collabN-local

collabN-local Skill - N-Round Local-Agent Consultation

Purpose

Run the collaborate3 consultation pattern entirely in-house: instead of consulting an external model (Gemini 2.5 Pro via Zen/OpenRouter), consult one of our own sub-agents (brainstorm, senior-engineer, oracle, Explore, etc.).

The decisive advantage over the external collaborate* commands: a local agent has the files. It can read the actual code, trace design docs and postmortems, and run real tests — none of which the external expert could ever do. You are effectively holding a structured, multi-round technical review with a second instance of yourself that can see the repo.

N is the number of consultation rounds (default 3), always followed by one independent final assessment (Round N+1). This command works exactly like collaborate3, with two swaps: the consultant is a local agent instead of Gemini, and the continuity handle is the agent's agentId + SendMessage instead of a Zen continuation_id. Everything else — the per-round files, the verbatim saving, the critical analysis, the independent final assessment — is identical.

Invocation & Arguments

/collabN-local [N] [agent] <topic / problem statement>

• N (optional integer) — number of consultation rounds. Default 3. If omitted, assume 3 (do not block to ask).
• agent (optional) — which sub-agent to consult. Default brainstorm. Recognized values map to the Agent tool's subagent_type:
  • brainstorm — default. Thought-partner that blends the senior-engineer's hands-on judgment with the oracle's doc/knowledge tracing, runs tests against the real repo, and delegates to either specialist when a sub-question is squarely theirs. Best general-purpose consultant for this command.
  • senior-engineer — pure expert-engineering role; architecture, debugging, hands-on fixes.
  • oracle — for "how/why does this work" questions; traces design docs, postmortems, MOCs, issues, git history with sourced citations.
  • Explore — read-only "where is X / what calls Y" codebase mapping; lighter weight, cannot run tests or edit.
  • any valid agent type (Plan, investigate, general-purpose, …) is also acceptable.
• <topic> — the problem/decision. If omitted, use the current conversation as the topic.

Parsing rule: if the first token is an integer, it's N; if the next token matches a known agent type, it's agent; the remainder is the topic.

Examples:

• /collabN-local 3 "export/import column alignment is off" → 3 rounds with brainstorm.
• /collabN-local oracle "why does the clone formula use MIN not subtraction" → 3 rounds (default) with oracle.
• /collabN-local 2 Explore "where is widget serialization handled" → 2 rounds with Explore.
• /collabN-local 1 "is asyncio.gather or create_task better here" → 1 round (quick) with brainstorm.

MANDATORY OUTPUT CONTRACT — read this before you start

This command is defined by the files it leaves on disk, exactly like collaborate3. A run that produces only a final synthesis has FAILED, no matter how good the analysis was. As you go, you MUST write:

• one Claude message file per round (what you sent the agent), and
• one agent response file per round (the agent's reply, saved VERBATIM), and
• one final assessment file.

For the default 3-round run that is 7 files (3 Claude + 3 agent + 1 final). For N rounds it is 2N + 1 files.

⚠️ THE TRAP THAT BREAKS THIS COMMAND. Unlike Gemini — whose replies arrive via mcp__zen__chat and exist nowhere else, forcing you to save them — a local sub-agent's reply arrives as an Agent/SendMessage tool result that is already sitting in your context. That makes it feel saved. It is NOT saved. You must write the file anyway, every round, before moving on.

Hard gate: treat each round's two files as a barrier — do not begin round k+1 until round k's Claude-message file AND agent-response file both exist on disk. Do not batch all the file-writing to the end. Write the agent's response immediately after each spawn/SendMessage returns, verbatim, the same way you would for Gemini.

Consultant mechanism (the local analog of continuation_id)

To hold a real multi-round conversation with one agent, spawn it once and keep talking to the same instance so its context persists across rounds:

1. Round 1 — spawn: call the Agent tool with subagent_type: <agent> (consider run_in_background: true so you get a stable agentId to resume). Capture the returned agentId (format a...-...) — this is your continuation handle; record it at the top of every response file.
2. Rounds 2..N — continue: use SendMessage(to: <agentId>, message: <next round>) to resume the same agent with its context intact. A fresh Agent call would start from zero and defeat the purpose.

Fallback (if resuming isn't available in this harness build): spawn a fresh agent each round but embed the verbatim transcript of all prior rounds (read it back from the round files you saved) so it reconstructs context. This is exactly what external collaborate3 does with Gemini, so it is a safe degradation — just more tokens. Note in the response file that you fell back.

Process Overview

Three rounds of exchange followed by final assessment (this is the default N=3 shape; see Adapting to other N below):

• Round 1: Present complete context, get initial analysis
• Round 2: Critical evaluation, probe weaknesses, run real tests, test viability
• Round 3: Final questions and proposed action plan
• Round 4: Independent final assessment using dev-workflow-process

Detailed Steps

Preparation

Gather all relevant context: problem description, the files/dirs involved (point the agent at real paths — it can open them), previous attempts and failures, current understanding, constraints, and the specific questions needing answers.

Round 1: Context & Initial Analysis

1. Prepare message (YYYY-MM-DD__hh-mm-ss__DISCUSS_Rnd1_Claude_(topic).md) — write this file:
   • Complete overview of the problem and relevant code structure (real paths).
   • Work attempted so far and why it failed.
   • Specific questions; request simple/quick vs. robust vs. long-term options plus considerations and future implications.
   • If the agent is NOT brainstorm, paste in the anti-echo-chamber standing instructions (below). brainstorm already has them baked in.
2. Spawn the agent via the Agent tool (subagent_type: <agent>); capture its agentId.
3. Save response (YYYY-MM-DD__hh-mm-ss__DISCUSS_Rnd1_(Agent)_(topic).md) — write this file now, before doing anything else. The reply is already in your context; save it to disk anyway.
   • Save the agent's response VERBATIM — no editing or summarizing.
   • CRITICAL: record the agentId and agent type at the very top (the local continuation_id). Format:

     # Consultant Response - [Topic]
     
     **AGENT_ID**: [a...-...]
     **AGENT_TYPE**: [brainstorm | senior-engineer | oracle | ...]
     **ROUND**: 1 of N
     
     ---
     
     [Full verbatim response]
     

4. Critical analysis (internal): categorize each suggestion Good / Neutral / Problematic; note effects on other code/data; identify weak points to probe.

Round 2: Critical Evaluation & Testing

1. Prepare message (YYYY-MM-DD__hh-mm-ss__DISCUSS_Rnd2_Claude_(topic).md) — write this file:
   • DO NOT blindly accept suggestions. Present your own analysis of the proposals.
   • Run real local tests for anything the agent asked you to validate, and include the actual output/data — use the scientific method to falsify ideas. (You can also ask the agent to run its own diagnostics, since it has tools; cross-check the two.)
   • Ask probing questions about weak points, challenge assumptions, explore edge cases.
2. Continue the same agent via SendMessage(to: <agentId>, message: <this message>) (fallback: re-spawn with full prior transcript embedded).
3. Save response (YYYY-MM-DD__hh-mm-ss__DISCUSS_Rnd2_(Agent)_(topic).md) — write this file now, VERBATIM, with AGENT_ID and ROUND: 2 of N at the top.
4. Deep analysis (internal): re-evaluate proposals against the new data; form preliminary conclusions.

Round 3: Final Questions & Action Plan

1. Prepare message (YYYY-MM-DD__hh-mm-ss__DISCUSS_Rnd3_Claude_(topic).md) — write this file:
   • DO NOT blindly accept Round 2 feedback. Present YOUR assessment and reasoning.
   • Ask remaining clarifying questions.
   • Propose YOUR action plan (not the agent's) and request validation of specific concerns.
2. Continue the same agent via SendMessage(to: <agentId>, message: <this message>).
3. Save response (YYYY-MM-DD__hh-mm-ss__DISCUSS_Rnd3_(Agent)_(topic).md) — write this file now, VERBATIM, with AGENT_ID and ROUND: 3 of N at the top.

Round 4: Final Assessment (Independent)

Create YYYY-MM-DD__hh-mm-ss__DISCUSS_Rnd4_FINAL_ASSESSMENT_(topic).md using the full DEV WORKFLOW PROCESS:

• Problem Analysis (SPCR: Story)
• Considerations Analysis (SPCR: Puzzle)
• Solutions Evaluation (SPCR: Content)
• Synthesis & Recommendation (SPCR: Result + PUVM)
• YOUR decision, informed by but not dictated by the consultant. Include risks and mitigations.

Adapting to other N

The four blocks above are the standard N=3 walk. For other values, keep Round 1 (context) and make the last consultation round the "final questions & action plan" round; the Final Assessment is always Round N+1:

• N=1: Round 1 → Final Assessment is Round 2. (3 files total.)
• N=2: Round 1, then Round 2 = "final questions & action plan" → Final Assessment is Round 3. (5 files.)
• N=3: as written above. (7 files.)
• N≥4: Round 1, then repeat the Round 2 (critical evaluation & testing) pattern for rounds 2..N-1, then Round N = "final questions & action plan" → Final Assessment is Round N+1. (2N+1 files.)

In every case the deliverable is 2N+1 files: N Claude-message files, N verbatim agent-response files, and 1 final assessment.

Closeout checklist (do this before declaring done)

List the files you created and confirm the count is 2N+1. If any round's two files are missing, you have not finished — go back and write them. Example for N=3:

[x] DISCUSS_Rnd1_Claude_(topic).md
[x] DISCUSS_Rnd1_(Agent)_(topic).md
[x] DISCUSS_Rnd2_Claude_(topic).md
[x] DISCUSS_Rnd2_(Agent)_(topic).md
[x] DISCUSS_Rnd3_Claude_(topic).md
[x] DISCUSS_Rnd3_(Agent)_(topic).md
[x] DISCUSS_Rnd4_FINAL_ASSESSMENT_(topic).md   → 7 files = 2(3)+1 ✓

Anti-echo-chamber: the consultant's standing instructions

Because the sub-agent is also Claude, the value comes from structure and framing, not from a different model. The default brainstorm agent already bakes the rules below into its persona — when you use it, you mainly pass context + the round number. When you point the command at a different agent (senior-engineer, Explore, etc.), paste these instructions into the round prompt. Every round's prompt to the agent must instruct it to:

• Act as an independent, skeptical reviewer — its job is to find what's wrong, not to agree.
• Read the actual files before opining; cite file:line, don't speculate.
• Disagree explicitly with the main thread's framing when warranted, and say why.
• Offer simple/quick vs. robust vs. long-term options and name the trade-offs.
• When it proposes something testable, run the test itself (if it has tools) and report real output — or specify the exact test for Claude to run.
• Flag assumptions and unknowns rather than papering over them.

Key Principles

1. Document everything as you go — full verbatim records, written to disk each round. The in-context tool result is not a substitute for the file. This is the principle this command most often violates; honor it.
2. Critical thinking over compliance — the agent is a sparring partner, not an oracle of truth; challenge everything.
3. Test before trust — validate suggestions against the real repo and real test output. This is the whole point of going local.
4. Persist the same agent — reuse the agentId so context carries across rounds; never silently start a fresh agent mid-consultation.
5. Independent decision — you make the final call in Round N+1.
6. Beware the echo chamber — a same-model consultant tends to agree; force genuine adversarial framing.

File Naming Convention

All files in ./private/claude/ (project work) or ~/claude/ (cross-project), matching the other collaborate commands. Use a short PascalCase token for (Agent) — e.g. Brainstorm, SeniorEng, Oracle, Explore:

• Round k (Claude): YYYY-MM-DD__hh-mm-ss__DISCUSS_Rnd{k}_Claude_(topic).md
• Round k (Agent): YYYY-MM-DD__hh-mm-ss__DISCUSS_Rnd{k}_(Agent)_(topic).md
• Final: YYYY-MM-DD__hh-mm-ss__DISCUSS_Rnd{N+1}_FINAL_ASSESSMENT_(topic).md

(Always run date to get the real timestamp for filenames.)

Example Usage

User: "We need to solve the export/import column alignment issue. Use collabN-local with 3 rounds."

Claude:
1. Writes Rnd1_Claude (context) → spawns brainstorm → writes Rnd1_(Agent) VERBATIM (agentId saved)
2. Runs the failing case, writes Rnd2_Claude (analysis + test output) → SendMessage(agentId) → writes Rnd2_(Agent) VERBATIM
3. Writes Rnd3_Claude (own action plan) → SendMessage(agentId) → writes Rnd3_(Agent) VERBATIM
4. Writes Rnd4_FINAL_ASSESSMENT (dev-workflow-process, YOUR decision)
5. Closeout: confirms 7 files exist

When to Use

• You want the collaborate3 rigor but the consultant needs to actually see/run the code.
• Offline / no external API, or you want to avoid OpenRouter cost and dependency.
• The question is repo-specific (architecture, debugging, "how does X connect to Y", "why was this designed this way").

When NOT to Use

• You specifically want an outside, different-model perspective to avoid groupthink — use collaborate3 (external Gemini) instead; that's the one thing a local agent can't give you.
• Trivial problems with obvious solutions, or pure-syntax lookups.

Relationship to the collaborate family

Remember: the local agent provides input and can prove things against the real repo — but YOU make the decision, and you must actively guard against it simply agreeing with you. And the run is only complete when all 2N+1 files exist on disk.