behavioral-evidence
Independently verify an accepted objective with real evidence (computer use, browser, the repository's verification CLI) and confirm the change adds no automated tests.
- License 许可
- MIT
- 最近更新
- 2026-11-14
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- behavioral-evidence
- description
- Independently verify an accepted objective with real evidence (computer use, browser, the repository's verification CLI) and confirm the change adds no automated tests.
03 SKILL.md 正文
Behavioral evidence
Real verification is the evidence. Computer use, the browser and the repository's verification CLI and feature map observe the accepted behavior through its public interface; a green check, count or coverage percentage never overrides the accepted objective.
No automated tests
The change adds no automated tests, test runner configuration, test scripts,
test dependencies or CI test steps. Run
development-system check-no-tests --root <repository> --json and report its
findings; automated tests found in the change are deleted, not kept or
strengthened. Report the actual missing behavior and the smallest real
observation that would prove it instead of proposing a test.
Independent verification
The implementer's own checks are never the only evidence. A context-isolated reviewer derives the oracle from the accepted objective and public interface, observes the behavior for real (computer use, the browser, the repository's verification CLI), rejects narrowed or unsupported runs, and reports the verdict against every acceptance criterion, including unproven items. Independent review returns the actual evidence gap; it does not invent a check per form or feature.
Web flows: QuickE2E with Jev (2.3.0)
Drive web flows with QuickE2E on its documented Jev engine. Jev is prepaid
(TypeSafe key at ~/.config/jevgrep/credentials.json), so cost bans on hosted
inference do not apply to it; data limits still do. QuickE2E calls Jev through
OpenRouter, so load the Development System preload that sends the same request
to TypeSafe.
QuickE2E is an external tool, not shipped in this release. Install it once,
pinned, outside any repository; skip this when
node ~/.development-system/tools/quicke2e/bin/quicke2e.mjs --help already
works and git -C ~/.development-system/tools/quicke2e rev-parse HEAD shows
that commit. The same commands move an existing clone to the pin:
[ -d ~/.development-system/tools/quicke2e/.git ] || git clone https://github.com/dmoka/quicke2e ~/.development-system/tools/quicke2e
git -C ~/.development-system/tools/quicke2e fetch origin
git -C ~/.development-system/tools/quicke2e checkout 98f74b6ceb44e8eb68c0bc176e522983639b9521 # 0.5.0
(cd ~/.development-system/tools/quicke2e && npm ci && npx playwright install chromium)
Then run a flow:
OPENROUTER_API_KEY=via-typesafe node --import "$HOME/.codex/development-system/runtime/quicke2e/typesafe-jev.mjs" \
"$HOME/.development-system/tools/quicke2e/bin/quicke2e.mjs" run <flow.mjs> --base <url> --engine jev --trace <private-dir> --video <private-dir>
- Log in with the repository's test accounts (names from its env files, never
printed) and pass a Playwright
storageState. Wait fornetworkidlebefore the first click on dev servers that hydrate late. - Keep the flow file under
$HOME/.development-system/private/verification/<run-id>/, never in the repository. Do not pass--emit: no replay spec is kept, so this is live verification, not an automated test (operator decision 2026-10-13); keeping a replay spec needs an operator test grant. A failed flow is re-run live, so renamed labels need no maintenance. - Prove permissions both ways: the allowed role reaches the end state and a role without the permission is observed refused.
- Assert side effects (rows, emails, payments in test mode), not just text.
- Clean up the records the flow created and stop every server you started.
The T3 browser remains the tool for visual judgment and exploration without a defined flow.
Lines of code, file counts, identifier length and minified formatting are not quality targets. Complexity signals may inform diagnosis but cannot fail a run.
Output
Return the check-no-tests result, each acceptance criterion with its real evidence (passed, failed or not reached), the independent verdict, every self-grading finding and any unproven behavior. This skill is read-only and never edits.
Authorized Claude Code account recovery (1.39.0)
For Claude Code review and computer use, the operator explicitly authorizes the local Codex credential-home sequence followed by a fresh native Claude role. Run codex-review once; it discovers existing profiles without copying credentials and records at most two sequential profile attempts. Read receipt.json as well as findings. When status is fallback_required and fallback.eligible is true, immediately dispatch reviewer (plan, diff or security), visual-reviewer (image review) or browser-qa (computer use) through the native Agent tool with the original packet. Copy fallback.prompt verbatim from the receipt; it contains the original packet and Codex fallback:, Codex fallback receipt: and Codex fallback packet: lines. The guard verifies the full dispatched packet, hash, root, mode and Task-Id and rejects resume so the native agent stays fresh. Computer-use fallback receipts are admitted once only; a consumed receipt requires reconciliation, not replay. Do not let a failed Codex attempt silently omit the independent review or absorb it into the coordinator. Preserve the endpoint and actual capability requirements. Cancellation, safety refusal and uncertain computer-use effects are not fallback permission: reconcile observed state before more actions. If the native role lacks vision or an authorized browser, retain the gap and continue independent work. Report the provider change and actual or unknown identity. This explicit Claude exception does not change restricted Codex-only tasks or Headroom's same-account per-invocation contract. HOME installation does not prove live host loading.
04 同仓库更多 Skills
working-backwards
Define a software feature through PRFAQ and staged documents when the user requests working backwards before coding.
flow-implement
Implement a requested software change through verification, corrections, and its authorized delivery endpoint.
development-steward
Run the scheduled, read-only weekly maintenance review for the six explicitly opted-in repositories and prepare one concise private report for Check-in.
grill-with-docs
Run one recoverable interview for functional, visual, or mixed software grills; reuse equivalent answers and offer the shared HTML questionnaire.
global-agent-guardrails
Install, audit, test, or update the Development System's global destructive-command guard for Codex, Claude Code and T3 Code. Use when configuring agent safety hooks or investigating whether dangerous shell operations are blocked.
impeccable
Design or refine frontend UI using project context and a focused design playbook; also handles visual critique and polish.