---
name: long-haul-parallel-repair
description: >
Use for multi-slice SDD when ≥2 independent streams exist, or when the user
authorizes subagents, fan-out, delegation, long-running autonomous work,
"work for hours", "iterate until satisfied", or solving every <8/10 confidence
item through evidence, fixes, and verification. Deep investigation alone is
not enough unless streams or autonomy language are present.
license: MIT
tags: [parallel, subagents, orchestration, investigation, repair, verification, confidence, codex]
agents: [claude-code, codex, autojack, grok]
category: orchestration
metadata:
version: "1.1.0"
capabilities:
network: false
filesystem: readwrite
tools: [Agent, Read, Edit]
requires-secrets: []
resources:
- path: agents/openai.yaml
type: file
---Long-Haul Parallel Repair
Activation Contract
Authorized when any of these hold:
- ≥2 independent streams that materially advance the goal (default multi-slice / SDD posture — no special phrase required).
- The user authorizes subagents, delegation, fan-out, or long-running autonomous work (phrases: "fan out", "dispatch agents", "subagents", "work for hours", "full autonomy", "no human input", long-haul time budget).
- Solving every item under 8/10 confidence through evidence, fixes, and verification was explicitly requested.
Requests for "thoroughness" or "deep investigation" alone are not enough
if there is only one stream and no autonomy language.
If authorization is still ambiguous (single stream, no fan-out language, and
spawning would be optional polish), ask one concise question before spawning.
Otherwise spawn without re-asking.
Operating Model
Act as the orchestrator and integrator. Keep the critical path local, and
delegate independent sidecar work that can run while you continue.
Use the host's available subagent interface: Codex multi-agent tools such as
spawn, wait, and close when exposed, or Claude Code's Agent tool when that is
the available mechanism.
Before spawning agents, write a short decomposition:
- Independent evidence questions.
- Independent implementation or test slices.
- Verification surfaces.
- The immediate task the orchestrator will do locally.
Spawn agents early when at least two streams are independent and materially
advance the goal. Do not delegate the task that blocks your next local step.
Fan-Out Rules
Use explorer-style agents for bounded read-only work:
- log and database forensics
- code ownership and blast-radius tracing
- existing issue or PR dedupe
- regression-test discovery
- independent verification of a suspected root cause
Use worker-style agents for bounded changes:
- tests in a disjoint file
- one implementation slice with a clear owner path
- docs or workflow updates
- small scripts or queries
Give every worker an explicit, path-bounded, disjoint write set. Tell workers
they are not alone in the codebase, must not revert unrelated edits, and must
adapt to concurrent changes.
Confidence Ledger
Maintain a visible ledger during the task:
| Finding | Confidence | Evidence | Owner path | Action | Verification | Risk |
|---|
Every item below 8/10 confidence must become one of:
- a concrete evidence query
- a failing regression test
- a bounded implementation task
- a delegated explorer task
- a delegated worker task
- an explicit blocker with proof
Do not final while any item below 8/10 remains merely speculative.
Long-Run Loop
Use long waits only when integration is blocked on a running agent result.
Otherwise continue non-overlapping local work immediately.
After an agent returns:
- Read the result and changed files.
- Verify claims with local commands or direct evidence.
- Integrate or reject the result.
- Update the confidence ledger.
- Close the agent when no longer needed.
If the runtime cannot keep working for the requested duration, leave a durable
handoff artifact such as an issue, task, workflow, or todo with evidence and
the next exact command or query to run.
Completion Gate
Do not send final until:
- all worker agents are complete and closed
- any unfinished agents are read-only and explicitly non-blocking
- all changed code has been linted and tested with fresh commands
- every sub-8/10 finding is fixed, evidence-solved, or blocked with proof
- the final answer names remaining risks honestly
- the final answer reports what was verified, not just what was attempted
Anti-Patterns
- Reporting a ranked list and stopping while repair was authorized.
- Spawning multiple agents on overlapping write scopes.
- Waiting idly while agents run and local non-overlapping work exists.
- Trusting agent self-report without checking evidence.
- Treating historical telemetry counts as reduced by a code change; only new telemetry can prove recurrence reduction.