tensor-grep-enterprise-agent

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.

tensor-grep for enterprise agents

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.

Guidance

• Default edit gate: 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.
• Prefer REPO/src. Whole-repo: tg prepare|agent REPO --deadline N → expect partial / ask_user.
• The WSL 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.
• At real workspace scale — native Windows, 300k+ files, a separate data point from the WSL caveat above: 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).

Hard stops

1. ask_user_before_editing.required
2. Full-coverage claims on partial / exit 2
3. Unscoped workspace search refuse
4. GPU promotion without search_ready
5. review-bundle create without --manifest
6. route-test.agreement == false (when not using prepare's floor)
7. Treating ledger overlaps as hard locks

Enterprise gaps (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

Recommended loop

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"

Published-wheel dogfood (2026-08-06)

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).

Retention (2026-08-15)

• DD-006: demand proven + design on main (#1015) is still an open row until PERF+HONESTY product code ships under deliberate authorization (A122). Do not treat the design packet as CUJ-complete.
• Skip-Fable waivers do not clear enterprise security gates on daemon acceptor code (A117/A3).