Real Subagent Protocol

Why ECL Requires Real Subagents

ECL uses real subagents to counter a predictable failure mode: the parent model anchors on its own framing and then unconsciously protects it.

Mandatory independent stages are therefore not advisory decoration. They are structural checks on plan quality.

Hard Rule

The following stages require real spawned agents:

• D / critique
• G / red-blue
• H / review
• J / compile-for-code

Stage A may use support agents, but the parent model remains the owner and writer of preprocess.

General Protocol

Every subagent follows the same base rules:

• receive only stage-local context
• do not receive the parent's preferred answer
• return structured deltas instead of final notes
• never write bundle files directly
• let the parent remain the sole bundle writer

Shared Delta Shape

All independent agents return the same shape:

{
  "facts": [],
  "challenges": [],
  "conflicts": [],
  "gaps": [],
  "discard_recommendations": [],
  "follow_up_questions": [],
  "confidence": "medium",
  "evidence_refs": []
}

This makes the parent integrate independent findings without giving up ownership of the canonical notes.

Stage A Support Agents

These are optional helpers. They are useful when the raw request is vague, solution-biased, or internally contradictory.

Intent Extractor

Purpose:

• infer likely real goals beyond the literal request
• surface hidden assumptions
• suggest examples and counterexamples worth eliciting

Use when:

• the user names a solution before naming the real problem
• the request sounds under-specified but emotionally certain

Reality Gap Checker

Purpose:

• challenge claims that may be false, incomplete, or repo-inconsistent
• identify contradictions between the request and observable reality

Use when:

• the user asserts repo behavior without evidence
• the proposed solution may rely on non-existent infrastructure

Blind Spot Scout

Purpose:

• identify dimensions the user has not named yet
• force consideration of non-goals, edge cases, acceptance meaning, and failure handling

Use when:

• the request jumps straight to feature construction
• important workflow edges are likely missing

D / Critique Agent

Stage D requires one independent critique agent.

Mission

• identify pseudo-requirements
• find contradictions
• remove unverifiable or wasteful requirements
• challenge bad decompositions before they harden

Why it exists

Without D, Stage C often preserves polished nonsense.

G / Red And Blue Agents

Stage G requires two independent agents.

Red

Mission:

• attack edge cases
• attack abuse paths
• attack dependency failure
• attack invalid state transitions
• attack ambiguity in recovery behavior

Blue

Mission:

• mitigate attacks
• convert attacks into explicit rules
• name residual risk when mitigation is out of scope

Why both are required

Red without blue creates panic. Blue without red creates optimism. The pair forces explicit defense.

H / Review Agent

Stage H requires one independent review agent.

Mission

• judge whether the next coder would still need to invent meaning
• return approved, approved_with_conditions, or rejected
• name the earliest stage to reopen if rejected

What H protects

H protects against fake readiness. A bundle can look detailed and still leave critical semantics undefined.

J / Compile-for-Code Agent

Stage J requires one independent compile-for-code agent.

Mission

• absorb the converged A-H package
• decide whether the package is really code-ready
• compile execution-facing companion docs
• confirm the direct next code command

Why J is separate from H

H judges readiness. J operationalizes readiness. A package can be review-approved and still poorly compiled for execution.

Failure Handling

If a mandatory agent cannot be created:

• write the stage note anyway
• set status=blocked_by_agent_unavailable
• set agent_mode=blocked
• stop before later dependent stages claim completion

ECL does not permit fake completion when independence is missing.

Practical Codex Guidance

When implementing this protocol in Codex:

• keep prompts stage-local
• do not forward your own solution as the default answer
• integrate deltas into notes only after checking they match repo reality
• preserve evidence references so validators and later readers can audit the reasoning trail