研究 skill 系统 2026-07 演进：审查加固 + evidence-first 五子系统 + 自更新 + workspace bundle#1
Merged
Merged
Conversation
Intentional behavior fix for B9: confirmed transitions now require confirmed-by/evidence and persist confirmation provenance.
F1: add lightweight retrieval ranking over record fields, payload leaves, and unit markdown; search_records now returns score-sorted hits.
Follow-up to the Top-6 commits (dbf6d7f..b5dc066) after an adversarial review (7 units, 18 findings, 0 blockers). Fixes the should-fix + cheap correctness items found: - docs: repo-analyst/blog-analyst SKILL.md confirm examples now include the B9-required --confirmed-by/--evidence (the documented commands previously failed argparse). - fix: require_confirmation_provenance accepts a bare-string evidence value instead of silently treating it as empty (list|str hint is now honest). - fix: kb.py review_sort_key orders by the true instant (parsed.timestamp()) rather than the isoformat string, so mixed-offset timestamps sort correctly. - test: add non-regression guard (auto_confirmed/pending/rejected still promote without provenance) + bare-string evidence acceptance. - docs: SCHEMAS.md documents the new confirmation{by,at,evidence,method} field and the confirm-provenance gate rule. - plan: OPTIMIZATION_PLAN.md marks Top 6 done (F1 partial: persistent index + alias deferred), records the review outcome and remaining test-coverage nits. 56 tests pass; all skill scripts compile. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Document what the workspace can actually do now, grounded in the real argparse CLI surface (every documented command validated via --help): - New docs/FEATURES.md: a task-oriented, plain-language tour of all 16 skills' capabilities, each with a one-line how-to and the principle behind it, plus the 3 core mental models (knowledge-unit-first, confirmation gate, program orchestration) and a directory map. - Highlights the round-4 additions (review-queue confirmation inbox, confirmation provenance, program dashboard/next, workflow lifecycle commands, ranked full-text search). - README: link FEATURES from Start Here. - SKILLS_GUIDE: refresh the CLI cheatsheet with the new commands and the confirm-provenance flags; note the confirmation-inbox workflow. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
- docs/RESEARCH_WORKFLOW.md: user-facing "what you do vs what the system does" guide — how to expand the KB, iterate ideas, iterate method/ experiments, and form a complete research output, framed around the four things only the user can do (provide sources, confirm, decide, give experiment feedback). - docs/GETTING_STARTED.md: add the missing "Step 0: run kb.py init" — onboarding docs previously never mentioned initialization (a first-run cliff for new users). - README: link both new docs from Start Here. - OPTIMIZATION_PLAN.md Part C: a UX/usability audit (7 user-journey friction lenses) validating "too much manual operation" with grounded evidence (batch-confirm absent → 54 required inputs/week; exact hash-id hand-carried across skills; empty-kb dead-ends), a 4-way burden taxonomy, ranked improvements, Top 5, and a north-star direction. No code behavior changed; docs + plan only. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Records the outcome of reviewing Codex's Top-5 UX commits and folds in the end-to-end user-simulation report: - R1-R6: must-fix defects in the Top-5 implementation (wikilink-lookup regression via widened fuzzy locate_record; batch-confirm parity + pending-guard; last/current ordering + case-insensitive id prefix + dead param; confirm_command triplication; SCHEMAS default-confirmer doc; --all-reviewed limit misnomer). - S1-S7: adopted findings from docs/RESEARCH_SKILLS_UX_SIMULATION_2026-07-04.md (idea select confirmation-semantics P0; sandbox --root safety P1; polished weekly renderer; next-step navigator; synthesizer fuzzy recall; paper-note coverage/reading-status; repo readiness grading). - Governance red line verified intact (evidence required; auto never self-signs). Next Codex batch scoped to R1-R6 + S1 + S2; S3-S7 adopted for a later batch. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Adversarial review of R1-R6+S1+S2 (90 tests, 0 blockers, governance verified intact). Records the F-group follow-ups: F1 orchestrate-auto must thread --root/RESEARCH_PROJECT_ROOT to its subprocess children (S2 gap in its own fan-out); F2 kb-browser scripts expose --project-root not --root; F3 add the missing end-to-end test pinning the empty-evidence batch-confirm rejection (governance red line); F4 test-strengthening nits. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
§18: agreed design (with user) for a lightweight capture→recall→promote memory that extends skill-evolution-advisor. Two audiences (skill-defect = record-only for user review, agent never auto-changes skills; user-preference /recurring-issue = agent memory). Confirmation-gated; recall surfaced at session start; confirmed prefs promoted into runtime-preferences. §19: necessity triage of all outstanding plan items into DO-NOW (must-fix + pure-positive low-risk cleanup/tests) vs HOLD-for-user (large refactors, new features, behavior-changing items). Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Verified: 384 green; reporting_style read from user-profile (config.py untouched): 详细=20 events/32 lines vs 简洁=5 events/17 lines, missing-markers preserved in both; diagnosis evidence verified against own run artifacts (valid passes, fabricated 'not verbatim' rejected), claimless diagnosis still works. Gates untouched. Real kb/ untouched. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
….1.0) Verified: 393 green; NO push anywhere (fetch/pull --ff-only/clone --depth only); offline-safe (status unknown, no crash); kb update check-only default, apply agent-mediated (NEXT FOR AGENT, never auto-applies); user output natural-language only (no raw commands); kb doctor shows skill_version; install.sh/ws_sync.py UNTOUCHED (invoked, not edited); real kb/ untouched. manifest version field deferred (install-UX track in flight). Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
…earer scope/confirm/finish Verified (acceptance): bash -n OK, git diff --check clean, test_installer.py 9/9, full suite 400 green; drove install/update/uninstall on a temp workspace — .agents + manifest land, update non-destructive, uninstall removes managed config while kb/ + .venv/ + user files survive; duplicate install cleanly rejected (use update). Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
…all + atomic + source/target Verified independently on temp workspaces: allowlist ships only skills+lib(no tests)+ VERSION+LICENSE, EXCLUDES tests/ + eval_research_value.py + untracked (git ls-files); merge-install preserves user's own skill + AGENTS.md prose (managed block); wrong-cwd non-interactive REFUSES without touching source repo (0 dirty); reinstall + update + uninstall all preserve kb/.venv/user files; no half-install (rollback). 404 green. Real kb/ untouched. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
…a + docs reposition + temp scrub + CI Verified: 405 green; skill_validator exits 0 over all 17 skills (blog-analyst 274->130 char description fixed; 25-64 short_description cap); Navigator hardcoded slug removed (kb.js title.trim() fallback); public temp/ refs scrubbed (README->SCHEMAS.md, evidence.py docstring, skill-evolution SKILL.md dev-eval section removed); KB coupling in SKILL.md examples decoupled (p-openvla->p-example, czx->research-lead) — metadata/example only, no skill logic/triggers changed; residual temp/ ref only in eval_research_value.py which is allowlist-EXCLUDED from installs. install.sh/ws_sync untouched. CI runs validator. Real kb/ untouched. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Found in whole-bundle acceptance: skill_validator.py (dev/CI tool, not imported by any runtime skill) shipped because the allowlist includes .agents/lib/research/ wholesale. Added to EXCLUDED_NAMES alongside eval_research_value.py. updater.py stays (runtime — kb update uses it). Verified: fresh install excludes validator, keeps updater, 17 skills, 405 green. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
The deferred 3.15 item: manifest now stamps the bundle semver (read from source .agents/VERSION) alongside source_commit, for local display/provenance without a fetch. schema stays 1 (additive — existing manifests still load; the field is optional on read). Wired at all 3 build_manifest call sites (install/update/reinstall) via read_source_version helper. Verified: fresh install records 0.1.0, update + reinstall preserve it, 405 green. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
Two CI-only failures (all 9 pass locally): 1. 4 git-checkpoint tests failed with 'empty ident name' — CI runner has no git identity. Added a git config step. 2. 5 installer PTY tests failed 'prompt not found' — install.sh's is_interactive_input() correctly returns false when $CI is set (real-CI safety guard), so the wizard was skipped under GitHub's CI=true even with a real PTY. Run pytest with 'env -u CI' for that step only; the guard itself is unchanged. Verified locally: git-identity set + CI unset -> all 9 pass. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
本 PR 汇总 2026-07 一整条演进线（256 commits，全部经 SSOT-first 设计 → codex 施工 → 独立验收 → 回写循环），最终套件 405 passed。真实
kb/用户数据全程零改动；每一波带撤销点 tag。采纳 codex 全系统审查
kb <verb>，无裸命令/--flag/${…}/内部路径/NEXT FOR AGENT；伪 CLI 不依赖 TTY。P0/P1 + 三大契约
kb undo/kb restore/kb resume。status误建 program、kb next吞参、输出清洗、installer PDF backend。你追加的两点
kb init偏好设置改 agent 中介问答 + headless 落盘（修isatty静默跳过）。scaffold → evidence-first（五子系统实质闭环）
survey / idea 陪练 / method（读资源缩放矩阵）/ experiment（强类型指标 + 验 artifact + baseline 对比）/ report（自包含 + outline），全部改成 paper.py 式 prepare/verify + 逐字证据。成熟度 scaffold→beta。
per-paper-type element sets
初筛 agent 分类 paper_type + 3 套要素集（method_system/benchmark/survey），未知兜底。
自更新 + 版本号
.agents/VERSION(0.1.0) +updater.py+kb update（check 默认、agent 中介 apply、只 fetch/pull --ff-only/clone、绝不 push）+kb doctor显示版本。multi-skill workspace bundle（8 点规格）
git ls-filestracked-only，排除 tests/eval/validator）；合并式安装（保留用户自带 skill + AGENTS.md prose，managed-block）；原子无半安装；错误 cwd 拒写不回退源仓。skill_validator.py过全 17 skill；公开面清temp/引用；SKILL.md 示例 KB 耦合解耦。version字段落盘（install/update/reinstall 三处）。验收纪律
每一波独立复现承重声明（实机跑 prepare/verify 证明证据门拦得住编造、实机驱动 install/uninstall 证明 kb 保留），从不 merge-on-summary。撤销点 tag
pre-*-merge-20260717覆盖每个 merge 点。🤖 Generated with Claude Code