session is cache — durable state lives in the vault (plan notes, checkpoint notes, [[Prompt Log]]). Treat /compact as a normal phase boundary, not a panic move.

When to /compact:

• Between groups, never mid-group. Group boundary is the safe point — all subagent dispatches for the group have returned, [[conducty-checkpoint]] has run, and the plan note records the group's state.
• Heuristic triggers:
  • 3+ groups remaining and the orchestrator session is already long (heavy tool output accumulated)
  • Claude Code warns context is approaching the limit
  • The just-finished group included large file reads, verbose code review output, or long verification logs
  • The just-finished group ran at full-review (three subagents per prompt — implementer + spec reviewer + quality reviewer — accumulates fast)
  • About to enter a group whose prompts will themselves return large outputs (codebase-wide refactors, multi-file generations)

Compact protocol:

1. Before compact — confirm durable state is written:
   • Plan note marks the just-finished group's prompts complete with evidence
   • Checkpoint note exists for the group
   • [[Prompt Log]] has entries for each prompt in the group
   • In-flight observations for [[conducty-improve]] are captured in the plan note's improvement-feedback section, not just in your head
2. Compact — /compact summarizes the session
3. After compact — restore cycle position before dispatching the next group:
   • Re-read the active plan note from the vault
   • Re-read the most recent checkpoint note
   • Re-read the last few entries in [[Prompt Log]]
   • Identify: which group is next, which prompt is its tracer, what no-go zones apply, what improvement experiments this plan is testing

Never compact mid-group. A compact between the tracer and the rest of a group risks losing the tracer's signal — the evidence that justified proceeding. Finish the group, checkpoint, then compact.

Subagents are unaffected. Each subagent dispatch is a fresh Task with a fresh window. Your /compact has no impact on already-dispatched or future subagents — they receive curated context per dispatch regardless.

Subagent Prompt Templates

• [[implementer-prompt]] — Dispatch an implementer subagent
• [[spec-reviewer-prompt]] — Dispatch spec compliance reviewer (for spec-review and full-review)
• [[quality-reviewer-prompt]] — Dispatch code quality reviewer (for full-review only)

Principles

• Tracer result is a plan signal — a tracer failure means "revise the plan," not "fix the prompt"
• Context is curated, not dumped — give subagents exactly what they need, nothing more
• Review rigor matches risk — ceremony for high-risk work, speed for low-risk work
• Time budgets are circuit breakers — they surface misestimates, not punish slow agents
• Every completion claim needs evidence — use [[conducty-verify]] before marking anything done
• Failures are learning data — every BLOCKED, every NEEDS_CONTEXT, every time overrun feeds back to [[conducty-improve]]
• The orchestrator doesn't implement — you coordinate, curate context, and maintain conceptual integrity. The subagents implement.