ultraqa
Skill4 starsChanged todayMITView on GitHub
[oh-my-agy] Adversarial dynamic end-to-end QA: build a hostile scenario matrix, run baseline + adversarial scenarios, diagnose, fix, clean up, and report, in up to 5 bounded cycles. Use for $ultraqa (--tests, --build, --lint, --typecheck, --interactive, --custom \"pattern\") when runnable behavior needs more than a shallow test checklist.
5 other projects use a version of this, 1 of them edited.
Adopted by
Most recently taken up by
- lza6/claw-code-tingfengtheir own version, first seen Nov 1, 2026
- listener-He/java-harness-agenttheir own version, first seen Oct 20, 2026
- zereight/gitlab-mcptheir own version, first seen Oct 8, 2026
- Yeachan-Heo/oh-my-codexedited, first seen Oct 8, 2026
- Jurel89/copilot-omnitheir own version, first seen Apr 17, 2026
Other versions5 projects use a version of this.
Word for word · 0
None found.
Edited · 1
- Yeachan-Heo/oh-my-codex★ 34k29% same
Wrote their own version · 4
- zereight/gitlab-mcp★ 2.1ksame name
- listener-He/java-harness-agent★ 16same name
- lza6/claw-code-tingfeng★ 5same name
- Jurel89/copilot-omni★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository23 other files it gives its agents.
- rules/AGENTS.mdAGENTS.md
- ai-slop-cleanerSkill
- analyzeSkill
- askSkill
- autopilotSkill
- cancelSkill
- code-reviewSkill
- deep-interviewSkill
- deep-researchSkill
- doctorSkill
- hudSkill
- loftSkill
- modernizeSkill
- performance-goalSkill
- prSkill
- ralplanSkill
- releaseSkill
- skillSkill
- teamSkill
- threat-modelSkill
- ultragoalSkill
- visual-ralphSkill
- wikiSkill
The file
We split this skill into the rules it gives an agent
3 rules. 0 are common, 1 is distinctive, and 2 are very specific to this project.
0
Common
Most of the 5 edited versions we read keep these.
1
Distinctive
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 8
Use this explicit opt-in when runnable behavior needs adversarial dynamic end-to-end QA.
- Very specific: line 14
If the user says continue, advance the current verified QA step rather than restarting.
- Very specific: line 20
Before running commands, record a matrix (in .omagy/artifacts/ultraqa-<slug>/matrix.md) with: scenario id, class, trigger status (triggered …
“Common” is measured: kept by at least 6 in 10 of the 5 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
UltraQA Task Card
Use this explicit opt-in when runnable behavior needs adversarial dynamic end-to-end QA. This card defines the QA matrix, evidence contract, and bounded cycling.
When to use and inputs
- Goals:
--tests,--build,--lint,--typecheck,--interactive, or--custom "pattern". Without a structured goal, derive a runnable behavior goal. - Inputs: goal, changed scope, acceptance criteria, runnable command/service, existing tests, and relevant state/cleanup paths.
- If the user says
continue, advance the current verified QA step rather than restarting. - A shallow build/lint/typecheck/test checklist does not satisfy UltraQA. Exercise the requested behavior through adversarial scenarios whenever it can be run, simulated, or harnessed safely.
- Standalone vs. Embedded Ownership: when invoked standalone (
$ultraqa), the current agent executes the QA matrix directly; when embedded inside$autopilotor$ultragoal, the orchestrator delegates QA scenario execution to an independent subagent worker (test-engineer,verifier, orexecutor) and verifies its evidence rather than self-grading its own implementation in the main lane.
Plan and trigger-mapped scenario matrix
Before running commands, record a matrix (in .omagy/artifacts/ultraqa-<slug>/matrix.md) with: scenario id, class, trigger status (triggered or not-applicable: <1-line reason>), intent, user/attacker model, setup, command or harness, expected signal, actual result, failure_class (none | product | harness), fixes, evidence, and cleanup. Evaluate all 9 adversarial classes against their concrete trigger conditions:
| # | Adversarial Class | Trigger Condition (run when true; otherwise log 1-line not-applicable) | Probes |
| 1 | Malformed input | Code parses user, CLI, file, network, JSON, or env input | Invalid JSON, missing fields, bad flags, oversized strings, unusual Unicode, path traversal (../), corrupted state |
| 2 | Prompt injection | Code ingests untrusted text, prompts, transcripts, or external content | Instruction overrides, secret exfiltration strings, verification-skip payloads, fake completion claims |
| 3 | Cancel / resume | Workflow persists state, checkpoints, or supports stop/resume | Mid-run abort, resume from partial state, idempotent re-entry, $cancel cleanup |
| 4 | Stale state | Code reads cached state, lockfiles, timestamps, or session IDs | Expired timestamps (> STALE_MS), foreign owner_conversation_id, schema-version mismatch, contradictory phases |
| 5 | Dirty worktree | Code runs git commands, file generators, or cleanup steps | Pre-existing modified/untracked files outside scope remain untouched after run and rollback |
| 6 | Hung or long-running commands | Code spawns child processes, network calls, or loops | Bounded timeout, clean process termination (manage_task kill), no orphan processes |
| 7 | Flaky tests | Async timing, concurrency, temp files, or randomness in tests | Capped reruns (3x), shared-state isolation, deterministic ordering (a single lucky green does not count) |
| 8 | Misleading success output | CLI/hook/API returns status text or exit codes | Non-zero exit with "PASS" text, zero exit with "ERROR"/test.skip, partial output truncation |
| 9 | Repeated interruptions | Interactive loop, hook nudge, or retry state machine | Repeated continue, rapid duplicate invocations, retry budget exhaustion |
Cycle (maximum 5)
- PLAN ADVERSARIAL QA: goal, success criteria, safety bounds, stop condition, runnable surfaces, matrix.
- RUN BASELINE VERIFICATION:
--testsruns the project tests;--buildruns the build plus built-artifact probes;--lint;--typecheckplus typed harnesses;--customchecks the pattern and exit status;--interactiveuses a bounded CLI/service or browser harness following references/browser-verification-safety.md (await document.fonts.ready+ doublerequestAnimationFramerender settle, console/network4xx/5xxtriage, and OWASP LLM01 untrusted DOM/cookie quarantine). - RUN ADVERSARIAL SCENARIOS: pre-scan custom harness scripts and teardown commands with
sh <plugin>/hooks/run.sh lookout <paths> --json(exitCode: 0required before execution), then capture exit codes, output, artifacts, and cleanup. Userun_commandwith a boundedWaitMsBeforeAsyncfor long runs, andmanage_task killfor hung processes. - CHECK RESULT: pass only when baseline, adversarial scenarios, evidence, and cleanup all pass. Otherwise diagnose and repeat.
- ARCHITECT DIAGNOSIS (via the
architectordebuggersubagent for non-trivial failures): root cause and safety impact. FIX ISSUES precisely. CLEAN UP AND ROLL BACK temporary harnesses, fixtures, logs, processes, state, and failed experiments before the next cycle.
Generate temporary tests, scripts, fixtures, or harnesses only when useful. Keep them under .omagy/artifacts/ultraqa-<slug>/ or the project's test directories. Use bounded runtimes and project-native tools. Write harness files with write_file / write_to_file (not interpolating heredocs). Use absolute repo paths in harness imports. Classify harness setup failures as harness debris: fix the harness and rerun before declaring a product defect.
Safety, state, and exit
No destructive commands, secret exfiltration, credential dumping, production writes, or unbounded process spawning. No unbounded waits. Preserve unrelated dirty work. If a scenario is unsafe, record it as blocked along with the safe substitute. Normalized Repeated-Failure Deduplication: normalize failure output via sh <plugin>/hooks/run.sh failure-sig "<errorOutput>" (which strips ANSI escape codes, pytest/vitest dot-progress & pass-summary lines, Start at clock timestamps, elapsed duration parentheticals & failing summary banner durations, and Node warning PIDs while preserving assertion literals) before comparing across cycles; three repeats of the same normalized failure signature stop the loop immediately with a diagnosis (ULTRAQA STOPPED: Same failure detected 3 times). Cycle 5 stops with residual risks. Success exits after a passing cycle.
State: standalone writes .omagy/state/ultraqa-state.json; inside Autopilot, use state.ultraqa and current_phase:"ultraqa". Phases: planning -> baseline -> adversarial-e2e -> diagnose -> fix -> cleanup -> complete. Update iteration, current_phase, and updated_at as you go (the Stop hook keeps an active run going). Terminal: active:false, current_phase:"complete"|"failed"|"blocked", completed_at.
Evidence/output contract
Return # UltraQA Report with: Goal and success criteria (stop condition, safety bounds); Scenario matrix (all columns); Commands run (exit code, purpose, timeout, key output); Failures found (root cause, user/safety impact); Fixes applied (files, rationale, scenarios, regression evidence); Cleanup and rollback (artifacts/processes/worktree before and after); Residual risks; Evidence (logs, harness output, screenshots where relevant, rerun/flake evidence).
Exit condition
ULTRAQA COMPLETE: Goal met after N cycles only after a passing baseline plus adversarial matrix, clean artifacts, and complete evidence. Otherwise return the exact bounded status: ULTRAQA STOPPED: Max cycles, ULTRAQA STOPPED: Same failure detected 3 times, ULTRAQA BLOCKED: ..., or ULTRAQA ERROR: ..., with the owner and next safe step.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

