菜单
Use when designing or evaluating tensor-grep as an enterprise agentic code-intelligence tool — tg prepare one-call edit readiness, PATH narrowing, agent --deadline, find/route-test/ledger, install-dense, EvidenceReceipts, review-bundle, GPU honesty, world-class readiness gaps.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/oimiragieo/tensor-grep --skill tensor-grep-enterprise-agent命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
Use when searching code, logs, or repositories with tensor-grep; validating rg or AST parity; using tg MCP tools; checking GPU/search routing; or producing agent-friendly context, source, refs, or blast-radius output.
Use when writing or editing any tensor-grep doc of record (AGENTS.md, CLAUDE.md, root SKILL.md, docs/SESSION_HANDOFF.md, docs/CONTRACTS.md, docs/CONTINUATION_PLAN.md, docs/PAPER.md, README.md, docs/benchmarks.md, docs/gpu_crossover.md, mkdocs.yml, or any .claude/skills/*/SKILL.md) — before adding a capability claim, syncing a release-line note, touching a version-stamped line, adding a new governed doc, or a docs-only pytest fails and you don't know which fragment broke it. Covers which doc owns which contract, the two governance layers (semantic-release version_variables + scripts/stamp_release_assets.py auto-stamping vs pytest content-pinning), the cross-doc fragment discipline, mkdocs --strict, and house style.
Use when deciding what counts as proof that a tensor-grep (tg) change works — before trusting a subagent's "tests pass", writing a new test, claiming a routing/docs/release fix is done, shipping a doc-drift/ranking/classification heuristic off green fixture tests, proving a red-green baseline on a reverted/pre-fix commit, reviewing a payload-byte/ratio governance test, de-flaking a timing-sensitive test, or running the pre-push gate. Covers TDD-first discipline, the CliRunner-vs-real-binary trap, the fixture-green-vs-real-corpus-dogfood trap for precision/heuristic features, the `capfd`-vs-`result.stdout` capture-surface trap on routing/delegation changes (needs `tests/integration/` run with the native `tg` binary rebuilt, not just `tests/unit/`), the `tests/conftest.py` `sys.path.insert`-outranks-`PYTHONPATH` trap that can falsify a red-green baseline even with `tensor_grep.__file__` verified, a shared-envelope field growth breaking a payload-ratio governance test plus its tmp-path-length platform sensitivit
| name | tensor-grep-enterprise-agent |
| description | Use when designing or evaluating tensor-grep as an enterprise agentic code-intelligence tool — tg prepare one-call edit readiness, PATH narrowing, agent --deadline, find/route-test/ledger, install-dense, EvidenceReceipts, review-bundle, GPU honesty, world-class readiness gaps. |
Verified against tg 1.110.14 (2026-07-26 refresh of the gap table only; the WSL workspace+GPU rows still date from the 2026-07-21 v1.91.0 dogfood and are marked as such. Individual gaps are re-verified by source inspection against the shipped line, not a re-run whole-workspace sweep — see the native-scale dogfood bullet below for a fresh large-repo data point).
2026-08-01 anchor-drift pass (local tg was 1.101.24, PyPI 1.101.27 at the time -- neither
re-run, this pass only re-derived cited file:line anchors against the current checkout): four
file:line citations had drifted (docs/BACKLOG.md #578 entry, repo_map.py's "Break + keep what
we have" comment, three codemap.py git-touching call sites, and bootstrap.py's
_run_rg_passthrough forwarding site) — each replaced below with a grep instruction plus its own
was -> now receipt instead of a re-stamped number. docs/CONTRACTS.md:144 was re-checked and is
unchanged. The language-tier split (Symbol-graph language coverage row) was re-verified live and
is unchanged; a THIRD hand-counting failure of that same claim is now recorded there.
Each row states when it was last checked; rows without a date are carried forward unverified. A row re-verified today and one inherited from a five-week-old dogfood are not the same evidence, and a gap table that flattens the difference is exactly the drift that let "8/10 languages, C/C++ deferred" survive three releases past the campaign that shipped them. When you re-verify a row, stamp it.
tg prepare REPO/src "task" --json (replaces orient→agent→route-test→callers→evidence argv guessing). Use --claim when multi-agent coordination is needed; use --out FILE to persist the capsule for tg evidence emit --capsule FILE with no manual save.REPO/src. Whole-repo: tg prepare|agent REPO --deadline N → expect partial / ask_user.tg agent REPO "empty TIMEOUT @75s" claim is DEBUNKED -- do not re-chase it.
docs/BACKLOG.md (grep -n '#578' docs/BACKLOG.md -- no line number is carried here on
purpose: this citation was re-stamped :603 -> :850 on 2026-08-01 and :850 was already
wrong by 2026-08-02. The file grows by insertion, so any stamp rots; run the grep) records #578
correcting this as one of TWO false WSL-/mnt/c "regression" claims; the native repro is ~26s.
The scan is already interruptible and already keeps a usable partial (repo_map.py,
grep -n "Break + keep what we have" src/tensor_grep/cli/repo_map.py -- the stamp chain was
:7400 -> re-stamped :7490 (2026-08-01) -> verified :7737 at this SHA, i.e. the re-stamp
rotted AGAIN, which is exactly why the grep is the citation: "Break + keep what we have -- never
raise, never zero the results"; measured 7/45 files at a
0.05s deadline still yields the correct primary_target).
What actually produces an apparently-empty result is the post-deadline overshoot, which is
deliberate: per docs/CONTRACTS.md:144 (re-grepped 2026-08-01, unchanged -- the partial_reason
"deadline" paragraph is still the correct anchor), --deadline is a stop-starting-new-work
bound, not a wall-clock guarantee -- each builder rechecks the deadline at its own return point
and would rather report a late-but-complete answer as partial. A 60s default plus an
FS-latency-proportional tail lands near 75s, at which point the CALLER's own timeout kills the
process. That is the harness returning nothing, not tg returning an empty payload -- a different
bug with a different fix.tg orient/tg search/tg inventory --deadline all bound gracefully ( ~4.9s via scan_limit+centrality; returns partial plus an honest "exceeded timeout" message, exit 124; bounds per-project). Known low-priority edge: a single non-lazy call in the shared can still blow on a pathological workspace-union tree (rare; not worth a load-bearing fix).ask_user_before_editing.requiredpartial / exit 2search_readyreview-bundle create without --manifestroute-test.agreement == false (when not using prepare's floor)world_class_readiness = not_claimed)| Gap | Status (version + date where re-verified) |
|---|---|
| Whole-repo agent/prepare default deadline reliability | CLOSED 2026-07-27 -- do not re-open from this row. Two successive framings here were both overtaken. The "bare agent TIMEOUT" was a debunked WSL-9p artifact (BACKLOG.md #578 entry, grep -n '#578' docs/BACKLOG.md, was :603 now :850, native ~26s), and the follow-on "unbounded session refresh" framing is ALSO stale: task #304 bounded the staleness-triggered rebuild with the same warm-daemon budget (session_daemon.py carries the Task #304: bound the staleness-triggered rebuild with the SAME budget comment and passes deadline_monotonic=monotonic() + WARM_DAEMON_DEFAULT_DEADLINE_SECONDS). The old citations are deliberately DROPPED rather than re-stamped -- they described a code shape that no longer exists, and re-pointing them would preserve a dead paragraph. If you believe a deadline gap remains, re-derive it from scratch against origin/main; do not carry this row forward |
| CUDA-native GPU promotion | Open (adjudicated HOLD, #169 CEO-gated; kernel is brute-force byte-compare, not PFAC) |
codemap on WSL | Debunked, not open (grep -n '#578' docs/BACKLOG.md, plus the older claim it corrects at grep -n '60-180s/no JSON' docs/BACKLOG.md -- no line stamps, they rotted twice; native 41s whole-repo complete, 33s in the earlier repro). Root cause was three git-touching call sites, fixed in e95abfa @ v1.82.1 |
| Mega-repo auto-narrow + accurate deadline primaries | Partial (suggested_scope/workspace_root_detected shipped, #684; deadline-primary accuracy still open) |
| Unscoped-search fast-refuse on the default flag-less path | Shipped (A9; generic 1500-file ceiling, ~1.7s, all 3 doors) |
| Dynamic-import / blast-radius decoy honesty | Shipped (A10/A15; dynamic_unresolved excluded from forward/reverse resolution and the blast-radius scoring prefilter) |
| One-call prepare CUJ | Shipped (prefer src/; --out FILE persists the capsule) |
| Packaged dense semantic |
tg install-dense --json # once per host
tg prepare REPO/src "task" --out /tmp/prep.json --json
# if ask_user / partial: narrow PATH or raise --deadline; do not edit yet
# optional: tg prepare REPO/src "task" --claim --json
# then edit from primary_target; run validation_commands; optionally:
tg evidence emit REPO --capsule /tmp/prep.json --query "task" --json --agent-id "$AGENT_ID"
Bare uvx --from tensor-grep==<ver> does not include dense embeddings. tg find will warn and
degrade (rank_fallback_reason). For enterprise launch-bar dogfood use prepare → evidence →
review-bundle → ledger (and plain tg search), or install tensor-grep[semantic] /
tg install-dense first. Isolate HOME when probing tg evidence emit --sign without a key
(AGENTS.md A70, A73).
orientsearchinventory --deadlineos.scandir_iter_repo_filesinventory --deadlinetg install-dense once per machine (never auto); then tg find. Every dense-absent hint across the CLI now leads with tg install-dense.tensor-grep-ledger. Claim/release/list now canonicalize to the nearest .git ancestor (worktree-aware); the PATH-mismatch footgun from 1.92.1-era dogfood is fixed.tg codemap WSL timeout is likewise DEBUNKED (docs/BACKLOG.md, same #578 entry as
above: native repro 41s whole-repo, partial=false, complete; and the older claim being
corrected -- grep -n '60-180s/no JSON' docs/BACKLOG.md -- "codemap '60-180s/no JSON' = WSL 9p
(native 33s complete)"). Line stamps deliberately dropped here too, same reason as above.
Its slow path is also a
DIFFERENT mechanism from tg agent's: codemap makes three git-touching call sites
(grep -n '\["git"\|_repo_revision_identity(' src/tensor_grep/cli/codemap.py; was :612,
:1054, :1353, now :568 the git ls-files subprocess, :1062 and :1361 the two
_repo_revision_identity calls) where tg agent makes zero, and that exact cost was already
root-caused and fixed in e95abfa (v1.82.1, confirmed present in history). Treating the two as
"one root cause" under-serves codemap and over-serves agent. GPU inventory ≠ acceleration; the WSL bare-shim cross-domain misclassification that produced a bogus path_not_found is fixed (v1.93.0).Shipped via install-dense (opt-in, once; every dense-absent hint now leads with it) |
| Ledger → CI / review-bundle bridge | Partial — review-bundle --receipt/--against CI gate chain shipped (#681); ledger itself stays advisory, not wired into a CI gate |
| Agent accuracy gate | Shipped (tests/eval/test_agent_accuracy.py, per-task-pinned, 16/16 golden tasks — the loop-4 measurement instrument that surfaced and fixed #250) |
| Symbol-graph language coverage | REGISTRY 10/10; CALLER-GRAPH 10/10 -- 10 parser-backed / 0 foundational: the foundational tier is EMPTY, there is no Y tier (2026-08-09 skills-accuracy audit vs origin/main e3feaf5; quote the descriptor split, never the bare 10/10). All ten register (c, cpp, csharp, go, java, javascript, php, python, rust, typescript; C/C++ via lang_c.py/lang_cpp.py, closed by the Task 10D/10E final waves of the top-10 language-support campaign). This skill's hard-stop list does NOT key on blast_radius_floor/callers_count directly (the ## Hard stops section above names neither -- earlier text claimed it did, and it did not) -- the connection is hard-stop #6 (route-test.agreement == false), whose only exemption is "when using prepare's floor", and that floor IS blast_radius_floor (carrying callers_count) built by prepare_service.py's _build_prepare_blast_radius_floor -- which needs the AST caller path: every registered language now HAS it -- in-file AST refs/callers are real for all ten. Task 11 update (1.110.x): cross-file caller confirmation shipped for Java/PHP/C#/C/C++ via LanguageSpec.file_imports_symbol_from_definition + blast-radius target-symbol binding (#950/#952/#955/#957). Residual reverse-import / text-prefilter gaps may still appear as resolution_gaps on other shapes — quote live JSON, do not assume zero gaps. Go belongs in the parser-backed tier, and this line used to say 4/10 with go as "PARTIAL" -- corrected 2026-07-27. repo_map._symbol_navigation_descriptor's own docstring warns about exactly that mistake: "this tier currently also includes go, which several PR-comment summaries lump in with the 'foundational-only' languages below -- this undercounts it", because lang_go.go_references_and_calls is a full tree-sitter extractor (package-alias resolution, node-type-based ref_kind), not a regex fallback. Never hand-count this -- the product derives it live, so ask the product: python -c "import sys;sys.path.insert(0,'src');from tensor_grep.cli import repo_map as r;print(r._symbol_navigation_descriptor())", which on 2026-08-09 (skills-accuracy audit vs origin/main e3feaf5) prints parser-backed-refs-callers:c-cpp-csharp-go-java-javascript-php-python-rust-typescript+foundational-defs-imports-only: -- 10/0, tier EMPTY -- the 7/3 string this sentence originally quoted was the C#-wave snapshot, now superseded; the dated SUPERSEDED chain at the end of this row records every intermediate state and ends at the C++ wave's 10/0. (The prior "5 of 8 registry specs" denominator was also wrong -- the registry holds 10, and when that note was written 5 had references_and_calls is None. Run it from a CURRENT checkout: a stale tree reports only 5 registered languages and 0 foundational, which looks like a clean answer and is not.) THIRD failure, 2026-08-01 (this skill's own audit pass): a claim framed as "3 parser-backed / 5 foundational / 2 unresolved" invented a third "unresolved" tier from a field's ABSENCE and shipped an unconfirmed hedge as if it were a measurement. There is no third tier and never has been: _symbol_navigation_descriptor partitions every registered LanguageSpec by exactly one boolean test (references_and_calls is not None), so each of the 10 registered languages lands in parser-backed OR foundational, never both, never neither, never "unresolved". Re-verified live 2026-08-01 against this checkout: python -c "import sys;sys.path.insert(0,'src');from tensor_grep.cli import repo_map as r;print(r._symbol_navigation_descriptor())" -> parser-backed-refs-callers:go-javascript-python-rust-typescript+foundational-defs-imports-only:c-cpp-csharp-java-php -- 5 parser-backed + 5 foundational, 10 total, no third bucket. Three failures of the same claim in a row is the tell: do not hand-derive this split by any method (grep, memory, arithmetic on a partial count) -- run the one-liner above and quote its output verbatim. SUPERSEDED 2026-08-04 by PR #927 (Java promoted foundational -> parser-backed, Task 10A, in-file refs/callers only): the same one-liner now prints parser-backed-refs-callers:go-java-javascript-python-rust-typescript+foundational-defs-imports-only:c-cpp-csharp-php -- 6 parser-backed + 4 foundational, 10 total, still no third bucket. SUPERSEDED AGAIN 2026-08-04 by the C# wave (Task 10B, csharp promoted foundational -> parser-backed, in-file refs/callers only): the one-liner now prints parser-backed-refs-callers:csharp-go-java-javascript-python-rust-typescript+foundational-defs-imports-only:c-cpp-php -- 7 parser-backed / 3 foundational. Two promotions landed the same day, so treat ANY quoted descriptor in this file as a dated snapshot and re-run the command. Fourth failure, this row itself, 2026-08-04: PR #927 silently swapped the 2026-08-01 quote above for the post-#927 string while leaving the "still 5 parser-backed + 5 foundational" prose attached to it and dated it to a day the split could not have measured that value -- a live instance of the exact false-receipt failure this row exists to warn against. Never trust a quoted descriptor string without checking its date is consistent with its content; run the one-liner yourself. SUPERSEDED AGAIN 2026-08-04 by the PHP wave (Task 10C, php promoted foundational -> parser-backed, in-file refs/callers only): the one-liner now prints parser-backed-refs-callers:csharp-go-java-javascript-php-python-rust-typescript+foundational-defs-imports-only:c-cpp -- 8 parser-backed / 2 foundational, 10 total, still no third bucket. Three promotions have now landed against this row (Java, C#, PHP), each same-day as the prior in this campaign -- treat ANY quoted descriptor in this file as a dated snapshot and re-run the command rather than trust the number. SUPERSEDED AGAIN 2026-08-04 by the C wave (Task 10D, c promoted foundational -> parser-backed, in-file refs/callers only, mirroring the csharp/php shape -- C owns its own _c_parser() factory, so _c_references_and_calls_for_registry forwards path/symbol only, same as the csharp/php adapters): the one-liner now prints parser-backed-refs-callers:c-csharp-go-java-javascript-php-python-rust-typescript+foundational-defs-imports-only:cpp -- 9 parser-backed / 1 foundational, 10 total, still no third bucket. C's confirmable population is narrower than PHP's: only a call_expression to a bare-identifier function with a real in-file function_definition/prototype confirms (0.9, c-infile-function-declared); a call through a struct member, a function-pointer-variable call, and a function-like-macro invocation are ALL demoted (0.6) -- C has no receiver-type confirmation analogue (no methods, no declared receiver types) and this module does not fake one. Four promotions have now landed against this row (Java, C#, PHP, C), each same-day as the prior in this campaign -- treat ANY quoted descriptor in this file as a dated snapshot and re-run the command rather than trust the number. Accepted ceiling: class MACRO Name in C++ misparses -- do not re-chase. SUPERSEDED AGAIN 2026-08-04 by the C++ wave (Task 10E, cpp promoted foundational -> parser-backed, in-file refs/callers only -- the FINAL wave, closing the top-10 language-support campaign): the one-liner now prints parser-backed-refs-callers:c-cpp-csharp-go-java-javascript-php-python-rust-typescript+foundational-defs-imports-only: -- 10 parser-backed / 0 foundational, 10 total -- the foundational tier is EMPTY. Every registered language now carries a real references_and_calls extractor; the "PARSER-BACKED for X; FOUNDATIONAL-TIER for Y" split framing at the start of this row is retired -- there is no more Y. The remaining gap for ALL ten languages is cross-file caller confirmation, which still relies on the same text prefilter (a resolution_gaps entry names it per language); the prepare floor (blast_radius_floor/callers_count) that hard-stop #6's exemption names keys on in-file AST confirmation, which every language now has. C++'s confirmable population is narrower than PHP's/Java's/C#'s by deliberate choice, not oversight: call_expression to a bare identifier or a qualified name (Foo::bar()) with a real in-file function/method declaration confirms (0.9, cpp-infile-function-declared), as does an explicit this->method() call (the one receiver-typed shape whose type is syntactically fixed, no local-declaration lookup needed); an arbitrary receiver call (w.method(), p->method()) NEVER confirms -- C++'s real inheritance and auto make a general receiver-type walk unsound for the common case, so this module chooses the honest narrower band over a plausible-but-wrong wider one (see lang_cpp.py's TASK 10E docstring block for the full reasoning). Five promotions have now landed against this row (Java, C#, PHP, C, C++), each same-day as the prior in this campaign -- treat ANY quoted descriptor in this file as a dated snapshot and re-run the command rather than trust the number. |
Beat-rg cold search | Closed — honest negative (#261). Startup is at parity (rg 6.2ms / tg 6.5ms), GPU is dead (3 proofs), .tg_index measured NET NEGATIVE (~10x slower), and tg's native walk is rg's walk (same ignore crate) — so widening it relocates cost, never removes it. The campaign's return was a defect family, not milliseconds. Do not re-measure. |
| LSP proof | Open |
| Trust: incompleteness disclosure across the CLI | Partial (#292). §0 of docs/CONTRACTS.md pins the completeness contract; a silent-loss census ratchet guards regressions; 7 disclosure defects fixed this cycle (codemap, checkpoint-undo data-loss, tg scan). Do NOT cite the trust-benchmark "tie" as a tg result. Adversarial review 2026-07-26 found the harness's tg row runs plain tg search (trust_benchmark.py:79, re-grepped 2026-08-01, unchanged), which forwards to real rg via _run_rg_passthrough whenever no native binary resolves -- grep -n 'resolve_ripgrep_binary()|_run_rg_passthrough(binary_name' src/tensor_grep/cli/bootstrap.py; the cited bootstrap.py:1497/:1269-1277 had drifted to unrelated code (a TimeoutExpired handler in a backward-compat shim), and the :1587/:1584 re-stamp that replaced them rotted too (verified at this SHA: the forwarding call _run_rg_passthrough(binary_name, passthrough_search_args) is :1685, gated by the resolve_ripgrep_binary() check at :1682 -- was :1587/:1584, now :1685/:1682; the grep is the citation, never the stamp) — the literal rg: prefix exists nowhere in our source, while tg's own is tg: {err}. So that row is ripgrep measured twice. The "tg is BEHIND on --json" verdict that stood here is now FALSE and is corrected in the row below -- re-measure before quoting any comparator claim from this table. |
Trust: --json/--ndjson incompleteness marker | CLOSED 2026-07-27 (#276, nine PRs) -- do not re-open this row from the stale verdict it replaced. It read "Open, and the load-bearing enterprise gap ... the native JSON envelope reports success while walk errors go only to stderr", and the row above concluded "on the --json surface that actually matters tg is BEHIND: rg --json exits 2 on an unreadable path, tg --json exits 0 with a success envelope". Both were true when written; both are false now, and a CEO-facing verdict that outlives its evidence is a decision input pointing the wrong way. Measured on the SHIPPED v1.101.4 (uvx --from "tensor-grep==1.101.4"), one ACL-denied directory beside a readable sibling, with the denial asserted to bite and the sibling asserted still listable before either arm ran: rg --json exit 2; tg --json exit 2 carrying result_incomplete: true, incomplete_reason_class: "unreadable_path", incomplete_paths_count: 1, routing_backend: "NativeCpuBackend". Exit-code parity, and tg additionally carries the cause IN-BAND where rg signals only via exit code plus stderr. Scope that claim honestly: one shape (unreadable directory, native CPU route, Windows) -- it establishes the gap is closed, not a general benchmark lead, and #72 (publishing benchmark claims) stays CEO-gated. |