@blxzer/cursor-trellis 0.1.2 → 0.2.1

package/dist/templates/trellis/scripts/cursor_retrieval_probe.py

@@ -0,0 +1,396 @@
+#!/usr/bin/env python3
+"""Cursor retrieval capability probe (Phase 1).
+
+Auto-detects the current Cursor environment (Native vs Cursor++ BYOK) by reading
+~/.ccursor/routes.json (byokMode), then runs the subset of retrieval-capability
+probes that are CLI-reachable with deterministic known answers:
+
+  P-01  Grep (rg)            — local IDE literal search
+  P-02  Read                 — local IDE file read
+  P-05  codegraph index      — on-disk .codegraph/ presence
+  P-07  smart-search CLI     — doctor / status reachability
+  D-01  Experiment D         — cursorEnv + BYOK fast-context install readiness
+
+Probes P-03/04/06/08/09/10/D-01(manual inventory) require a live agent session and are documented in
+cursor_retrieval_probe_prompt.md; their results are filled into
+retrieval_probe_matrix_template.json manually.
+
+This script does NOT depend on Cursor being open and does NOT modify any state.
+"""
+
+from __future__ import annotations
+
+import argparse
+import io
+import json
+import os
+import re
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+
+_SCRIPTS = Path(__file__).resolve().parent
+if str(_SCRIPTS) not in sys.path:
+    sys.path.insert(0, str(_SCRIPTS))
+
+from common.cursor_retrieval_env import (  # noqa: E402
+    ENV_BYOK,
+    ENV_NATIVE,
+    ENV_UNKNOWN,
+    detect_cursor_retrieval_env_info,
+)
+
+if sys.platform.startswith("win"):
+    for _stream_name in ("stdin", "stdout", "stderr"):
+        _stream = getattr(sys, _stream_name, None)
+        if _stream is None:
+            continue
+        if hasattr(_stream, "reconfigure"):
+            try:
+                _stream.reconfigure(encoding="utf-8", errors="replace")
+            except Exception:
+                pass
+        elif hasattr(_stream, "detach"):
+            try:
+                setattr(
+                    sys,
+                    _stream_name,
+                    io.TextIOWrapper(_stream.detach(), encoding="utf-8", errors="replace"),
+                )
+            except Exception:
+                pass
+
+PROBE_VERSION = 1
+
+
+def _utc_now() -> str:
+    return datetime.now(timezone.utc).isoformat().replace("+00:00", "Z")
+
+
+def detect_environment() -> dict[str, object]:
+    """Alias for router/probe: same as detect_cursor_retrieval_env_info()."""
+    return detect_cursor_retrieval_env_info()
+
+
+def _probe_grep(repo_root: Path) -> dict[str, object]:
+    """P-01: Grep a known literal in a known file under the repo.
+
+    Known answer: 'TRELLIS_B2_WPeLc8' must appear in
+    .trellis/local/cursor2plus/patch_wpelc8.py (the MARKER constant).
+    """
+    marker = "TRELLIS_B2_WPeLc8"
+    target = repo_root / ".trellis" / "local" / "cursor2plus" / "patch_wpelc8.py"
+    ok_file = target.is_file()
+    found = False
+    matches: list[str] = []
+    if ok_file:
+        try:
+            text = target.read_text(encoding="utf-8", errors="ignore")
+            for i, line in enumerate(text.splitlines(), 1):
+                if marker in line:
+                    matches.append(f"{i}:{line.strip()[:120]}")
+            found = bool(matches)
+        except OSError:
+            pass
+    return {
+        "id": "P-01",
+        "label": "Grep (rg) — local IDE literal search",
+        "status": "pass" if found else "fail",
+        "expected": marker,
+        "target_file": str(target),
+        "matches": matches[:5],
+        "note": "Local IDE operation; not routed through BYOK proxy." if found else (
+            "Target file missing or marker not found." if ok_file else "Target file missing."
+        ),
+    }
+
+
+def _probe_read(repo_root: Path) -> dict[str, object]:
+    """P-02: Read a known file and verify a known line content.
+
+    Known answer: .cursor/mcp.json must contain a 'codegraph' server entry.
+    """
+    target = repo_root / ".cursor" / "mcp.json"
+    ok_file = target.is_file()
+    has_codegraph = False
+    has_fast_context = False
+    servers: list[str] = []
+    if ok_file:
+        try:
+            data = json.loads(target.read_text(encoding="utf-8"))
+            mcp = data.get("mcpServers", {})
+            if isinstance(mcp, dict):
+                servers = sorted(mcp.keys())
+                has_codegraph = "codegraph" in mcp
+                has_fast_context = "fast-context" in mcp
+        except (OSError, json.JSONDecodeError):
+            pass
+    ok = has_codegraph
+    return {
+        "id": "P-02",
+        "label": "Read — local IDE file read + JSON parse",
+        "status": "pass" if ok else "fail",
+        "expected": "mcpServers.codegraph present",
+        "target_file": str(target),
+        "configured_servers": servers,
+        "has_codegraph": has_codegraph,
+        "has_fast_context": has_fast_context,
+        "note": "Local IDE operation; verifies MCP config on disk." if ok else "mcp.json missing or codegraph not configured.",
+    }
+
+
+def _probe_codegraph_index(repo_root: Path) -> dict[str, object]:
+    """P-05: codegraph index on disk (.codegraph/ directories).
+
+    Mirrors codegraph_session_smoke.py logic.
+    """
+    found: list[Path] = []
+    direct = repo_root / ".codegraph"
+    if direct.is_dir():
+        found.append(direct.resolve())
+    for child in sorted(repo_root.iterdir()):
+        if not child.is_dir():
+            continue
+        nested = child / ".codegraph"
+        if nested.is_dir() and nested.resolve() not in found:
+            found.append(nested.resolve())
+    ok = len(found) > 0
+    return {
+        "id": "P-05",
+        "label": "codegraph index — on-disk .codegraph/ presence",
+        "status": "pass" if ok else "fail",
+        "expected": "at least one .codegraph/ dir under workspace root or top-level subproject",
+        "index_paths": [str(p) for p in found],
+        "note": "Local disk; MCP must also be enabled in Cursor for live calls." if ok else "Run codegraph init to index.",
+    }
+
+
+def _probe_smart_search(repo_root: Path) -> dict[str, object]:
+    """P-07: smart-search CLI reachability (resolve, do not execute search).
+
+    Reuses common.smart_search_resolve to find the executable; reports resolved
+    argv without running a network search. A separate `doctor` invocation is
+    attempted if the resolved entrypoint is the smart-search package CLI.
+    """
+    scripts_dir = repo_root / ".trellis" / "scripts"
+    if str(scripts_dir) not in sys.path:
+        sys.path.insert(0, str(scripts_dir))
+    resolved: list[str] | None = None
+    resolve_error = ""
+    try:
+        from common.smart_search_resolve import resolve_smart_search_argv  # type: ignore[import-not-found]
+
+        resolved = resolve_smart_search_argv(repo_root)
+    except Exception as exc:
+        resolve_error = f"{type(exc).__name__}: {exc}"
+
+    doctor_status = "skipped"
+    doctor_output = ""
+    if resolved:
+        import subprocess
+
+        entry = resolved[-1] if resolved else ""
+        try:
+            proc = subprocess.run(
+                resolved + (["doctor"] if entry.endswith((".js", ".cmd", ".bat")) or "smart-search" in str(resolved) else []),
+                capture_output=True,
+                text=True,
+                timeout=12,
+                cwd=str(repo_root),
+                encoding="utf-8",
+                errors="replace",
+            )
+            doctor_output = (proc.stdout or "") + (proc.stderr or "")
+            doctor_status = "ok" if proc.returncode == 0 else f"exit={proc.returncode}"
+        except (subprocess.SubprocessError, OSError) as exc:
+            doctor_status = f"error: {type(exc).__name__}"
+            doctor_output = str(exc)
+
+    ok = bool(resolved)
+    return {
+        "id": "P-07",
+        "label": "smart-search CLI — resolve + doctor",
+        "status": "pass" if ok else "fail",
+        "expected": "smart-search argv resolved",
+        "resolved_argv": resolved,
+        "resolve_error": resolve_error,
+        "doctor_status": doctor_status,
+        "doctor_output_excerpt": doctor_output[:400],
+        "note": "Local CLI; not routed through BYOK proxy." if ok else "smart-search not resolvable.",
+    }
+
+
+def _probe_d01_semanticsearch_drop(
+    env_info: dict[str, object],
+    mcp_config: dict[str, object],
+) -> dict[str, object]:
+    """D-01 (auto): Experiment D readiness — BYOK needs fast-context in mcp.json.
+
+    Does not prove SemanticSearch absence (that is manual D-01 + P-08). Pass when:
+    - Native: env is native (built-in semantic path expected in manual probes).
+    - BYOK: fast-context configured in .cursor/mcp.json.
+    """
+    env = str(env_info.get("env") or ENV_UNKNOWN)
+    byok_mode = env_info.get("byokMode")
+    fc = bool(mcp_config.get("fast_context_configured"))
+    cg = bool(mcp_config.get("codegraph_configured"))
+    if env == ENV_BYOK:
+        ok = fc and cg
+        status = "pass" if ok else "fail"
+        note = (
+            "BYOK: fast-context + codegraph in mcp.json — ready for P-08-SA / REC-11."
+            if ok
+            else "BYOK: missing fast-context and/or codegraph in mcp.json; "
+            "select codebase-retrieval at init or add servers manually."
+        )
+    elif env == ENV_NATIVE:
+        ok = True
+        status = "pass"
+        note = (
+            "Native: use manual D-01 / P-08 for built-in SemanticSearch inventory; "
+            f"fast-context configured={fc} (optional on Native)."
+        )
+    else:
+        ok = fc
+        status = "degraded" if fc else "fail"
+        note = (
+            "cursorEnv unknown; fast-context in mcp.json supports BYOK concept recall if env is byok."
+            if fc
+            else "cursorEnv unknown and fast-context not in mcp.json."
+        )
+    return {
+        "id": "D-01",
+        "label": "Experiment D — SemanticSearch drop readiness (env + MCP)",
+        "status": status,
+        "expected": "BYOK → fast-context+codegraph in mcp.json; Native → env native",
+        "cursor_env": env,
+        "byokMode": byok_mode,
+        "fast_context_configured": fc,
+        "codegraph_configured": cg,
+        "manual_followup": "cursor_retrieval_probe_prompt.md PROBE D-01 (tool inventory)",
+        "related_probes": ["P-08", "P-08-SA"],
+        "note": note,
+    }
+
+
+def _probe_mcp_config(repo_root: Path) -> dict[str, object]:
+    """Supplemental: read .cursor/mcp.json to report configured MCP servers.
+
+    Tells the user which MCP-based probes (P-03/04/06) should be runnable.
+    """
+    target = repo_root / ".cursor" / "mcp.json"
+    servers: dict[str, object] = {}
+    if target.is_file():
+        try:
+            data = json.loads(target.read_text(encoding="utf-8"))
+            mcp = data.get("mcpServers", {})
+            if isinstance(mcp, dict):
+                servers = mcp
+        except (OSError, json.JSONDecodeError):
+            pass
+    return {
+        "codegraph_configured": "codegraph" in servers,
+        "fast_context_configured": "fast-context" in servers,
+        "all_servers": sorted(servers.keys()),
+        "mcp_json_path": str(target),
+    }
+
+
+def run_auto_probes(repo_root: Path) -> dict[str, object]:
+    env_info = detect_environment()
+    mcp_config = _probe_mcp_config(repo_root)
+    return {
+        "probe_version": PROBE_VERSION,
+        "env": env_info,
+        "probed_at": _utc_now(),
+        "repo_root": str(repo_root),
+        "auto_results": [
+            _probe_grep(repo_root),
+            _probe_read(repo_root),
+            _probe_codegraph_index(repo_root),
+            _probe_smart_search(repo_root),
+            _probe_d01_semanticsearch_drop(env_info, mcp_config),
+        ],
+        "mcp_config": mcp_config,
+        "manual_probe_ids": ["P-03", "P-04", "P-06", "D-01", "P-08", "P-09", "P-10"],
+        "manual_probe_doc": "cursor_retrieval_probe_prompt.md",
+        "matrix_template": "retrieval_probe_matrix_template.json",
+    }
+
+
+def _format_human(report: dict[str, object]) -> str:
+    env_info = report.get("env", {})
+    env = env_info.get("env", "unknown")
+    lines = [
+        f"Cursor retrieval probe (Phase 1)",
+        f"  environment : {env}  (source: {env_info.get('source', '?')}, byokMode={env_info.get('byokMode')})",
+        f"  repo_root   : {report.get('repo_root')}",
+        f"  probed_at   : {report.get('probed_at')}",
+        "",
+        "Auto probes (CLI-reachable):",
+    ]
+    for r in report.get("auto_results", []):
+        status = r.get("status", "?")
+        tag = "PASS" if status == "pass" else "FAIL"
+        lines.append(f"  [{tag}] {r.get('id')} {r.get('label')}")
+        if r.get("note"):
+            lines.append(f"        {r['note']}")
+    mcp = report.get("mcp_config", {})
+    lines.append("")
+    lines.append("MCP config (for manual probes P-03/04/06 and D-01 BYOK readiness):")
+    lines.append(f"  codegraph     : {'configured' if mcp.get('codegraph_configured') else 'MISSING'}")
+    lines.append(f"  fast-context  : {'configured' if mcp.get('fast_context_configured') else 'MISSING'}")
+    lines.append(f"  all servers   : {', '.join(mcp.get('all_servers', [])) or '(none)'}")
+    lines.append("")
+    lines.append("Manual probes (run via cursor_retrieval_probe_prompt.md):")
+    for pid in report.get("manual_probe_ids", []):
+        lines.append(f"  {pid}")
+    lines.append("")
+    redirect = env_info.get("redirect_endpoints", [])
+    if env == "byok" and isinstance(redirect, list) and redirect:
+        lines.append(f"BYOK redirect endpoints ({len(redirect)}):")
+        for ep in redirect[:8]:
+            lines.append(f"  - {ep}")
+        if len(redirect) > 8:
+            lines.append(f"  ... +{len(redirect) - 8} more")
+        lines.append("")
+    return "\n".join(lines)
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="Cursor retrieval capability probe (Phase 1)")
+    parser.add_argument("--root", type=Path, default=None, help="Workspace root (default: Trellis repo root)")
+    parser.add_argument("--json", action="store_true", help="Machine-readable JSON report")
+    parser.add_argument("--out", type=Path, default=None, help="Write report JSON to this path")
+    args = parser.parse_args()
+
+    if args.root is not None:
+        repo_root = args.root.resolve()
+    else:
+        try:
+            from common.paths import get_repo_root  # type: ignore[import-not-found]
+            repo_root = get_repo_root()
+        except Exception:
+            repo_root = Path.cwd().resolve()
+
+    report = run_auto_probes(repo_root)
+
+    if args.out:
+        args.out.parent.mkdir(parents=True, exist_ok=True)
+        args.out.write_text(
+            json.dumps(report, indent=2, ensure_ascii=False) + "\n",
+            encoding="utf-8",
+        )
+        print(f"wrote {args.out}", file=sys.stderr)
+
+    if args.json:
+        print(json.dumps(report, indent=2, ensure_ascii=False))
+    else:
+        print(_format_human(report))
+
+    any_fail = any(r.get("status") != "pass" for r in report.get("auto_results", []))
+    return 1 if any_fail and args.json else 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

package/dist/templates/trellis/scripts/cursor_retrieval_probe_prompt.md

@@ -0,0 +1,300 @@
+# Cursor retrieval capability probe — manual prompts (Phase 1)
+
+> **Purpose:** Measure retrieval capabilities that require a live agent session
+> (MCP tools, built-in semantic search, LSP). Run each probe in **both**
+> environments and record results in `retrieval_probe_matrix_template.json`.
+>
+> **Environment A (Native):** ccursor proxy OFF (or `byokMode=0`).
+> **Environment B (BYOK):** ccursor proxy ON, `byokMode=1`.
+
+## How to use this file
+
+1. Switch to the target environment (Native or BYOK) in Cursor.
+2. Copy the **probe block** (from `--- PROBE` to the next `---`) into a fresh
+   Cursor Agent chat. Use the same block in both environments.
+3. The agent answers with a structured reply; copy its answer into the matching
+   `native`/`byok` slot of the matrix JSON.
+4. Do **not** coach the agent. It must use whichever tools it actually has.
+5. Record the **actual tool name** the agent called (from its tool-use output),
+   not the name we asked for — this is the BYOK detection signal.
+
+## Deterministic known answers
+
+Every probe has a verifiable answer derived from the `D:\MyHarness` workspace.
+The agent either returns the right answer (pass) or does not (fail/degraded).
+
+---
+
+--- PROBE P-03: codegraph MCP — codegraph_search (symbol lookup)
+
+You are a retrieval capability probe. Do NOT read any files directly. Use ONLY
+the **codegraph_search** MCP tool to answer.
+
+Task: Call `codegraph_search` with the query `patch_wpelc8` and report:
+1. The file path(s) returned (expected to contain `patch_wpelc8.py` under
+   `.trellis/local/cursor2plus/`).
+2. The symbol kind (expected: `function`).
+3. The exact tool name you invoked (from your tool-use record).
+
+Reply in this exact format:
+
+```
+P-03 RESULT
+tool_invoked: <exact tool name from your tool-use record>
+file_path: <path returned>
+symbol_kind: <kind returned>
+verdict: pass  # pass if patch_wpelc8.py appears, else fail
+notes: <anything notable>
+```
+
+---
+
+--- PROBE P-04: codegraph MCP — codegraph_callers (caller chain)
+
+You are a retrieval capability probe. Do NOT read any files directly. Use ONLY
+the **codegraph_callers** MCP tool to answer.
+
+Task: Call `codegraph_callers` with the symbol `route_codebase_retrieval` and
+report:
+1. The list of caller sites returned (expected: at least one caller in
+   `route_codebase_retrieval.py` under `.trellis/scripts/`).
+2. The count of callers.
+3. The exact tool name you invoked.
+
+Reply in this exact format:
+
+```
+P-04 RESULT
+tool_invoked: <exact tool name from your tool-use record>
+caller_count: <number>
+caller_sites: <semicolon-separated file:line list, or "none">
+verdict: pass  # pass if at least 1 caller found, else fail
+notes: <anything notable>
+```
+
+---
+
+--- PROBE P-06: fast-context MCP — semantic search
+
+You are a retrieval capability probe. Do NOT read any files directly. Use ONLY
+the **fast-context** MCP tool (fast_context_search) to answer.
+
+Task: Call the fast-context semantic search with the query:
+`Cursor++ BYOK subagent model routing patch`
+
+Report:
+1. Whether the tool returned any results.
+2. If so, the top file path (expected to contain something under
+   `.trellis/local/cursor2plus/` or `.trellis/spec/guides/cursor-subagent-policy.md`).
+3. The exact tool name you invoked.
+
+Reply in this exact format:
+
+```
+P-06 RESULT
+tool_invoked: <exact tool name from your tool-use record>
+result_count: <number>
+top_file: <path or "none">
+verdict: pass  # pass if tool executed and returned results, else fail
+notes: <anything notable>
+```
+
+---
+
+--- PROBE D-01: Experiment D — codebase semantic tool inventory (manual)
+
+**Experiment D** documents why **Cursor++ BYOK** often lacks built-in **SemanticSearch**
+(P-08 `tool_invoked: none`) and why **fast_context_search** is the designed Primary.
+
+Before running P-08, inventory what **this Agent session** can use for **codebase-wide
+concept recall** (not Grep on a known string). Do **not** invoke tools yet — list from
+your **available tool table / MCP list** only.
+
+Also run locally (or ask the user to confirm):
+`python .\.trellis\scripts\cursor_retrieval_probe.py --json` and note `env.env`,
+`mcp_config.fast_context_configured`, and auto `D-01` status.
+
+Reply in this exact format:
+
+```
+D-01 RESULT
+cursor_env: native|byok|unknown  # from probe JSON env.env or your knowledge of ccursor/BYOK
+builtin_semantic_tools: <comma-separated names, or "none">  # e.g. SemanticSearch, @codebase
+fast_context_mcp_listed: yes|no
+fast_context_search_listed: yes|no
+auto_d01_status: pass|fail|degraded|not_run  # from probe auto_results D-01 if you ran it
+verdict: pass|fail|degraded
+notes: <BYOK expects builtin none + fast-context; Native expects SemanticSearch or equivalent>
+```
+
+verdict rules:
+- **Native:** pass if `builtin_semantic_tools` includes SemanticSearch or equivalent built-in codebase semantic.
+- **BYOK:** pass if `builtin_semantic_tools` is `none` (or empty) **and** `fast_context_search_listed` is yes **and** auto D-01 is pass when mcp.json has fast-context+codegraph.
+- **degraded:** BYOK with fast-context MCP missing from config/list, or Native without built-in semantic but fast-context present.
+- **fail:** cannot determine tool table, or BYOK with neither builtin nor fast-context path.
+
+Pair with **P-08** (actual concept question) and **P-08-SA** (fast-context only).
+
+---
+
+--- PROBE P-08: @codebase semantic search (Cursor built-in)
+
+You are a retrieval capability probe. Do NOT use Grep, Read, or any MCP tool.
+Use ONLY Cursor's **built-in codebase / semantic search** (the `@codebase`
+capability or any built-in semantic search tool the host exposes) to answer.
+
+The question below is deliberately written with NO literal file names, symbol
+names, or paths — it can only be answered by semantic understanding of the
+codebase, not by literal string matching.
+
+Semantic question:
+> "Where does the system decide which model a background sub-agent will run on
+> when the user is routing their own API keys through a local proxy instead of
+> the official cloud?"
+
+Expected answer (for verification, do NOT reveal this to the solver path):
+the relevant logic lives in `.trellis/local/cursor2plus/patch_wpelc8.py` (the
+`WPeLc8` function patch) and is documented in
+`.trellis/spec/guides/cursor-subagent-policy.md` and
+`.trellis/spec/guides/cursor-subagent-reverse-engineering-report.md`.
+
+Reply in this exact format:
+
+```
+P-08 RESULT
+tool_invoked: <exact tool name from your tool-use record — e.g. @codebase, codebase_search, SemanticSearch, or "none">
+answer: <your answer to the semantic question>
+matched_expected: yes|no  # yes if answer references patch_wpelc8.py or cursor-subagent-policy.md
+verdict: pass|fail|degraded
+notes: <did the tool execute at all? any error? did you fall back to another tool?>
+```
+
+verdict rules:
+- pass: built-in semantic tool executed AND answer references the expected files.
+- degraded: tool executed but answer wrong, OR you had to fall back to Grep/codegraph.
+- fail: built-in semantic tool did not execute / not available / errored out.
+
+---
+
+--- PROBE P-08-SA: Concept probe — fast-context only (BYOK Primary path)
+
+You are a retrieval capability probe. Answer the semantic question below using
+**only** the fast-context MCP tool **`fast_context_search`** (CallMcpTool on the
+fast-context server). Do NOT use Grep, Read, SemanticSearch, @codebase,
+codegraph_*, Task subagent, or WebSearch.
+
+Use a natural-language query in English (add Chinese business terms only if helpful).
+Set `project_path` to the workspace root (e.g. D:\MyHarness).
+
+Semantic question (same as P-08 — no literal file/symbol/path clues in your query):
+> "Where does the system decide which model a background sub-agent will run on
+> when the user is routing their own API keys through a local proxy instead of
+> the official cloud?"
+
+Expected (for verification only — do not paste paths into fast_context_search query):
+Top hits should include `.trellis/local/cursor2plus/patch_wpelc8.py` and/or
+`.trellis/spec/guides/cursor-subagent-policy.md`.
+
+Reply in this exact format:
+
+```
+P-08-SA RESULT
+tool_invoked: fast_context_search|none|other
+top_paths: <comma-separated paths from fast_context_search results, or "none">
+matched_expected: yes|no  # yes if top_paths includes patch_wpelc8.py or cursor-subagent-policy.md
+verdict: pass|fail|degraded
+notes: <errors, empty results, or if you had to use another tool>
+```
+
+verdict rules:
+- pass: fast_context_search executed AND matched_expected yes.
+- degraded: executed but wrong/empty top paths.
+- fail: did not invoke fast_context_search, or tool missing in this session.
+
+**Native:** optional control — pass confirms MCP works; Primary for concept recall
+remains built-in SemanticSearch (P-08). **BYOK:** required — this is the designed
+Primary when P-08 reports `tool_invoked: none`. Pair with D-01 inventory and P-08.
+
+---
+
+--- PROBE P-09: DEEP_SEARCH / wide cross-cutting semantic explore
+
+You are a retrieval capability probe. Use Cursor's **DEEP_SEARCH** capability
+(or the equivalent wide cross-cutting semantic explore tool the host exposes)
+to answer. Do NOT use Grep or Read until after you have attempted DEEP_SEARCH.
+
+Semantic question (no literal clues):
+> "How does the workspace stitch together per-query retrieval plans, agent
+> execution policy, and per-turn hook injection across Cursor, given that one
+> of the injection channels is known-broken?"
+
+Expected (for verification only): the answer should weave together
+`.cursor/hooks/inject-retrieval-plan.py` (beforeSubmitPrompt injection),
+`.cursor/rules/retrieval-routing.mdc` (alwaysApply policy), and the
+`sessionStart` bug #158452 noted in `.trellis/spec/guides/cursor-context-injection-guide.md`.
+
+Reply in this exact format:
+
+```
+P-09 RESULT
+tool_invoked: <exact tool name — DEEP_SEARCH, Explore subagent, or "none">
+answer: <your answer>
+matched_expected: yes|no  # yes if answer references at least 2 of: inject-retrieval-plan.py, retrieval-routing.mdc, sessionStart bug #158452
+verdict: pass|fail|degraded
+notes: <did DEEP_SEARCH run? did you fall back to subagent explore / Grep?>
+```
+
+verdict rules:
+- pass: DEEP_SEARCH (or equivalent) executed AND answer references >=2 expected anchors.
+- degraded: fell back to Explore subagent / codegraph / Grep, or answer partial.
+- fail: no wide explore tool executed.
+
+---
+
+--- PROBE P-10: Definition / reference (codegraph — Agent LSP not exposed)
+
+Cursor Agent does **not** expose GO_TO_DEFINITION in the tool table (Native/BYOK).
+Trellis routes definition/reference to **codegraph**. This probe measures the
+**product path**, not raw LSP.
+
+You are a retrieval capability probe. Use **codegraph_node** or **codegraph_search**
+as the **primary** path. Do NOT use Grep as the first step. Use **Read** only to
+confirm line numbers after codegraph returns a definition.
+
+Task: Locate the **definition** of `route_codebase_retrieval` (Python function).
+
+Expected: definition in `.trellis/scripts/common/codebase_retrieval_router.py`
+(`def route_codebase_retrieval(`).
+
+Reply in this exact format:
+
+```
+P-10 RESULT
+tool_invoked: <exact tool name — codegraph_node, codegraph_search, or "none">
+definition_file: <path>
+definition_line: <line number or "unknown">
+verdict: pass|fail|environment_limitation
+notes: <if you tried GO_TO_DEFINITION and it was unavailable, say so; pass when codegraph + Read confirm def>
+```
+
+verdict rules:
+- **pass**: codegraph (or equivalent MCP) returned the correct `def route_codebase_retrieval` file.
+- **environment_limitation**: only when codegraph MCP/index is missing; not when LSP is missing.
+- **fail**: wrong file or no structural tool executed.
+
+---
+
+## After running all probes
+
+For each probe, fill the matching `native` and `byok` objects in
+`retrieval_probe_matrix_template.json`:
+
+```json
+"native": { "status": "pass", "evidence": "<answer excerpt>", "actual_tool": "<tool_invoked>", "probed_at": "<ISO8601>" }
+```
+
+Then set `verdict.summary`, `verdict.degraded_capabilities` (list of probe IDs
+where byok differs from native), and `verdict.probed_at`.
+
+Degraded capabilities drive Phase 2 code adaptations — only those that actually
+differ between environments get changed.