Skip to content

cskwork/supergoal-skill / skills/supergoal

supergoal

Unverified

Route objective-driven development, verification, planning, "teach/explain", codebase learning, and skill evaluation through the Supergoal workflow. Use when Supergoal is requested or an autonomous delivery workflow is needed.

FileKindSize
SKILL.mdskill md12.8 KB
agents/analyst.mdmarkdown1.5 KB
agents/architect.mdmarkdown1.7 KB
agents/code-reviewer.mdmarkdown2.5 KB
agents/db-reader.mdmarkdown2.0 KB
agents/debugger.mdmarkdown2.0 KB
agents/designer.mdmarkdown6.3 KB
agents/executor.mdmarkdown2.8 KB
agents/explore.mdmarkdown2.0 KB
agents/qa-auditor.mdmarkdown4.3 KB
agents/qa-tester.mdmarkdown3.2 KB
agents/security-reviewer.mdmarkdown1.2 KB
agents/skill-forger.mdmarkdown1.6 KB
agents/skill-miner.mdmarkdown1.6 KB
reference/agent-browser.mdmarkdown2.8 KB
reference/arch.mdmarkdown5.1 KB
reference/archify.mdmarkdown3.6 KB
reference/db-access.mdmarkdown2.9 KB
reference/debugging.mdmarkdown6.5 KB
reference/delivery-gate.mdmarkdown6.6 KB
reference/domain-context.mdmarkdown7.8 KB
reference/domain-rules.mdmarkdown1.6 KB
reference/engagement.mdmarkdown3.8 KB
reference/functional-ui.mdmarkdown6.4 KB
reference/harness-eval.mdmarkdown28.7 KB
reference/interview.mdmarkdown7.7 KB
reference/learn-domain.mdmarkdown12.7 KB
reference/market-research.mdmarkdown1.7 KB
reference/observability.mdmarkdown5.0 KB
reference/plan-grounding.mdmarkdown4.3 KB
reference/playwright-cli.mdmarkdown2.6 KB
reference/prototype.mdmarkdown4.5 KB
reference/qa-only.mdmarkdown12.7 KB
reference/qa.mdmarkdown7.9 KB
reference/reporting.mdmarkdown1.5 KB
reference/research.mdmarkdown2.7 KB
reference/review-only.mdmarkdown2.4 KB
reference/role-loop.mdmarkdown26.7 KB
reference/rules.mdmarkdown982 B
reference/skill-mine.mdmarkdown7.6 KB
reference/taste-aesthetics.mdmarkdown7.5 KB
reference/taste-skill-v2.mdmarkdown18.8 KB
reference/taste-sources.mdmarkdown3.0 KB
reference/teach.mdmarkdown28.0 KB
reference/ui-ux.mdmarkdown4.0 KB
reference/vercel-host.mdmarkdown3.6 KB
reference/wayfinder.mdmarkdown7.2 KB
templates/GOAL.mdmarkdown1.3 KB
templates/PLAN.mdmarkdown1.4 KB
templates/QA.mdmarkdown1.2 KB
templates/R-LOOP.mdmarkdown569 B
templates/Z-DONE.mdmarkdown301 B
templates/arch-report.htmlother8.1 KB
templates/commit-gate.shexecutablescript11.6 KB
templates/contrast-gate.mjsexecutablescript4.3 KB
templates/db-access/.env.exampleother451 B
templates/db-access/check-connection.shexecutablescript134 B
templates/db-access/db-access.cmdscript39 B
templates/db-access/db-access.mjsexecutablescript5.2 KB
templates/db-access/read-only-query.shexecutablescript138 B
templates/db-access/schema-summary.shexecutablescript132 B
templates/domain-agent/code-map.mdmarkdown744 B
templates/domain-agent/config.jsonconfig291 B
templates/domain-agent/decisions/README.mdmarkdown270 B
templates/domain-agent/flows/README.mdmarkdown800 B
templates/domain-agent/freshness.mdmarkdown1.2 KB
templates/domain-agent/glossary.mdmarkdown781 B
templates/domain-agent/index.mdmarkdown1.3 KB
templates/domain-agent/invariants.mdmarkdown504 B
templates/domain-agent/qa/README.mdmarkdown879 B
templates/domain-agent/qa/nav-map.mdmarkdown1.2 KB
templates/domain-agent/test-map.mdmarkdown432 B
templates/domain-agent/tickets/README.mdmarkdown204 B
templates/domain-onboarding.htmlother9.2 KB
templates/harness-eval-case.yamlconfig4.4 KB
templates/harness-eval-cases/authored/authored-underspec-001-deepmerge.yamlconfig2.2 KB
templates/harness-eval-cases/authored/authored-underspec-002-csvline.yamlconfig2.0 KB
templates/harness-eval-cases/authored/authored-underspec-003-authz-cache.yamlconfig1.6 KB
templates/harness-eval-cases/fixtures/README.mdmarkdown2.8 KB
templates/harness-eval-cases/fixtures/revfactory-case-002-async-race/package.jsonmanifest124 B
templates/harness-eval-cases/fixtures/revfactory-case-002-async-race/src/cache.mjsscript494 B
templates/harness-eval-cases/fixtures/revfactory-case-002-async-race/test/cache.hidden.test.mjsscript2.2 KB
templates/harness-eval-cases/fixtures/revfactory-case-002-async-race/test/cache.visible.test.mjsscript950 B
templates/harness-eval-cases/fixtures/revfactory-case-003-refactoring/package.jsonmanifest125 B
templates/harness-eval-cases/fixtures/revfactory-case-003-refactoring/src/order.mjsscript1.3 KB
templates/harness-eval-cases/fixtures/revfactory-case-003-refactoring/test/order.hidden.test.mjsscript2.2 KB
templates/harness-eval-cases/fixtures/revfactory-case-003-refactoring/test/order.visible.test.mjsscript1.2 KB
templates/harness-eval-cases/fixtures/underspec-001-deepmerge/package.jsonmanifest124 B
templates/harness-eval-cases/fixtures/underspec-001-deepmerge/src/merge.mjsscript84 B
templates/harness-eval-cases/fixtures/underspec-001-deepmerge/test/merge.hidden.test.mjsscript1013 B
templates/harness-eval-cases/fixtures/underspec-001-deepmerge/test/merge.visible.test.mjsscript276 B
templates/harness-eval-cases/fixtures/underspec-002-csvline/package.jsonmanifest122 B
templates/harness-eval-cases/fixtures/underspec-002-csvline/src/csv.mjsscript77 B
templates/harness-eval-cases/fixtures/underspec-002-csvline/test/csv.hidden.test.mjsscript804 B
templates/harness-eval-cases/fixtures/underspec-002-csvline/test/csv.visible.test.mjsscript240 B
templates/harness-eval-cases/fixtures/underspec-003-authz-cache/package.jsonmanifest126 B
templates/harness-eval-cases/fixtures/underspec-003-authz-cache/src/authorizer.mjsscript601 B
templates/harness-eval-cases/fixtures/underspec-003-authz-cache/test/authorizer.hidden.test.mjsscript3.8 KB
templates/harness-eval-cases/fixtures/underspec-003-authz-cache/test/authorizer.visible.test.mjsscript1.2 KB
templates/harness-eval-cases/revfactory-case-001-rest-api.yamlconfig2.1 KB
templates/harness-eval-cases/revfactory-case-002-bug-fix.yamlconfig2.4 KB
templates/harness-eval-cases/revfactory-case-003-refactoring.yamlconfig2.4 KB
templates/harness-eval-cases/revfactory-case-004-documentation.yamlconfig2.1 KB
templates/harness-eval-cases/revfactory-case-005-complex.yamlconfig2.0 KB
templates/harness-eval-cases/revfactory-case-006-research.yamlconfig2.1 KB
templates/harness-eval-cases/revfactory-case-007-interpreter.yamlconfig2.1 KB
templates/harness-eval-cases/revfactory-case-008-microservice.yamlconfig2.1 KB
templates/harness-eval-cases/revfactory-case-009-sql-engine.yamlconfig2.0 KB
templates/harness-eval-cases/revfactory-case-010-crdt.yamlconfig2.0 KB
templates/harness-eval-cases/revfactory-case-011-raft-kv.yamlconfig2.1 KB
templates/harness-eval-cases/revfactory-case-012-spreadsheet.yamlconfig2.1 KB
templates/harness-eval-cases/revfactory-case-013-bytecode-vm.yamlconfig2.1 KB
templates/harness-eval-cases/revfactory-case-014-event-sourcing.yamlconfig2.1 KB
templates/harness-eval-cases/revfactory-case-015-lsp.yamlconfig2.3 KB
templates/harness-eval-cases/run-local-eval.mjsscript43.3 KB
templates/harness-eval-cases/run-underspec-n3.mjsscript41.1 KB
templates/harness-eval-external/deepswe/README.mdmarkdown9.9 KB
templates/harness-eval-external/deepswe/prepare-supergoal-arm.mjsscript4.4 KB
templates/harness-eval-external/deepswe/run-default-suite.mjsscript6.1 KB
templates/harness-eval-external/deepswe/run-full-cycle.mjsscript25.3 KB
templates/harness-eval-external/deepswe/task-set.yamlconfig12.2 KB
templates/harness-eval-external/featadd/README.mdmarkdown3.9 KB
templates/harness-eval-external/featadd/features.tsvother1.6 KB
templates/harness-eval-external/featadd/gen.shscript4.0 KB
templates/harness-eval-external/featadd/package.pyscript6.2 KB
templates/harness-eval-gate.mjsscript24.9 KB
templates/harness-eval-report.mdmarkdown5.8 KB
templates/harness-eval-result.jsonconfig11.9 KB
templates/harness-eval-runner.mjsscript19.9 KB
templates/harness-eval-stats.mjsscript1.8 KB
templates/learn-grounding-gate.mjsscript6.1 KB
templates/observability/heartbeat.schema.jsonconfig2.5 KB
templates/observability/sg-emit.shexecutablescript5.4 KB
templates/qa-gate.shexecutablescript6.9 KB
templates/qa-only-gate.shexecutablescript5.1 KB
templates/qa-report.mdmarkdown2.8 KB
templates/rules.mdmarkdown301 B
templates/run-state-gate.mjsscript2.2 KB
templates/run-state.jsonconfig474 B
templates/skill-frontmatter-gate.mjsscript3.7 KB
templates/skill-install-audit.mjsexecutablescript3.7 KB
templates/skill-mine/mine.mjsscript7.8 KB
templates/skill.md.templateother568 B
templates/teach-lesson-gate.mjsscript5.4 KB
templates/teach/README.mdmarkdown4.1 KB
templates/teach/assets/lesson-book.jsscript5.8 KB
templates/teach/assets/lesson-template.htmlother6.1 KB
templates/teach/assets/lesson.cssother11.8 KB
templates/teach/assets/quiz.jsscript3.2 KB

templates/harness-eval-external/deepswe/README.md

View on GitHub
# DeepSWE External Harness Eval

Use this lane when local fixtures ceiling out or when the user asks for an independently replayable,
public source-code task. DeepSWE supplies public repositories, pinned base commits, isolated task
containers, and verifier artifacts.

## Decision

Do not claim `u3` proves a harness win. `u3` proves the fixture discriminates: visible checks can pass
while hidden authorization/cache checks fail. A harness win needs paired public benchmark artifacts.

## Forced Default Suite

For difficult SWE or harness-effectiveness claims, do not stop at a single public task unless the user
explicitly names one. Run the forced default DeepSWE suite:

1. `etree-xml-diff-patch` - go; leaderboard rank 76/113 (40.5% overall) but the anchor task: the only one
   with a completed local paired run showing baseline headroom and directional harness lift at codex
   gpt-5.5 low.
2. `cliffy-config-file-parsing` - typescript; rank 31/113 (17.2% overall, gpt-5.5 1/12). Hardest gpt-5.5
   task in the suite; expect binary floors and rely on partial-reward deltas.
3. `csstree-shorthand-expansion-compression` - javascript; rank 15/113 (11.2% overall, gpt-5.5 6/12).
4. `skrub-duration-encoding` - python; rank 14/113 (10.3% overall, gpt-5.5 4/12).
5. `termenv-preserve-ansi-resets` - go; rank 11/113 (9.5% overall, gpt-5.5 5/12); small repo, fast tests.

Selection method (snapshot 2026-07-06): 13,108 included-in-score leaderboard trials across 21 models and
all 113 tasks, scraped per task page. Suite tasks need a LOW overall pass rate (difficult) AND nonzero
gpt-5.5 passes (so the codex gpt-5.5 lane is not a guaranteed floor). Rust tasks are excluded from the
900s low-effort lane because compile time eats the agent budget. Held-out escalation pool when the suite
saturates or floors: `updo-policy-alerting`, `happy-dom-deterministic-intersectionobserver`,
`testem-bail-on-test-failure`, `koota-deferred-mutation-buffer`, `dateutil-rfc5545-timezone-interop`.

`yjs-map-conflict-detection` was demoted from this suite on 2026-07-06: leaderboard rank 101/113 (55.2%
overall, gpt-5.5 11/12) and the local Spark-low paired run had a perfect baseline the harness regressed
by one preservation check - no headroom, so it cannot show harness lift.

Use the suite runner so the same full-cycle baseline/harness protocol is applied to each task:

```bash
node templates/harness-eval-external/deepswe/run-default-suite.mjs \
  --agent codex \
  --model gpt-5.5 \
  --reasoning-effort low \
  --codex-auth-json auto \
  --timeout-seconds 900 \
  --run-root /tmp/sg-deepswe-default-suite
```

One paired run per task is directional only, but it is the minimum default for this lane because it catches
single-task ceiling effects and regressions. A proven claim still needs repeated seeds.

## Default Public Tasks

Primary scoring task: `etree-xml-diff-patch`

- Benchmark task: `https://deepswe.datacurve.ai/data/v1/tasks/etree-xml-diff-patch`
- Upstream repo: `https://github.com/beevik/etree`
- Base commit: `4032e04c8f2e2f35e43ce5d772fcef14a5df4d74`
- Benchmark ref used for this manifest: `3cda4081fed96103a6395de39c85e9b20275e307`
- Why this first for scoring: it is a larger Go feature task covering XML diff, patch application,
  reverse patches, three-way merge, and summaries. It is intentionally broader than the saturated Happy
  DOM bugfix, so it is the default candidate for baseline-headroom/effectiveness runs.

Smoke task: `happy-dom-abort-pending-body-reads`

- Benchmark task: `https://deepswe.datacurve.ai/data/v1/tasks/happy-dom-abort-pending-body-reads`
- Upstream repo: `https://github.com/capricorn86/happy-dom`
- Base commit: `82a0888cb2c87a6123e05424b528f8e8c9b3e426`
- Benchmark ref used for this manifest: `3cda4081fed96103a6395de39c85e9b20275e307`
- Why keep it: it is a real TypeScript async lifecycle bugfix with public source and verifier artifacts,
  so it is useful for checking that Pier, Codex auth, patch capture, and reports work end to end.

Important: the 2026-07-03 Codex pilot is a setup artifact, not a valid paired correctness result. The
harness arm was manually interrupted after elapsed time was observed. A later no-interrupt full-cycle run
completed both arms, but current Codex `gpt-5.5` low reasoning saturated the task: baseline and harness
both reached `reward=1`, `f2p=14/14`, `p2p=165/165`. Use Happy DOM only as a public full-cycle smoke
task, not as proof of harness lift under saturated settings.

## Protocol

Preferred path: use the full-cycle runner. It pins the benchmark, prepares the harness arm, runs baseline
then harness serially through Pier, enforces the declared stop policy, and writes `manifest.json`,
`summary.json`, `report.md`, logs, and Pier job artifacts.

```bash
node templates/harness-eval-external/deepswe/run-full-cycle.mjs \
  --task etree-xml-diff-patch \
  --agent codex \
  --model gpt-5.5 \
  --reasoning-effort low \
  --codex-auth-json auto \
  --timeout-seconds 900 \
  --run-root /tmp/sg-deepswe-etree-full-cycle
```

Use `--dry-run` first to inspect the exact commands without spending model time. Use `--force` only for a
`/tmp/...` run root or a repo-local `docs/experiments/...` run root; the script refuses broader deletion
targets.

Codex runs default to `--reasoning-effort low` in this lane. That holds the compute budget equal across
baseline and harness arms. If baseline reaches perfect public verifier score, the runner reports
`not_proven_no_headroom` instead of a generic win/loss. A scoring run should not stop at a no-headroom
task; use the harder default task or another held-out DeepSWE task until baseline headroom is visible.
The runner also defaults to `--codex-auth-json auto`: when `OPENAI_API_KEY` is absent and
`~/.codex/auth.json` exists, it passes `CODEX_FORCE_AUTH_JSON=1` into Pier without logging the secret.
It also records a non-secret `chatgpt.com` allowlist hint because Codex auth.json uses the ChatGPT Codex
transport rather than the API-key transport.

Manual path:

1. Clone and pin DeepSWE.

```bash
git clone https://github.com/datacurve-ai/deep-swe /tmp/deep-swe
git -C /tmp/deep-swe checkout 3cda4081fed96103a6395de39c85e9b20275e307
```

2. Install the public runner.

```bash
uv tool install datacurve-pier
```

3. Prepare the harness arm. This copies one task and prepends the shipped supergoal reference to
   `instruction.md`; it does not alter verifier files or reference solutions.

```bash
node templates/harness-eval-external/deepswe/prepare-supergoal-arm.mjs \
  /tmp/deep-swe \
  /tmp/deep-swe-supergoal \
  etree-xml-diff-patch \
  .
```

4. Declare the stop policy before either arm runs. Manual post-hoc interruption invalidates paired
   correctness. A budgeted timeout is valid only when the timeout, patch-capture behavior, and scoring
   treatment are written into the run metadata before launch.

```yaml
stop_policy:
  timeout_seconds: 900
  valid_outcomes: [completed, budget_timeout, error]
  manual_interrupt: invalid_paired_correctness
  capture_patch_on_budget_timeout: true
  score_patch_after_terminate_timeout_or_error: true
```

5. Run baseline and harness with the same agent, model, environment, seed policy, and stop policy. The
   commands below
   use the leaderboard-style `mini-swe-agent`; a Codex-native run is allowed only if both arms use the same
   adapter and the report records that choice. If the adapter cannot commit because `.git` is read-only,
   capture the working-tree diff for both arms after the process terminates, times out by the declared
   budget, or errors; do not score `HEAD` for one arm and a working tree for the other.

```bash
pier run -p /tmp/deep-swe/tasks/etree-xml-diff-patch \
  --agent mini-swe-agent \
  --model openai/gpt-5.5

pier run -p /tmp/deep-swe-supergoal/tasks/etree-xml-diff-patch \
  --agent mini-swe-agent \
  --model openai/gpt-5.5
```

6. Record the artifacts for every seed:

- `verifier/reward.json`
- `verifier/ctrf.json`
- `verifier/test-stdout.txt`
- `verifier/run.log`
- `model.patch`
- Pier trajectory metadata
- `harness-metadata.json`
- stop policy and actual process outcome: `completed`, `budget_timeout`, `error`, or `invalid_manual_interrupt`

The full-cycle runner records `runner_timeout` separately when the outer safety timeout kills Pier. That
outcome is `Not proven`; it is not a valid completed or budget-timeout arm.

## Scoring

Report:

- Baseline pass rate: `baseline_reward_sum / baseline_runs`
- Harness pass rate: `harness_reward_sum / harness_runs`
- Absolute lift: `harness_pass_rate - baseline_pass_rate` in percentage points
- Relative lift: `absolute_lift / baseline_pass_rate` when baseline is non-zero
- Partial reward delta when `reward.json` exposes pass fractions
- Token, duration, and tool-call deltas

Minimum language:

- `n=1`: fixture smoke only, `Not proven`.
- `n>=3 paired seeds`: directional pilot.
- `n>=6 paired seeds`: first point where the paired sign-flip test can reach p < 0.05; use paired
  McNemar for binary pass/fail and report the exact p-value.
- Manual interruption after observing progress: invalid paired correctness. Report only diagnostic
  artifacts and rerun with a predeclared timeout.
- Completed tie where baseline is already perfect: `not_proven_no_headroom`. Report it as a valid
  full-cycle reliability check only, then run harder public tasks before claiming harness effectiveness.

## Controls

- Same benchmark ref, task id, upstream base commit, model, agent, and environment for both arms.
- Original DeepSWE task body hash recorded for the harness arm.
- Baseline gets no supergoal reference.
- Harness gets only the approved supergoal files embedded by the adapter.
- Hidden tests, verifier code, and solution patches are not copied into the prompt.
- A costlier harness loses unless correctness or partial reward improves enough to justify the overhead.
- `happy-dom-abort-pending-body-reads` is smoke only after the completed no-interrupt Codex run saturated
  both arms.
- `cliffy-config-file-parsing` remains secondary; it previously exceeded the low-turn budget without a
  patch under the Claude adapter.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall