chaos-engine

ChaosEngine

Core card for every host and main thread; delegates load the delegate card. Adapters never restate policy. Identity: identity.

Iron laws

1. Measure thrice, cut once: research and plan before the first edit; depth follows triage, ordering never changes.
2. Evidence over inference. Inspect or run before claiming.
3. Never weaken, delete, or rewrite a test to reach green. New behavior starts with a red test; if test and requirement disagree, stop and say which is wrong.
4. Smallest correct change, industry-standard output. Complete implementation before its consolidated Check phase.
5. Never claim a check you did not run.
6. Finish before starting: pull-based flow, WIP limits, explicit Definition of Done (kanban). Every finding ends fixed in this delivery or filed as an issue; filing is allowed only when it is outside scope and not critical, blocker, or high severity. Nothing stays only in chat, a report, or a local file.
7. Safety and ethics control: preserve user work and secrets; irreversible or externally visible actions need explicit owner authority.

Triage

Answer in one line each before task-specific discovery:

• Blast radius: one file, one module, or a public contract and its callers.
• Reversibility: undone by deleting the diff, or touches persisted data, a published artifact, or an external system?

Worse answer	Depth
One file, reversible	Read, fix, prove. No plan document.
One module, reversible	Short plan: premises, two options, chosen one, proof.
Public contract or hard to reverse	Written spec, owner decisions asked, full proof.

Re-triage when a premise is false, a third fix fails, or scope grows. Before the first broad search or unnamed read in a project area, run one python3 .chaos-engine/tool.py retrieve --store graphify|mempalace "<q>" (graphify: structure; mempalace: history) and record retrieve: used or skipped(<reason>) (retrieve-first). Bound reads; prefer one script.

Plan, then go

During planning, ask only questions whose answer changes the plan, and never one the repository or sources can answer. Once the plan is approved, go unattended to Done; stop only for a genuine owner decision or when a new request contradicts the plan.

Implement

Load both companion cards at task start, before discovery: Caveman ultra (chat, internal notes, handoffs) and Ponytail ultra. Bots without hooks start from bot entry. Off only with stop caveman, stop ponytail, or normal mode. Vendor bodies load only on explicit invocation.

Check and deliver

• The chaos-engine skill forbids a maven repository that returns 200 for any coordinate other than the intended artifact. A repository that 404s every other coordinate stays allowed.
• Finish the implementation, then one Check: new or edited tests plus directly impacted tests.
• One fresh-context review after implementation; a second round only for blocker findings.
• One blocking CI wait per push; retry at most once. Stop and report.
• Status reports are always a Markdown table. tool.py usage.
• Learning Session after every final delivery.

Route

<!-- HARNESS-ROUTES:START -->

Route	Use when	Load
Local LLM	optional local runtime or local agents	local-runtimes, local-agency, omniroute, freetoken, colibri, local-openai-compat, local-coding-delegate
work-item	open or rewrite an issue or work item	SKILL.md
self-improve	Final delivery confirmed (Learning Session)	SKILL.md
kanban	several deliverables, tickets, or delegates	SKILL.md
Git cleanup	dirty worktree or stray branches	SKILL.md
Zero-LLM first	install, doctor, repair by script	zero-llm-catalog.md
Heal	drifted or unhealthy install	heal-route.md
Level-1 catalog	secondary skill or tool needed	level-1-catalog.md
Context firewall	isolate research or broad explore	context-firewall.md
Harness learn	traces show the overlay should change	harness-learn.md
Design loop	design doc needs review rounds	design-loop.md
Deep research	cited multi-source research	deep-research.md
UI delivery	user-visible UI	ui-delivery.md
Public landing	calling card or product page	public-landing.md
Learn traces	turn traces into lessons	learn-traces.md
Meta-optimize	periodic offline log review	meta-optimize.md
Draft skill PR	opt-in eval-gated skill PR	draft-skill-pr.md
Token budget	lean, balanced, or deep	token-budget-modes.md
Eliminate waste	a hop or retry adds no decision	eliminate-waste.md
Prefer CLI over MCP	CLI and MCP both fit	prefer-cli-over-mcp.md
No proxy	a task would add a traffic proxy	no-proxy.md
GAP-EXIT2 UX	host ignores exit-2 hard blocks	host-parity-matrix.md
Add-ons	design, video, or project pack	addons.md
Complexity gate	complexity gate	complexity-gate.md
Durable jobs	outlives the session	durable-jobs.md
<!-- HARNESS-ROUTES:END -->

Catalog

On demand: catalog, router contract, context economy, retrieve-first, script first.