feat(skills): hệ skill v1 — ngăn Skill, registry, chạy qua runner (B5) - #120
Merged
Merged
Conversation
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
…-run 8/10, gap-probe, ban bam duoc Ho so B5 truoc Cong Pham vi: design doc (dac ta UX), contract (Vat truoc vong, Tien de TD-1..TD-10 kem ket qua dry-run), 22 eval, 11 dong so quyet dinh, gap-probe 1 P0 + 4 P1 da sua, design-pass 33 khung, ba hinh quyet dinh, the Cong 1. Component ngan skill (trinh bay, chua noi du lieu) + route proto. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…y check Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Real routes over a temp SQLite: closed-set param refusals, missing plugin, unknown skill, the 3-run ceiling with its suppression half, a business-only prompt, per-node steps with the failed node from the engine's envelope, the instance graph for view/edit plan, and missing slots read by the real scanner. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ine delegate Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Stateful panel over the S1-D presentational components: list from the real registry, manifest-generated form with upload, per-step progress from the task's node events, outputs and failure copy by error code, rerun with the same params, and view/edit plan with a replace confirmation. Tests drive the real routes, runner and engine delegate over a temp DB. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… registry entry Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The first real end-to-end run produced both files but an empty result screen:
the engine persists results keyed by node id, while the reader (and E7's
hand-typed fixture) assumed WorkflowOutput names. Manifest outputs now address
{nodeId, field}, the run view reads the engine's real shape, and E7 uses the
result recorded from that run. E16's no-special half no longer scans the
skill's own dir (glob pathspec), and its teeth now include a realistic manifest.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…E15 not-run executed), E14 red Run log, evidence report and review findings from the first verify round, the E12 frames moved into the dossier's evidence dir (the capture agent saved them at the repo root), and the first «thước:» edit of the round. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
S4 round 1 in-contract finding: the integrity check accepted a config target on an existing node with a field its slot does not declare, and E1 only broke input targets. E1 now covers all three target cases plus a real-field control. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…hird «thước:» edit In-contract: config targets on handle-bound fields pass the integrity check and the new control locks that in. E14 red again under S4 load; sdk_pytest red because uv resolves the system Python 3.9. Fixing the in-contract finding needs a third measurement edit this round, which the owner's experiment turns into a stop-and-report. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…Python; E14 wait ceiling Owner decision 16/09 after the S4-r2 stop: change the measurement (third «thước:» edit) — E1 reads the field kind from the template binding, controls on threshold, and refuses a config target on an edge-fed field by name. The shared sdk_pytest key now asks uv for Python >=3.10 as the SDK declares, and the a11y guard waits up to 300s for its server under full S4 load. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…rastructure only No in-contract findings; every contract eval green except E14, whose own server did not come up within 300s under full S4 load. pnpm test red on one out-of-scope test (SDK wheel build racing in sdk/build). Stops at the Evidence Gate as the owner decided. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… signable: BLOCKED) Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
… both green Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ing Evidence Gate signoff Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…ouched runner.ts and api/skills Lane run_id repin-20260917T000600Z-94801, run with UV_PYTHON=">=3.10": the first lane went red on cache-l3-tier-b/E11 only because uv picked the system Python 3.9 (PEP 604 annotations in the test file); no measurement file changed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
run_id repin-20260917T001713Z-38026, 12/12 machine evals + 9 suites green. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
phanlemanh
force-pushed
the
feat/skill-system-v1
branch
from
September 17, 2026 01:02
24686b2 to
310b8f6
Compare
uv picked the system Python 3.9 for every `uv run --no-project` executor in _acceptance/config.yaml, and the SDK tests use 3.10 syntax, so eval lanes went red on infrastructure. uv discovers .python-version from the working directory's parents even with --no-project, so one file covers every SDK key without editing them. CI already sets up Python 3.10 and is unaffected. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Owner decision 17/09. The plugin's source repository has not existed publicly since 26/08, so E14b, E17a and E17b can never go green again. The dossier is not re-signed and the evals are not declared not-run. Kit 2.14.0 has no retired status, so status stays signed-off and the dossier keeps blocking pre-merge as recorded debt that the owner bypasses on merge. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
…tem-v1): re-pin after adding .python-version run_id repin-20260917T010826Z-59027, 59 machine evals + 9 suites green at fcdc6df, run without UV_PYTHON so the lane itself proves the pin file suffices. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Owner
Author
Acceptance Gate còn đỏ vì đúng một nợ đã ghi — owner bỏ qua có ghi nhận khi mergeCI trên Trước khi PR này chạm hồ sơ, cùng hồ sơ đó bị báo bằng dòng «re-pin lane … recorded no evals_exit» (pin suite-only đời cũ, có sẵn trên Vì sao không ghim lại được
Các hồ sơ khác đã sạch
Làn cuối chạy 59 eval máy + 9 suite, xanh, không đặt Merge PR này là việc của owner, với bỏ-qua-có-ghi-nhận cho đúng vi phạm |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Tóm tắt
Roadmap B5 ★ · dossier
skill-system-v1· T3 (chạmsrc/app/api) · Cổng 2 ký 17/09 (Mạnh), phán quyết PASS. Pre-merge của chính hồ sơ: sạch.Người dùng bấm nút Skill ở thanh trái, chọn một việc làm sẵn, đưa video vào và nhận kết quả mà không phải tự dựng đồ thị. Khi muốn, họ mở đúng đồ thị vừa chạy lên canvas để sửa. Bản này có hai skill: Cắt cảnh video và Tách tiếng khỏi video.
Kiến trúc (ADR-0002)
ExecutableWorkflowdựng bằng exporter thật + manifest tham số. Engine không đổi.instantiateghi tham số vào cả bản thực thi lẫnoriginalFlow; test tương đương với exporter.tasksvớifeature: "skill". Runner dựng lại bản thực thi lúc dispatch, từ chối khi phiên bản skill đổi hoặc thiếu plugin.GET /api/skills,POST /api/skills/[id]/run,GET /api/skills/runs/[taskId],GET /api/skills/runs/[taskId]/plan.Bằng chứng
pnpm test(1041 đạt). Hai lệnh này chỉ đỏ khi chạy cùng làn giao diện.failedở bướca1._acceptance/skill-system-v1/(contract, evals, evidence-report, evidence-page.html, decisions.jsonl).Hạn chế đã biết (ký nhận ở Cổng 2)
17 mục ngoài hợp đồng, liệt kê trong
contract.md› Notes. Nặng nhất: đóng rồi mở lại ngăn Skill khi lượt đang chạy làm lượt đó chạy lại, vì luồng chờ mở lại không kèmreconnect=true.Ghim lại kèm theo
cache-l3-tier-bvàconformance-l0bị nhánh này làm cũ (runner.ts,src/app/api/skills/*). Đã ghim lại bằng làn eval, run_idrepin-20260917T000600Z-94801, xanh 30/30 eval + 9 suite. Làn chạy vớiUV_PYTHON=">=3.10": làn đầu đỏ ởcache-l3-tier-b/E11chỉ vìuvchọn Python 3.9 hệ thống.Còn chặn merge (có sẵn trên main, ngoài PR này)
pre-merge-check.sh --base maincòn 3 vi phạm:chong-doc-sai-em-ru,local-cpu-plugins,normalize-text-viđược ghim bằng làn suite-only, thiếuevals_exit. Cần ghim lại bằng làn eval trước khi merge.Lưu ý môi trường
Hơn trăm lệnh SDK trong
_acceptance/config.yamlgọiuv runkhông ghim bản Python. Máy mặc định Python 3.9 sẽ đỏ giả.🤖 Generated with Claude Code