skillissue.sh

house / Long-Haul Parallel Repair

✓ hosted skill bundleorchestrationv1.1.0MIT5.1 KB
skills/long-haul-parallel-repair/SKILL.md

Agents would either under-delegate or fan out without authorization. This skill is the contract for when "work for hours" is real.

InstallCLI + MCP
CLIRun from your local shell.
$autovault add jack-arturo/skillissue@775e1e1620d2730c8a67e8b7958448e3e888dfc2:skills/long-haul-parallel-repair/SKILL.md --sync-profiles
MCPPaste into an agent MCP tool call.
>add_skill({ source: "github", identifier: "jack-arturo/skillissue@775e1e1620d2730c8a67e8b7958448e3e888dfc2:skills/long-haul-parallel-repair/SKILL.md" })
Choose CLI for a shell install or MCP for an agent tool call.
Source
Example type
bundle
hosted skill bundle
Bundle files
2
SKILL.md + resources
Resources
1
inspectable source files
Declared agents
4
from frontmatter
Source
house
pinned GitHub bundle
SKILL.mdview raw →
---
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:

  1. ≥2 independent streams that materially advance the goal (default multi-slice / SDD posture — no special phrase required).
  2. 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).
  3. 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:

  1. Independent evidence questions.
  2. Independent implementation or test slices.
  3. Verification surfaces.
  4. 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:

FindingConfidenceEvidenceOwner pathActionVerificationRisk

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:

  1. Read the result and changed files.
  2. Verify claims with local commands or direct evidence.
  3. Integrate or reject the result.
  4. Update the confidence ledger.
  5. 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.
story.md

Why it exists

Agents would either under-delegate or fan out without authorization. This skill is the contract for when "work for hours" is real.

History

1.0.0

Activation requires explicit autonomy language — thoroughness alone is not enough.

How Jack actually uses it

Only when fan-out/autonomy is authorized. Orchestrate; verify; do not claim done early.

What it is not

Not a generic tutorial. It's the house runbook for this machine and these agents.

Bundle contents

Every file declared by this skill is inspectable here. Static resources are previewed from same-origin hosted files; script-like files are shown as text only.

2files
markdownSKILL.md
view raw →

SKILL.md

Primary agent instructions, frontmatter, workflow, and declared resource manifest.

Select a package file to inspect it.

Declared capabilities

network
false
filesystem
readwrite
tools
[Agent, Read, Edit]

No secrets declared by this package.

Public, pinned, and inspectable

This house package is installed from the pinned Git commit shown here. Inspect the source and every bundled file before you run it.

package pin775e1e1sourceGitHub packagecompatibilityclaude-code, codex, autojack, grok
VersionBundleSourcePinRaw
v1.1.0latest2 files · 5.1 KBhouse775e1e1SKILL.md