fix(collection): identity dedup + chunk IN() + kill tree + help TTL + safe CDP cleanup [修复组⑥]

.claude/CLAUDE.md

@@ -0,0 +1,123 @@
+# CLAUDE.md
+
+<!-- Add your custom instructions below. Repowise will never modify anything outside the REPOWISE markers. -->
+<!-- Examples: coding style rules, test commands, workflow preferences, constraints -->
+
+<!-- REPOWISE:START — Do not edit below this line. Auto-generated by Repowise. -->
+## IMPORTANT: Codebase Intelligence Instructions for opencli-admin
+
+> This repository is indexed by [Repowise](https://repowise.dev).
+> Use the MCP tools below for orientation, discovery, and enriched context
+> (documentation, ownership, history, decisions). **Always verify against
+> actual source files before making changes** — the index may be stale.
+
+Last indexed: 2026-07-12 (commit c326a2a)
+### Entry Points
+- `backend/main.py`
+- `odp-rs/crates/odp-ingest/src/main.rs`
+- `odp-rs/crates/odp-store/src/main.rs`
+- `iii/workers/collector-discord/src/main.py`
+- `iii/workers/collector-opencli/src/main.py`
+- `iii/workers/odp-ingest-bridge/src/main.py`
+- `iii/workers/schedule-bootstrap/src/main.py`
+- `odp-rs/crates/odp-contracts/src/lib.rs`
+### Tech Stack
+**Languages:** Python, TypeScript
+**Frameworks:** FastAPI, HTTPX, Pydantic
+**Databases:** Alembic, SQLAlchemy
+**Infra:** Celery, Docker, Docker Compose### Architectural Layers
+| Layer | Files | Purpose |
+|-------|-------|---------|
+| Application | 431 |  |
+| API | 48 |  |
+| UI | 84 |  |
+| Config | 4 |  |
+| Types | 22 |  |
+| Data | 25 |  |
+| Service | 83 |  |
+| Docs & Tooling | 197 |  |
+| Test | 170 |  |
+
+### Guided Tour (12 steps)
+1. `README_HANDOVER.md`
+2. `main.py`
+3. `main.rs`
+4. `main.rs`
+5. `frontend.mjs`
+6. `schema.ts`
+... and 6 more steps
+### Hotspots (High Churn)
+| File | Churn | 90d Commits | Owner |
+|------|-------|-------------|-------|
+| `backend/channels/opencli_channel.py` | 100.0th %ile | 8 | xujinghua |
+| `tests/integration/test_workflow_compile_api.py` | 99.9th %ile | 15 | lunnt |
+| `backend/workflow/opencli_hda_tracer.py` | 99.8th %ile | 6 | lunnt |
+| `tests/integration/test_workflow_opencli_hda_trace_api.py` | 99.6th %ile | 9 | lunnt |
+| `frontend/lib/flow/store.ts` | 99.5th %ile | 7 | lunnt |
+
+## Code health
+Three signals: **defect risk** (the overall score), **maintainability** (smells that hurt readability/change-cost without predicting bugs), and **performance** (static performance RISK: I/O-in-loop / N+1 shapes that waste work, high-precision/low-recall). Maintainability and performance are co-equal views, never blended into the defect headline. See `docs/CODE_HEALTH.md`.
+
+Defect risk, Hotspot health: 5.42/10 (stable) ·
+Average: 8.0/10 ·
+Worst: 1.0/10 (`frontend/lib/flow/store.ts`)
+Maintainability, Average: 8.88/10
+Performance risk, Average: 9.93/10
+
+### Critical biomarkers
+- `frontend/components/flow/workflow-editor.tsx` — churn risk — impact −2.3
+- `tests/integration/test_workflow_opencli_hda_trace_api.py` — change entropy — impact −2.3
+- `backend/schemas/control.py` — change entropy — impact −2.1
+- `backend/config.py` — untested hotspot — impact −2.0
+- `frontend/components/ui/card.tsx` — untested hotspot — impact −2.0
+
+### Repowise MCP Tools
+
+This repo has the Repowise MCP server configured. The tools below answer questions `grep`/`Read` cannot. Every response carries an `_meta` envelope with `index_age_days`, `indexed_commit`, and a `stale_warning` only when the index has actually diverged from HEAD — silence means the index is current.
+
+**When to call which tool:**
+
+| Tool | What only this tool answers |
+|------|------------------------------|
+| `get_answer(question)` | Synthesised answer with citations and a content-grounded `confidence`. First call for "how does X work" / "where is Y" / "why is Z" — and the one-shot for "explain function/method/class X": when the question names an indexed symbol, get_answer anchors its defining file (even when fuzzy retrieval would miss it) and returns the symbol's full live body in `symbol_bodies` — read that instead of a `get_symbol` follow-up. Value questions may return `grounding: "extracted"` (verbatim source line). On low confidence returns `best_guesses` with one-line justifications, plus `code_rationale` (rationale comments mined live from the candidate source) when the "why" lives in a code comment the wiki never captured. |
+| `get_context(targets=[...])` | Triage card for files/modules/symbols — title, summary, signatures, `hotspot` bit, `decision_records` titles, `symbol_id`s. File targets auto-upgrade to a `verified` skeleton (every signature, ~37% of a full Read). `include=["callers"]` works on file targets too (import + call rollup). |
+| `get_symbol(...)` | Source bytes with live-verified bounds. Three forms: `"path.py::Name"` (indexed symbol), `"path.py:140-180"` (live range read, ≤200 lines), `"repowise#<hex>"` (omission ref). Index misses return `fallback_lines` from a live grep instead of a dead end. |
+| `search_codebase(query, mode?, kind?, symbol_kind?)` | Hybrid code search. `mode="auto"` (default) routes by query shape: an identifier → indexed symbol hits (`symbol_id`/`file`/line bounds — pipe into `get_symbol`), a path → file pages (pipe into `get_context`), prose → wiki-semantic search, mixed → hybrid (symbols first). Force a branch with `mode=symbol\|path\|concept\|hybrid`. Concept hits carry `search_method` (`embedding` vs `bm25` fallback); decision records rank below file pages unless the query is why-shaped. |
+| `get_why(query, targets?)` | Architectural decision archaeology — *why* the code is shaped this way. Call before refactors or pattern divergences. Falls back to git archaeology when no ADRs exist for a file, and to `code_rationale` (rationale comments mined from the source) when neither decisions nor git history explain it. |
+| `get_risk(targets, changed_files?)` | What history says about touching these files: churn, owners, blast radius. Pass `changed_files` for PR mode → returns a `directive` (`will_break`, `missing_cochanges`, `missing_tests`). |
+| `get_health(targets?, include?)` | Code-health scores + biomarker findings (defect / maintainability / performance pillars). Self-check before a PR — read the same signals the merge-gate judges your change on. Default is lean; opt in with `include`: `["accuracy"]` (does the score rank the buggy files first — precision@K + `lift`), `["signals"]` (per-file prior-defects / churn / owners / degree, targeted mode), `["churn_complexity"]` (volatile-and-complex danger-zone files), `["biomarkers"]` (all findings), and a dimension name `["performance"]` / `["defect"]` / `["maintainability"]` to filter findings to one pillar. |
+| `get_dead_code(...)` | Tiered unreachable / unused-export / zombie-package findings. Run before a cleanup sprint, not before a targeted fix. |
+| `get_overview(repo?)` | Architecture map + `tool_guide` recipes. One-time orientation; skip on subsequent calls in the same session. |
+
+**Trust protocol — when a response replaces reading the source:**
+- `verified: true` on any response means the served content was checked against the live working tree. **Never follow a verified response with a Read of the same lines** — you would be paying twice for identical bytes.
+- `get_answer` with `confidence: "high"` or `grounding: "extracted"` is content-grounded (asserted values were verified against retrieved source; ≥1 citation is source-backed). Cite it directly. `quotes` entries `{path, lines, quote}` are verbatim live source — quote them instead of re-reading. `symbol_bodies` entries `{path, name, lines, source}` are the full live body of a named definition — read that body, do not call `get_symbol` for the same symbol (a `continuation` token names the next range if it was truncated).
+- `code_rationale` entries `{path, lines, comment}` (on a low-confidence `get_answer` or a fallen-back `get_why`) are rationale comments read live from the source. When the "why" you asked for is a code comment rather than an ADR, the comment is right there: cite it instead of opening the file to hunt for it.
+- Reading code: `get_context` skeleton first (~37% of a full Read), then `get_symbol` for bodies, `"path.py:a-b"` range reads for anything between symbols. Raw `Read` is for files the index marks `mostly_full` or cannot serve. On a **large file** `get_symbol("path::Name")` is the win — it serves one verified body instead of pulling the whole file into context.
+- The **only** re-read triggers: `bounds: "approximate"`, `_meta.stale_warning`, `search_method: "bm25"`, or `confidence: "low"`/`retrieval_quality: "weak"`.
+- Disallowed rationalizations for extra reads: "just to be safe", "to double-check the tool", "to see the full context" (use the skeleton / a range read), "the file might have changed" (that is what `verified` already checked).
+- When a plain `Grep` is the right tool, use it: an exhaustive literal-token sweep (rename every call site, find all occurrences) is one grep and unbeatable on cost. MCP's edge there is not fewer tokens — it is the `callers_total`/`callers_truncated` honesty signal that tells you whether a list is complete; reach for `get_context(include=["callers"])` when silent incompleteness would bite.
+
+**Composition tips:**
+- `get_answer` → if `confidence` is `medium`/`low`, follow `best_guesses[0].file` or `fallback_targets[0]` into `get_context`, then `get_symbol` for bytes.
+- `get_context` returns `decision_records` titles → `get_why(targets=[...])` for the rationale; `hotspot: true` → `get_risk` before editing.
+- PR review → `get_risk(targets=[...], changed_files=[...])`; read the `directive` block first.
+- A `tombstone` error means the file was deleted/renamed since indexing — follow `successor_paths`.
+
+### Output Distillation
+
+- Prefer `repowise distill <cmd>` for noisy commands — test runs, builds, `git status`/`log`/`diff`, searches, file listings. It runs the command unchanged (exit code preserved) and prints a compact, errors-first rendering; every error line survives.
+- Output may contain a marker like `[repowise#a1b2c3d4e5f6: 230 lines omitted (~6.1k tokens); restore: repowise expand a1b2c3d4e5f6]`. The omitted content is fully preserved — run `repowise expand <ref>` to retrieve it, or `repowise expand <ref> -q <regex>` for just the matching lines.
+- Never re-run a command to see omitted output; expand the marker instead.
+- For structure-level questions about a large indexed file ("what's in here", "which function handles X"), `get_context(["path"], include=["skeleton"])` returns the file with bodies elided — every signature plus the bodies of the most central symbols — at a fraction of the cost of a full Read.
+
+### Codebase Conventions
+**Commands:**
+- Build: `npm run build`
+- Test: `npm run test`
+- Lint: `npm run lint`
+- Dev: `npm run dev`
+- Format: `npm run format`
+- Typecheck: `npm run typecheck`
+
+<!-- REPOWISE:END -->

.claude/launch.json

@@ -0,0 +1,17 @@
+{
+  "version": "0.0.1",
+  "configurations": [
+    {
+      "name": "frontend-dev",
+      "runtimeExecutable": "pnpm",
+      "runtimeArgs": ["-C", "frontend", "exec", "next", "dev", "-p", "8030"],
+      "port": 8030
+    },
+    {
+      "name": "backend-stack",
+      "runtimeExecutable": "docker",
+      "runtimeArgs": ["compose", "up", "api", "agent-1"],
+      "port": 8031
+    }
+  ]
+}

.omc/state/sessions/b4f7d02c-1da1-4b4d-9835-33e44401355e/hud-cache.json

@@ -0,0 +1 @@
+{"session_id":"b4f7d02c-1da1-4b4d-9835-33e44401355e","context_samples":[{"ts_ms":1783878055117,"tokens":109500},{"ts_ms":1783878057926,"tokens":109939},{"ts_ms":1783878059350,"tokens":109939},{"ts_ms":1783878062169,"tokens":109939},{"ts_ms":1783878065137,"tokens":112312},{"ts_ms":1783878066527,"tokens":112312},{"ts_ms":1783878069643,"tokens":112437},{"ts_ms":1783878070036,"tokens":112437},{"ts_ms":1783878070541,"tokens":112437},{"ts_ms":1783878071081,"tokens":112437},{"ts_ms":1783878077649,"tokens":113096},{"ts_ms":1783878078111,"tokens":113096},{"ts_ms":1783878083683,"tokens":113513},{"ts_ms":1783878084794,"tokens":113513},{"ts_ms":1783878091897,"tokens":114044},{"ts_ms":1783878093626,"tokens":114044},{"ts_ms":1783878095116,"tokens":114044},{"ts_ms":1783878100551,"tokens":114749},{"ts_ms":1783878117149,"tokens":114749},{"ts_ms":1783878475480,"tokens":116909},{"ts_ms":1783878477603,"tokens":116909},{"ts_ms":1783878481720,"tokens":118189},{"ts_ms":1783878483660,"tokens":118189},{"ts_ms":1783878488025,"tokens":118636},{"ts_ms":1783878488449,"tokens":118636},{"ts_ms":1783878495026,"tokens":119031},{"ts_ms":1783878514349,"tokens":119031},{"ts_ms":1783878521236,"tokens":121158},{"ts_ms":1783878586833,"tokens":121464},{"ts_ms":1783878587475,"tokens":121464},{"ts_ms":1783878597857,"tokens":121836},{"ts_ms":1783878604356,"tokens":122080},{"ts_ms":1783878604705,"tokens":122080},{"ts_ms":1783878614717,"tokens":122341},{"ts_ms":1783878621949,"tokens":122341},{"ts_ms":1783878630484,"tokens":123658}],"last_updated_ms":1783878630484,"session_start_tokens":72839}

.omc/state/sessions/dc752497-c4fd-4221-ae35-e78f40e96767/hud-cache.json

@@ -0,0 +1 @@
+{"session_id":"dc752497-c4fd-4221-ae35-e78f40e96767","context_samples":[{"ts_ms":1783877236367,"tokens":0},{"ts_ms":1784175633941,"tokens":0}],"last_updated_ms":1784175633941,"session_start_tokens":null}

.omc/state/sessions/de0b7a77-e9c7-46ac-961a-f0e90ffa1f07/hud-cache.json

@@ -0,0 +1 @@
+{"session_id":"de0b7a77-e9c7-46ac-961a-f0e90ffa1f07","context_samples":[{"ts_ms":1783867603857,"tokens":0},{"ts_ms":1783867659239,"tokens":0}],"last_updated_ms":1783867659239,"session_start_tokens":null}

backend/channels/cli_channel.py

@@ -2,15 +2,19 @@
 
 import asyncio
 import json
+import logging
 import os
 import re
 import shlex
 import shutil
 from typing import Any
 
 from backend.channels.base import AbstractChannel, ChannelResult
+from backend.channels.opencli_channel import _kill_subprocess, _process_group_kwargs
 from backend.channels.registry import register_channel
 
+logger = logging.getLogger(__name__)
+
 _TEMPLATE_RE = re.compile(r"\{\{(\w+)\}\}")
 
 
@@ -73,16 +77,24 @@ async def collect(
                 stdout=asyncio.subprocess.PIPE,
                 stderr=asyncio.subprocess.PIPE,
                 env=env,
+                **_process_group_kwargs(),
             )
             stdout, stderr = await asyncio.wait_for(proc.communicate(), timeout=timeout)
         except asyncio.TimeoutError as exc:
             # Don't orphan the child: wait_for only cancels communicate();
-            # the subprocess itself keeps running until explicitly killed.
-            proc.kill()
+            # the subprocess itself keeps running until explicitly killed. A
+            # bare proc.kill() only reaches the direct child — shell-wrapped
+            # or forked grandchildren survive it — so kill the whole process
+            # tree instead (same helper opencli_channel uses: taskkill /T /F
+            # on Windows, killpg on POSIX; the process group is set up above).
             try:
-                await proc.wait()
+                await _kill_subprocess(proc)
             except Exception:
-                pass
+                logger.warning(
+                    "failed to reap CLI subprocess (pid=%s) after timeout kill",
+                    getattr(proc, "pid", "?"),
+                    exc_info=True,
+                )
             return ChannelResult.fail(
                 f"CLI command timed out after {timeout}s", error_type=type(exc).__name__
             )