agentic-flow-kit 1.8.7 → 2.0.0

package/locales/en/overrides/.agents/skills/verify-ui/SKILL.md

@@ -0,0 +1,116 @@
+---
+name: verify-ui
+description: UI change verification (read-only verification, no dirty data left). First choice chrome-devtools MCP — take_snapshot to explore shape + single-step drive + screenshot evidence; harness .mjs scripts are only a fallback path for regression/unattended runs, and any fallback must record one line of reason in the verification report.
+---
+
+# verify-ui — UI verification guide
+
+## Division of labor
+- This file carries only the method (environment / selection points / evidence / cleanup); the execution contract (environment safety / no self-repair / HALT etiquette / deliverable shape) is single-sourced in `.agents/roles/ui-verifier.md`.
+- Delegated subagents follow the contract; the main agent verifying itself follows this file.
+
+## Trigger conditions
+- User requests: `verify` / `UI verification` / `browser verification`.
+
+## Tool path priority
+- Interactive UI verification defaults to chrome-devtools MCP: the browser instance is reused throughout; first `take_snapshot` to explore shape, then drive single-step, retry locally.
+- harness scripts are used only for: regression / anti-recurrence assets; batch assertion summary + exit-code gates (`makeRecorder`); unattended / pipeline batches.
+- Degradation ladder (established 2026-11-04, revised 2026-11-11): chrome-devtools MCP is the first choice; harness probes are only a fallback path. **Host has no chrome-devtools tools (MCP not installed / not connected)**: when attended, HALT and prompt the user to install / enable chrome-devtools MCP, do not auto-degrade (probes are too slow); only unattended / overnight batches degrade directly to harness headless probes. **MCP installed but not working** (browser cannot start and still fails after handling per runtime-env §6.4 / critical steps blocked by MCP errors): degrade to harness headless probes. Every degradation must record one line of "degradation reason" in the verification report — a probe verification without a degradation reason = path violation, the report is not accepted.
+- Script technical details single source = `scripts/verify/e2e/harness.mjs` (export list: `grep '^export' scripts/verify/e2e/harness.mjs`); when writing scripts import its exports, do not hand-write equivalent logic.
+
+### Capability mapping (harness ↔ MCP)
+
+| Goal | MCP | harness |
+|---|---|---|
+| Session seeding | API fetch token → `evaluate_script` write `auth_token` / `auth_user` → `navigate_page reload` | `getToken` + `seedSession` |
+| Navigation / waiting | `navigate_page` + `wait_for` | `goto` / `hardGoto` / `reload` |
+| Shape exploration | `take_snapshot` (a11y tree + uid) | `inventory` / one-off probe |
+| Driving | `click` / `fill` / `press_key` | `clickText` / `setInput` / `pressKey` |
+| Network capture | `list_network_requests` | `installNetCapture` + `apiCalls` |
+| Console capture | `list_console_messages` | `installConsoleCapture` + `consoleErrors` |
+| Metrics / screenshots | `evaluate_script` / `take_screenshot` | `ctx.evaluate` / `ctx.screenshot` |
+
+### Session and capture rules (same criteria for both paths)
+- Reach the FE origin before injection; same-document hash switches do not reload the document, after seeding a full-page reload is required.
+- Default identity = JWT claims of the seeded token; only when testing "an identity different from the token" manually seed `auth_user` (take the login response's `jobNo` / `displayName` / `dept`).
+- Synthetic JWT only applies to pages calling zero-authorization APIs: `exp = Math.floor(Date.now()/1000) + 1800`, never write far-future; after injection go through about:blank and then hard-navigate.
+- If the page issues token-bearing APIs and 401 clears the session back to `/login`: for real data dependencies use account-pool login (account pool single source `.agents/notes/runtime-env.md` §6.1a).
+- Exclude OPTIONS preflight from request counts; assertion counts use API responses, do not count DOM (virtual scrolling / virtual columns render only the visible window).
+- Before asserting "zero console errors", filter the application's own noise (AG Grid license banner, Network Error from API not started); filter by content features, do not use `^` line-start anchors.
+
+### MCP session differences
+- MCP uses a persistent profile (localStorage retained across rounds): clear localStorage before seeding, or pass `isolatedContext` to `new_page`.
+- On "The browser is already running for …chrome-profile" → kill residual processes whose command line contains `chrome-devtools-mcp` (handling recorded in runtime-env §6.4).
+- `take_screenshot`'s `filePath` is limited by the MCP workspace whitelist (some hosts reject both repo paths and session directories, empirically 2026-10-31): when writing is rejected do not retry paths repeatedly, switch to **inline screenshot without `filePath`** for archiving; evidence validity is based on the `.md` report's DOM/API assertions (spec acceptance criteria never recognize image paths anyway).
+- `upload_file` triggering `input[type=file]` may likewise be fully rejected by the whitelist for repo/session paths (empirically 2026-11-07): do not try paths, inject in-page instead — a temporary read-only HTTP server (e.g., `npx serve` on a random port) serves the fixture, `evaluate_script` fetches it into a File, puts it into `DataTransfer`, then `dispatchEvent(new Event('change', {bubbles:true}))` to the antd upload input; before injection record `modalCount` as 1 to prevent stale modal DOM misalignment (MTO modal residual after close previously caused PML injection misalignment; reload clears residue). Verify: the uploaded item tag/preview API receives file rows matching the fixture row count.
+
+## Workflow
+
+### 1. Environment preparation
+- Start API / frontend per `.agents/notes/runtime-env.md`; record all started processes.
+- Multiple worktrees in parallel: set `SHIPYARD_SLOT` so the three port groups API / frontend / CDP shift together (port literal single source `scripts/verify/e2e/ports.mjs`, port table in runtime-env §6.1b).
+- Only harness / fallback paths start their own browser: newer Chromium kernel headless (e.g., Edge `--headless=new`); on the MCP path the browser instance is managed by MCP, do not start another manually.
+- Start long-lived services as separate processes (e.g., PowerShell `Start-Process`); no pipe at the end of the start command.
+
+### 2. Execute verification
+- Select points from the feature map (`features/`); run at least one normal path and one failure/boundary path each (validation / empty state / permissions).
+- Navigation: hash mode uses `goto`; deep links (Query changes) use `hardGoto` (about:blank relay).
+- Data-condition assertions for paginated tables inside modals: click `.ant-pagination-item` page by page, scan all, then merge assertions; do not look only at the first screen.
+- For layout assertions first probe the real DOM (snapshot / probe reading `outerHTML` + `getComputedStyle`); do not infer expected values from CSS source; when unsure use invariants decoupled from spec numbers, downgrade spec reconciliation to a NOTE.
+- Evidence capture: screenshots, network payloads, console logs; record paths in the task document's "Acceptance criteria" item afterwards.
+
+### 3. Cleanup and wrap-up
+- Delete temporary data precisely by the ID returned in the POST response; fuzzy-match deletion forbidden.
+- Close all processes started by this flow.
+- Report by `page / step / expected vs actual / evidence`; append "feature map coverage M/N" (M = features covered this time, N = total `features/` entries) on the last line, listing uncovered items one by one.
+- Record defects only; do not directly fix code.
+- Write the UI verification report to `workflow/evidence/<YYYY-MM-DD>-<topic>.md`: header records environment / reproduction commands, assertion table each row carries a "corresponding acceptance An" column (shape reference `workflow/evidence/2026-10-24-purchase-contract-template-print-ui.md`). A report left in the session or `tmp/` = verification not closed out.
+
+## Feature map (features/)
+- `features/` holds one file per core feature; four sections, none may be missing when creating: reach path / drive / observable end state / Gotchas.
+- Verification and close-out spot checks take coverage from the map; current coverage follows the `features/` directory.
+- Create files synchronously when a new page or new main chain goes live; source waves from `frontend/src/router/index.ts` routes: route exists, map missing = to add; map exists, route missing = to delete or rewrite.
+- Coverage gaps / skeleton: `node .agents/skills/verify-ui/scripts/gen-verify-feature.mjs --check` (or `--json`) lists uncovered routes; `--route <path>` writes only the four H2 sections + TODO skeleton, inventing drive steps forbidden.
+- Monthly drift maintenance: ① take routes / component diffs of the recent month as source waves; ② run verification once only for touched features; ③ fix in at most one commit.
+- For each failure first decide "map drift" or "code defect": the former fixes the map, the latter enters an incident; mixing them is forbidden.
+- Missing prerequisites → record as a failure and write the reason into that feature's Gotchas; weakening the verification path to pass is forbidden.
+- If a page / behavior does not match the map → fix the map first, then conclude.
+- Run `check-regression-consistency.mjs` before close-out: `#/route` / `frontend/src/*` / `workflow/*.md` / `scripts/verify/*.mjs` referenced by the map dangling = red; dangling code symbols are only advisory.
+
+## Technical requirements and constraints
+
+### Shell and processes
+- Protocol prefers HTTP, avoid self-signed certificate interception.
+- "Partial changes take effect" → restart the dev server; single file not taking effect while others are normal → first touch that file.
+- Probes must hit this worktree's dev server: if the vite port is occupied it silently shifts — after starting the service confirm the actual port and run with `E2E_FE_URL=http://localhost:<actual port>`; login-page navigation uses the same `FE` (localStorage is isolated by origin).
+- Truncate output with `| tail -N`, never `| head -N` (SIGPIPE terminates the process before finally, leaving a residual browser occupying the CDP port).
+- Reset the Console / Log collector before every navigation; install the toast observer before navigation actions.
+- The three-piece verification script set is mandatory: `installNetCapture(ctx)`, `seedSession(token, { reload: true })`, `hardGoto(url)`.
+- Wrap the probe body in `try { … } finally { await ctx.close?.() }`; before script start, kill residual processes by `--remote-debugging-port=<port>` for self-healing.
+- Before using `harness.dbQuery`, confirm the database-explorer skill is available on this machine; prefer criteria that can be established without DB connection.
+
+### Assertion summary
+- Wrap the assertion body in try/catch: in catch append a failed assertion (with error message) and count it in the summary, with exit code 1 accordingly; "run halfway + print GREEN + exit code 0" is forbidden.
+
+### DOM location and interaction
+- Form location inside overlays / drawers / modals must always be scoped to the overlay root (e.g., `.pml-detail .ant-form-item`); global `querySelector` / `querySelectorAll` forbidden.
+- Locate input boxes by placeholder within the target container.
+- Normalize both sides of text comparisons with `.replace(/\s+/g,'')` (antd inserts spaces into two-character Chinese buttons).
+- Visibility criterion uses `getClientRects().length > 0` or `getBoundingClientRect().width > 0`; `offsetParent` forbidden (always null under a `position: fixed` ancestor).
+- The `a-select` clear icon renders only on hover: dispatch mouseenter / mouseover first, wait one tick, then click.
+- The antd pagination size-changer dropdown cannot be opened via synthetic events (focus + dispatched keydown Enter is also unstable): the real path = click the combobox uid with the click tool; measure with the invariant "each option `scrollWidth <= clientWidth+1` and dropdown width ≥ widest option scrollWidth"; if the dropdown animation has not finished, measuring yields 0 size, wait ≥600ms before measuring (2026-11-06 PML pagination verification).
+- Small controls inside modals (e.g., paginator) are squeezed out of the a11y snapshot by the large table behind (output over limit truncation makes the uid unavailable): temporarily `setAttribute('aria-hidden','e2e-hide')` on the background table container to shrink the tree, obtain the uid, perform the real click, then remove the attribute; be careful not to hide the modal's own table as well (2026-11-06 PML verification).
+- If an `a-select` cannot be selected via synthetic events → use real mouse + keyboard: CDP `Input.dispatchMouseEvent` opens it, `ArrowDown` + `Enter` selects; between multiple selects press `Escape` first to close the previous dropdown; query options limited to the last visible dropdown (`.ant-select-dropdown:not(.ant-select-dropdown-hidden)`).
+- Locate segmented / radio / tabs by rendered form (e.g., `.ant-radio-button-wrapper`), do not only look for `<button>`.
+- Table selection state: clear leftover checkmarks by deduplicating rows by `data-row-key` and clicking row by row, and execute in two ticks separately from subsequent selection; clicking the header select-all in indeterminate / none state is "select all", not "clear".
+- The first tbody row of an antd table with `:scroll={y}` is `.ant-table-measure-row` (height 0, no `ant-table-cell` class): skip it for row-height / cell measurements.
+- `ctx.evaluate` already returns an object, do not wrap it again in `JSON.parse`.
+- "Back"-type buttons on the detail overlay use `history.back()`: empirically clicking it goes back to the previous step or even `about:blank` (hit during 2026-11-03 purchase order detail verification) — to return to the list use refresh or direct re-navigation; never click "Back" as a close action.
+- Before asserting, confirm the overlay content has rendered (header button group appears): a document created in a parallel session may fail to load detail, leaving only the "Back" button in the overlay (empirically 2026-11-03) — switch to a known sample document, do not judge a feature defect from this.
+
+### AG Grid list page
+- Cell overhead is 32px (padding 15×2 + border 2), not 16 / 24.
+- Measure width with a hidden DOM span (copy `font` / `letter-spacing`), never canvas; fixed width `ceil10(max(header requirement, content p90 requirement) + 32 + 8)`, outliers delegated to `tooltipShowMode:'whenTruncated'`.
+- Judge truncation directly by comparing `.ag-cell`'s `scrollWidth > clientWidth`; measure horizontal overflow with `.ag-center-cols-viewport`, not `.ag-body-viewport`.
+- Under column virtualization columns outside the viewport are not in the DOM: `.ag-header-cell` count ≠ total column count.
+- Nulls rendered as the literal `-` via `valueFormatter`: cell emptiness checks must explicitly exclude `'-'`.

package/locales/en/overrides/.agents/skills/visual-parity/SKILL.md

@@ -0,0 +1,35 @@
+---
+name: visual-parity
+description: Visual alignment playbook — for pixel-level alignment of two implementations (prototype / old page / design), fix viewport and data, change only styles, classify differences with evidence; for UI restoration tasks.
+---
+
+# visual-parity — Visual alignment
+
+## Trigger
+
+Prototype → implementation, old page → new page, design → page, requiring "looks consistent".
+
+## 0. Fix comparison conditions
+
+Same viewport (width/height + DPR), zoom 100%, same theme, same data (seed IDs fixed). Service online and auth per `.agents/skills/verify-ui/SKILL.md` §1; technical details single source `scripts/verify/e2e/harness.mjs`.
+
+## 1. Capture comparison
+
+Screenshot the same region on both sides (including boundary elements), named `tmp/e2e-evidence/parity-<topic>-{before,after}.png`. Before screenshotting reset scroll position, disable animations and tooltips.
+
+## 2. Diff classification (count item by item)
+
+Spacing / font size and weight / color / layering and shadows / border radius / wrapping and overflow / scrollbars. Rendering noise (font antialiasing, subpixel) listed separately, not counted as to-fix.
+
+## 3. Iterate
+
+Change only styles and layout, not data, not backend, do not import prototype fonts. Re-screenshot the same region each round; diff items only decrease, never increase.
+
+## 4. Criteria and evidence
+
+To-fix items = 0 counts as aligned; do not claim "consistent" before clearing. Evidence: both sides' screenshot paths + viewport + diff table (item / before / after / status).
+
+## Red lines
+
+- Do not weaken the verification path (`.agents/skills/verify-ui/SKILL.md` "weakening the verification path to pass is forbidden").
+- Do not substitute "it looks the same to the eye" for the diff table.

package/locales/en/overrides/.agents/skills/why/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: why
+description: Trace "why it looks like this" — search source history, process trail, domain docs, session records, issues/PRs, and external evidence in parallel, giving a pointer for each finding; use before changing code you do not understand.
+---
+
+# why — Why it looks like this
+
+## Triggers
+
+- "Why was it designed this way / why is this flag off / why was this option chosen".
+- You are about to change code you do not understand and first need to know why it exists.
+- During review you are uneasy about the intuition "this line cannot be deleted" and need evidence.
+
+## Core
+
+**No memory, no guessing.** Search in parallel by evidence category, and give every conclusion a pointer (`file:line` / commit / doc / issue / metric). Distinguish **measured / inferred / guess**; not finding anything is also an answer—write "no record", do not invent.
+
+## Evidence categories (search in parallel)
+
+| Category | Entry point in this repo |
+|---|---|
+| Source and history | `git log -S<term>` (pickaxe), `git log --follow -- <path>`, `git blame`, `xd://lsp` references |
+| Process trail | `workflow/{intents,specs,plans,incidents}`, `workflow/archive/`, `workflow/delegations.md`, `workflow/papercuts.md`, `workflow/regression-checklist.md` |
+| Domain knowledge | `wiki/`, `node .agents/scripts/kb-search.mjs <term>` |
+| Session records | `history://`, `archive` (eval global), `.agents/skills/session-resume/SKILL.md` |
+| Issue / PR | `issue://<N>`, `pr://<N>` (when a GitHub remote exists) |
+| External docs | `microsoft_docs_search`, `web_search`, `read <url>` |
+| Runtime / monitoring | available MCPs (chrome-devtools, context-mode, etc.), logs, `workflow/metrics.md` |
+
+## Steps
+
+1. **Break down the subject**: split the question into searchable terms—symbol names, file paths, feature names, decision keywords. If one term finds nothing, try synonyms/Chinese/English.
+2. **Local evidence**: run `node .agents/skills/why/scripts/why-evidence.mjs <term> [--path <file>]` to get pickaxe, path history, workflow/wiki hits, and kb-search summaries in one pass.
+3. **External evidence**: search issues/PRs, external docs, and monitoring in parallel via available MCPs; skip if unavailable—do not pretend you searched.
+4. **Synthesize**: give the conclusion first (why), then the evidence chain (each claim with a pointer), and finally unknowns/conflicts.
+5. **Mark certainty**: `measured` (ran it / read the original), `inferred` (derived from evidence), `guess` (no evidence). Predictions and unobserved causes are always `guess`.
+6. **Timeline** (optional): when a decision spans multiple commits/docs, order a chain by time and mark the turning points.
+
+## Output format
+
+- **Conclusion**: one sentence answering "why".
+- **Evidence chain**: each = claim + pointer (`file:line` / commit / doc / issue) + certainty.
+- **Unknowns / conflicts**: parts that cannot be found, evidence that contradicts itself, or that needs the owner to decide.
+- **Implications for the change**: whether this can be touched, and what must hold before touching it.
+
+## Where it lives in this repo
+
+- Takeover/continuation: after `session-resume` rebuilds the current state, use this skill to fill in "why".
+- Bug fixing: before `diagnosing-bugs` finds the root cause, confirm why this code exists.
+- Review: "Impact proof" in `review.md` uses `blast-radius`; this skill adds "historical intent".
+- Evidence collector: `.agents/skills/why/scripts/why-evidence.mjs` (collects only, does not conclude).
+
+## Boundaries
+
+- This skill answers "why it looks like this"; `blast-radius` answers "what it will break"; `how` answers "how it works". The three do not replace each other.
+- Do not write speculation as fact; without evidence, write "no record".

package/locales/en/overrides/.agents/skills/wiki/SKILL.md

@@ -0,0 +1,53 @@
+---
+name: wiki
+description: Project wiki organization and verification skill. Responsible for moving scattered documents into wiki/ and maintaining INDEX.md, 知识沉淀总览.html, and archive consistency.
+---
+
+# Wiki Organization and Verification Skill
+
+## Goal
+Build and maintain a structured project knowledge base, ensuring three-way consistency among disk files, index files, and the visual board.
+
+## Triggers
+- User command: `整理 wiki` / `归类文档` / `建知识库` / `同步知识沉淀看板` / `新增文档到 wiki`.
+
+## Scope and constraints
+- **Inclusion scope**:
+    - **In**: master-data knowledge documents, project-wide data-maintenance SQL.
+    - **Out**: the `workflow/` directory, non-master-data knowledge documents such as MTO/procurement/inventory/UI, dependency lock files, IDE configuration.
+- **Directory structure**: `wiki/<主题>/<关键词文件>`.
+- **Naming convention**: `<主题>-<关键词>[-范围/版本].ext`.
+- **Authority**: `.md` is the content source $\rightarrow$ `.html` is the display snapshot. On conflict, `.md` wins.
+- **Operation limits**: only copy (`Copy-Item`/`cp`). Deleting original files requires separate user confirmation.
+
+## Standard operating procedure
+
+### 1. Organize and place (batch/new)
+1. **Topic determination**: read the first 8 lines of the file and determine the topic from the content. New topics must be aligned with the user.
+2. **Directory creation**: create `wiki/<主题>/` and the corresponding `wiki/drafts-archive/YYYY-MM-DD_<主题>/`.
+3. **File migration**:
+    - Copy the file to `wiki/<主题>/` and rename it per the convention.
+    - Copy the original file into `drafts-archive/` for provenance.
+4. **Index update**: run `node .agents/scripts/gen-wiki-board.mjs`.
+5. **Description completion**: fill in a description for the new topic in the "Purpose" column of the quick-view table in `wiki/INDEX.md`, then run the generator again.
+
+### 2. Add a single document
+1. Determine the topic $\rightarrow$ copy to `wiki/<主题>/` $\rightarrow$ run the generator $\rightarrow$ run the verification script.
+
+### 3. Fix misplacement
+1. Move the file to the correct topic directory $\rightarrow$ run the generator $\rightarrow$ run the verification script. Leave `drafts-archive` untouched.
+
+## Verification mechanism
+Run after every change:
+```bash
+node .agents/scripts/verify-wiki-consistency.mjs
+```
+**Acceptance criteria**:
+- `exit 0` and output `✅ wiki 三方一致`.
+- Checks cover: disk $\leftrightarrow$ INDEX $\leftrightarrow$ board DATA $\leftrightarrow$ data-maintenance ledger $\leftrightarrow$ link reachability.
+
+## Maintenance notes
+- **Generated regions**: the counts/mapping table in `wiki/INDEX.md` and the DATA block in `知识沉淀总览.html` are auto-generated; do not modify them manually.
+- **Manual edit spots**: maintain only the "Purpose" column in `wiki/INDEX.md` and the ledger in `wiki/数据维护/README.md`.
+- **Conflict handling**: when the wiki conflicts with `AGENTS.md` or `workflow`, the latter wins.
+- **Commit convention**: adjusting wiki content (adding/modifying/deleting/renaming/placing knowledge documents, INDEX, boards, this skill file) is a docs-level change, **exempt from the workflow trio** (no intents/specs/plans, no confirmation door); commit directly (`docs(wiki): <action>`); the single rule source = root `AGENTS.md` §4 "wiki exemption". Deleting files still requires separate user confirmation (see "Operation limits").

package/locales/en/overrides/.claude/agents/implementer.md

@@ -0,0 +1,9 @@
+---
+name: implementer
+description: "Implements an approved, tightly scoped code change or test when the task includes owned files and executable acceptance criteria"
+model: inherit
+---
+
+# Implementer
+
+Shared role contract: `.agents/roles/implementer.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.claude/agents/independent-reviewer.md

@@ -0,0 +1,10 @@
+---
+name: independent-reviewer
+description: "Independently reviews a fixed spec, plan, or code diff for correctness and project-rule violations without modifying files"
+tools: Read, Grep, Glob, Bash
+model: inherit
+---
+
+# Independent Reviewer
+
+Shared role contract: `.agents/roles/independent-reviewer.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.claude/agents/pr-adjudicator.md

@@ -0,0 +1,10 @@
+---
+name: pr-adjudicator
+description: "Read-only adjudicator for PR panel review; handles escalated candidates and returns verdicts"
+tools: Read, Grep, Glob, Bash
+model: inherit
+---
+
+# PR Adjudicator
+
+Shared role contract: `.agents/roles/pr-adjudicator.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.claude/agents/pr-finder.md

@@ -0,0 +1,10 @@
+---
+name: pr-finder
+description: "Read-only finder for PR panel review; finds bugs independently per lens and returns candidate JSON"
+tools: Read, Grep, Glob, Bash
+model: inherit
+---
+
+# PR Finder
+
+Shared role contract: `.agents/roles/pr-finder.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.claude/agents/pr-verifier.md

@@ -0,0 +1,10 @@
+---
+name: pr-verifier
+description: "Read-only verifier for PR panel review; independently verifies a candidate and returns a vote JSON"
+tools: Read, Grep, Glob, Bash
+model: inherit
+---
+
+# PR Verifier
+
+Shared role contract: `.agents/roles/pr-verifier.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.claude/agents/ui-verifier.md

@@ -0,0 +1,10 @@
+---
+name: ui-verifier
+description: "Runs browser-based verification for an implemented UI change and returns reproducible evidence without editing repository files"
+tools: Read, Grep, Glob, Bash, mcp__chrome-devtools__navigate_page, mcp__chrome-devtools__new_page, mcp__chrome-devtools__list_pages, mcp__chrome-devtools__select_page, mcp__chrome-devtools__close_page, mcp__chrome-devtools__take_snapshot, mcp__chrome-devtools__take_screenshot, mcp__chrome-devtools__click, mcp__chrome-devtools__fill, mcp__chrome-devtools__fill_form, mcp__chrome-devtools__type_text, mcp__chrome-devtools__press_key, mcp__chrome-devtools__hover, mcp__chrome-devtools__drag, mcp__chrome-devtools__evaluate_script, mcp__chrome-devtools__wait_for, mcp__chrome-devtools__resize_page, mcp__chrome-devtools__emulate, mcp__chrome-devtools__list_console_messages, mcp__chrome-devtools__get_console_message, mcp__chrome-devtools__list_network_requests, mcp__chrome-devtools__get_network_request, mcp__chrome-devtools__handle_dialog, mcp__chrome-devtools__upload_file, mcp__chrome-devtools__get_css_styles, mcp__chrome-devtools__performance_start_trace, mcp__chrome-devtools__performance_stop_trace, mcp__chrome-devtools__performance_analyze_insight, mcp__chrome-devtools__take_heapsnapshot, mcp__chrome-devtools__lighthouse_audit, mcp__chrome-devtools__screencast_start, mcp__chrome-devtools__screencast_stop
+model: inherit
+---
+
+# UI Verifier
+
+Shared role contract: `.agents/roles/ui-verifier.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.claude/agents/wiki-compiler.md

@@ -0,0 +1,10 @@
+---
+name: wiki-compiler
+description: "Translates a code commit into strict JSON describing what business capability the system gained, written for non-coding readers (product/customer service); read-only analysis, never edits files"
+tools: Read, Grep, Glob, Bash
+model: inherit
+---
+
+# Wiki Compiler
+
+Shared role contract: `.agents/roles/wiki-compiler.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.claude/commands/wf-deploy.md

@@ -0,0 +1,8 @@
+---
+description: "Deploy: merge checklist (static gates + pair checks + regression + DB + authorization + PR review/merge)"
+argument-hint: [发布范围]
+---
+
+Authoritative command: `.agents/commands/deploy.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.claude/commands/wf-maintain.md

@@ -0,0 +1,8 @@
+---
+description: "Maintain: incident record + retrospective trio (fix / prevention / rule entries)"
+argument-hint: [事故或问题描述]
+---
+
+Authoritative command: `.agents/commands/maintain.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.claude/commands/wf-new-task.md

@@ -0,0 +1,8 @@
+---
+description: "New Task: closed-loop entry — classify first, then route (L0/L1 fast lane; L2/L3 → wf-stages)"
+argument-hint: [任务描述]
+---
+
+Authoritative command: `.agents/commands/new-task.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.claude/commands/wf-pipeline-run.md

@@ -0,0 +1,8 @@
+---
+description: "pipeline-run automation entry: single command drives the six-stage loop (experimental; state machine unverified)"
+argument-hint: [工单或任务描述]
+---
+
+Authoritative command: `.agents/commands/pipeline-run.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.claude/commands/wf-pr-review.md

@@ -0,0 +1,8 @@
+---
+description: "PR panel review: multi-agent finders + independent votes + ledger"
+argument-hint: [PR 编号或范围]
+---
+
+Authoritative command: `.agents/commands/pr-review.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.claude/commands/wf-review.md

@@ -0,0 +1,8 @@
+---
+description: "Code review standard: risk-graded machine gates + AI self-review checklist; user decides severity and merge"
+argument-hint: [评审范围]
+---
+
+Authoritative command: `.agents/commands/review.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.claude/commands/wf-stages.md

@@ -0,0 +1,8 @@
+---
+description: "L2/L3 staged flow: Plan intent → Design spec → Build plan + implementation → Test static gates + UI verification + independent review + closeout (L1 → wf-new-task)"
+argument-hint: [阶段或任务描述]
+---
+
+Authoritative command: `.agents/commands/stages.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.cline/agents/implementer.yml

@@ -0,0 +1,13 @@
+---
+name: implementer
+description: "Implements an approved, tightly scoped code change or test when the task includes owned files and executable acceptance criteria"
+tools:
+  - read_files
+  - search_codebase
+  - run_commands
+  - editor
+---
+
+# Implementer
+
+Shared role contract: `.agents/roles/implementer.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.cline/agents/independent-reviewer.yml

@@ -0,0 +1,11 @@
+---
+name: independent-reviewer
+description: "Independently reviews a fixed spec, plan, or code diff for correctness and project-rule violations without modifying files"
+tools:
+  - read_files
+  - search_codebase
+---
+
+# Independent Reviewer
+
+Shared role contract: `.agents/roles/independent-reviewer.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.cline/agents/pr-adjudicator.yml

@@ -0,0 +1,11 @@
+---
+name: pr-adjudicator
+description: "Read-only adjudicator for PR panel review; handles escalated candidates and returns verdicts"
+tools:
+  - read_files
+  - search_codebase
+---
+
+# PR Adjudicator
+
+Shared role contract: `.agents/roles/pr-adjudicator.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.cline/agents/pr-finder.yml

@@ -0,0 +1,11 @@
+---
+name: pr-finder
+description: "Read-only finder for PR panel review; finds bugs independently per lens and returns candidate JSON"
+tools:
+  - read_files
+  - search_codebase
+---
+
+# PR Finder
+
+Shared role contract: `.agents/roles/pr-finder.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.cline/agents/pr-verifier.yml

@@ -0,0 +1,11 @@
+---
+name: pr-verifier
+description: "Read-only verifier for PR panel review; independently verifies a candidate and returns a vote JSON"
+tools:
+  - read_files
+  - search_codebase
+---
+
+# PR Verifier
+
+Shared role contract: `.agents/roles/pr-verifier.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.cline/agents/ui-verifier.yml

@@ -0,0 +1,8 @@
+---
+name: ui-verifier
+description: "Runs browser-based verification for an implemented UI change and returns reproducible evidence without editing repository files"
+---
+
+# UI Verifier
+
+Shared role contract: `.agents/roles/ui-verifier.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.cline/agents/wiki-compiler.yml

@@ -0,0 +1,14 @@
+---
+name: wiki-compiler
+description: "Translates a code commit into strict JSON describing what business capability the system gained, written for non-coding readers (product/customer service); read-only analysis, never edits files"
+tools:
+  - read_files
+  - search_codebase
+  - run_commands
+  - editor
+  - skills
+---
+
+# Wiki Compiler
+
+Shared role contract: `.agents/roles/wiki-compiler.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.cline/workflows/wf-deploy.md

@@ -0,0 +1,8 @@
+---
+name: wf-deploy
+description: "Deploy: merge checklist (static gates + pair checks + regression + DB + authorization + PR review/merge)"
+---
+
+Authoritative command: `.agents/commands/deploy.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.cline/workflows/wf-maintain.md

@@ -0,0 +1,8 @@
+---
+name: wf-maintain
+description: "Maintain: incident record + retrospective trio (fix / prevention / rule entries)"
+---
+
+Authoritative command: `.agents/commands/maintain.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.cline/workflows/wf-new-task.md

@@ -0,0 +1,8 @@
+---
+name: wf-new-task
+description: "New Task: closed-loop entry — classify first, then route (L0/L1 fast lane; L2/L3 → wf-stages)"
+---
+
+Authoritative command: `.agents/commands/new-task.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.cline/workflows/wf-pipeline-run.md

@@ -0,0 +1,8 @@
+---
+name: wf-pipeline-run
+description: "pipeline-run automation entry: single command drives the six-stage loop (experimental; state machine unverified)"
+---
+
+Authoritative command: `.agents/commands/pipeline-run.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.cline/workflows/wf-pr-review.md

@@ -0,0 +1,8 @@
+---
+name: wf-pr-review
+description: "PR panel review: multi-agent finders + independent votes + ledger"
+---
+
+Authoritative command: `.agents/commands/pr-review.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.cline/workflows/wf-review.md

@@ -0,0 +1,8 @@
+---
+name: wf-review
+description: "Code review standard: risk-graded machine gates + AI self-review checklist; user decides severity and merge"
+---
+
+Authoritative command: `.agents/commands/review.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.cline/workflows/wf-stages.md

@@ -0,0 +1,8 @@
+---
+name: wf-stages
+description: "L2/L3 staged flow: Plan intent → Design spec → Build plan + implementation → Test static gates + UI verification + independent review + closeout (L1 → wf-new-task)"
+---
+
+Authoritative command: `.agents/commands/stages.md`
+
+This file is a pointer shell. Read the authoritative source at execution time; do not copy the body here.

package/locales/en/overrides/.codebuddy/CODEBUDDY.md

@@ -0,0 +1,7 @@
+# Project Memory
+
+Single source of truth: root `AGENTS.md` (read it fully before starting). Host surface: this directory's `commands/`, `agents/`, `rules/`; hooks are wired via this directory's `settings.json` to `.agents/hooks/` — the host layer does not own scripts.
+
+- Before touching `backend/` or `frontend/`, read the corresponding `AGENTS.md` (lazy load).
+- **Do not** place `CODEBUDDY.md` at the repository root (it takes precedence over `AGENTS.md` without merging and would override the authority).
+- English is the language for this generated pack; commits follow Conventional Commits.

package/locales/en/overrides/.codebuddy/agents/implementer.md

@@ -0,0 +1,9 @@
+---
+name: implementer
+description: "Implements an approved, tightly scoped code change or test when the task includes owned files and executable acceptance criteria"
+tools: Read, Glob, Grep, Edit, Write, Bash, LSP
+---
+
+# Implementer
+
+Shared role contract: `.agents/roles/implementer.md`. This file keeps only the host frontmatter.

package/locales/en/overrides/.codebuddy/agents/independent-reviewer.md

@@ -0,0 +1,10 @@
+---
+name: independent-reviewer
+description: "Independently reviews a fixed spec, plan, or code diff for correctness and project-rule violations without modifying files"
+tools: Read, Glob, Grep
+disallowedTools: Edit, Write, Bash
+---
+
+# Independent Reviewer
+
+Shared role contract: `.agents/roles/independent-reviewer.md`. This file keeps only the host frontmatter.