Tensor Grep

Tensor-Grep Skill

Use this skill when you need to locate code precisely, understand likely edit impact, or prepare a minimal patch in a real repository.

When To Use

• You need the primary definition or source block for a symbol.
• You need references or likely callers before editing code.
• You need an edit plan, blast radius, or validation target instead of ad hoc grep loops.
• You are preparing a patch and want a smaller, more accurate context bundle.
• You need a fast codebase orientation capsule (central files, entry points, symbol map) before diving into symbol lookup.
• You need to find code by text/content relevance rather than an exact symbol name.
• You need to resume or persist cross-session repo-map context — use tg session to cache the repo-map, then call session-scoped commands (tg session context-render, tg session edit-plan, tg session blast-radius-render) without re-indexing on each invocation.

Argument Order

All symbol commands are path-first: tg <command> <REPO_PATH> <SYMBOL>. A reversed <SYMBOL> <REPO_PATH> call is auto-corrected with a stderr hint, and a single tg <command> <SYMBOL> resolves against the current directory — but prefer the canonical path-first form.

Default Workflow

1. Confirm the installed CLI is available:
   • tg --version
2. (Unfamiliar repo) Orient — single repo preferred; workspace root works (~4.9s cold-scan, last measured v1.95.0; a warm session-daemon hit is faster still):
   • tg orient REPO_PATH
   • tg inventory REPO_PATH --json
3. File deps (cheap):
   • tg imports FILE / tg importers FILE [ROOT] — absolute paths; importers may be deadline-partial
4. Content search then source:
   • tg search PATTERN REPO_PATH --rank
   • Vocabulary mismatch: tg find "intent" REPO_PATH/src --deadline 20 --json (run tg install-dense once; see tensor-grep-find-and-route)
   • Multi-project: tg search PATTERN . --glob "*.py" --max-depth 3 (bare text/--json without PATH: exit 1 + note; --json also sets path_was_defaulted/scope_note @ 1.101.31 — always pass PATH; unscoped multi-project parent refuses exit 2 with incomplete_reason_class/error.code=workspace_root_refused @ 1.110.x — not generic scan_limit)
   • tg source REPO_PATH/src SYMBOL
5. Symbol navigation — prefer src/:
   • tg callers REPO_PATH/src SYMBOL --deadline 15 --json
6. Edit readiness — prefer tg prepare REPO/src (~5–22s PASS on gotcontext-saddle / tg src @ 1.110.14+; ~27s on larger src @ 1.91.0):
   • tg prepare REPO_PATH/src "task" --json # primary + blast floor + validation + coordination hooks
   • tg prepare REPO_PATH/src "task" --out capsule.json --json # also persists the full capsule to FILE (byte-identical to stdout JSON; symlink/dangling-symlink/dir refused; feeds tg evidence emit --capsule FILE directly, no manual redirect)
   • tg prepare REPO_PATH/src "task" --claim --json # also submit advisory ledger claim; anonymous claims stamp coordination.claim.agent_id_hint unless TG_LEDGER_AGENT_ID is set
   • Fallback loop: tg agent + tg route-test (budget 90s) if prepare unavailable
   • Whole-repo: explicit --deadline N on prepare/agent; bare tg agent REPO still TIMEOUT empty @75s
   • Mega-repos: narrow PATH; deadline partials often null symbol 5a. Multi-agent ledger — see tensor-grep-ledger. Claim/release/list now canonicalize to the nearest .git ancestor (worktree-aware, one store per repo) — list rolls scope UP, so the PATH-mismatch footgun from earlier dogfood rounds is fixed:
   • tg ledger claim REPO --symbol SYM --agent-id AGENT --json
   • tg ledger list REPO --json # or any subtree PATH under REPO — rolls up to the same store
   • tg ledger record REPO --receipt receipt.json --symbol SYM --agent-id AGENT --json
   • tg ledger find REPO --symbol SYM --fresh-only --json then release (a zero-match release with --claim-id/--symbol emits unmatched_reason + live_claims_elsewhere; a bare-path release with neither fails closed) 5b. Evidence receipt:
   • tg evidence emit REPO_PATH --capsule capsule.json --query "task" --json --agent-id AGENT
   • Note TG_CAPSULE_INLINE_CALLERS (default-OFF): when set, tg agent/tg prepare prepend # tg: callers=N (top: a, b) to the primary snippet and add snippets[i].inline_structural_annotation (~+2.8% tokens) — reuses already-collected blast-radius evidence, no new scan. 5c. Skip tg codemap on WSL (TIMEOUT 90s) 5d. GPU: see tensor-grep-gpu — default loops stay CPU
7. Make the smallest correct edit from primary targets.
8. Run only the returned validation commands.
9. Cached loops: tg session open --json REPO_PATH then tg session context-render SESSION_ID ABS_ROOT "query"
10. Enterprise: tg review-bundle create --manifest … --json

Registration-Audit Workflow (blast-radius before claiming done)

When you add an entity that must be registered in multiple places (a command, a flag, a route, a hook), enumerate ALL its registration sites BEFORE claiming the change is done — missing one fails quietly. The default audit path:

1. Blast radius — tg callers PATH SYMBOL --json lists every call site (file:line). On a real billing repo it surfaced 2 webhook handlers + 1 reconcile cron in ~1s — a 10-minute grep-and-read became a one-second decision. When the JSON has "result_incomplete": true, the call-site list was TRUNCATED by a scan/output cap — treat coverage as partial; do not conclude unlisted sites are safe. Human mode emits a loud stderr caveat.
2. Pattern bugs — tg scan PATH --ruleset RULESET runs a built-in security/compliance rule pack across those sites (see tg rulesets for pack names). --config sgconfig.yml and --rule FILE are separate options for a custom ast-grep config or a single rule file — not for built-in packs.
3. Diagnostics — tg doctor --with-lsp.

For registration-completeness specifically: tg callers PATH REGISTRATION_FUNCTION lists callable registrations — but the call graph can't see set/list/decorator registrations (allow-lists, @router.post, dispatch tables), which are often the missed site, so grep / tg scan those too. Your new entry must appear in ALL sites. (General principle: verify-plan-against-code Hard Rule 6; call-graph blind spots: tensor-grep-code-audit P7.) A resolved zero-caller result is NOT dead code either — the call graph can't see set/list/decorator/dispatch-table registrations; cross-check with tg scan or grep before removing a zero-caller symbol. As of v1.17.1 the registration-completeness checker (extract_members) is string/comment-aware, so #-commented entries are no longer surfaced as false members.

tg imports/tg importers/tg blast-radius now report a relative dynamic import (import_module(".x", package=...), __import__(..., level>=1)) as dynamic_unresolved — the literal text is preserved in unresolved, and it is NEVER silently resolved to a same-named decoy top-level file (both forward and reverse directions, and excluded from blast-radius's reverse scoring prefilter too). A wrong edge is worse than a missing one; treat dynamic_unresolved as "re-check yourself," not as a resolved dependency.

Non-Interactive Mode

• When running in claude -p or other non-interactive automation, do not ask for confirmation.
• Use tg against the repository path that was added via --add-dir.
• After tg identifies the likely file/span, make the change directly instead of stopping at analysis.
• If direct editing is unavailable, emit a clean git-style unified diff only.

Rules

• Prefer tg over repeated manual grep loops when working inside a real repository.
• Run tg orient REPO_PATH first when entering an unfamiliar repo — it gives centrality, entry points, and a symbol map in one call, and costs no API key or GPU.
• Use tg search PATTERN PATH --rank for content/text search; prefer it over raw grep loops when relevance ranking matters. The --bm25 flag is an alias for --rank.
• Keep edits narrow and grounded in the files tg ranks highest.
• Do not expand context blindly if tg already identified the primary file and span.
• Use provider-backed modes only when a task is clearly about semantic ambiguity.
• In non-interactive mode, do not return “want me to apply this?” style responses.

Known Issues

Latest CUJ dogfood: v1.110.14 (unchanged through v1.110.16; re-run the CUJ on the published wheel before the next restamp) (2026-08-11, Windows uvx, artifact C:\Users\Public\tg-dogfood-111013.json — 21/21 PASS). Core CUJ + M16/M17 still green. A90 shipped: reserved Phase-2 names edit-ready / verify-edit / workspace with a flag (--help/--json) fail closed — exit 2, stderr unknown_command (JSON on stderr for --json, nearest: []). Typo + --help suggests nearest (searhc → search). Bare tg edit-ready without a flag still searches that token by design (escape: tg search edit-ready). Live callers coverage still 10/10 parser-backed. Parent refuse still workspace_root_refused. See tensor-grep-workspace-dogfood.

Parent refuse vs scoped empty: unscoped tg search needle C:\dev\projects --json → exit 2 + workspace_root_refused. Scoped … --glob "*.py" --max-depth N does not hard-refuse (may exit 1 with zero matches = complete empty). Skill text that said parent class=scan_limit is stale.

Historical: last full workspace+GPU sweep was v1.91.0 (WSL). Language campaign closed through Task 10E + F7 Task 11 (#957). Prefer the 1.110.13 CUJ table over older 1.95.0 “7 of 10” prose.

Prefer tg prepare REPO/src over the multi-step agent loop for routine edits. Whole-repo prepare/agent with --deadline still partial/null-symbol; bare agent TIMEOUT empty @75s.

tg install-dense: once per host; post-install tg find drops the BM25-only rank_fallback_reason (the fallback message itself now leads with tg install-dense when dense is absent).

tg ledger: claim/release/list/record/find — see tensor-grep-ledger. Slice 1 + Slice 2 PATH-canonicalization fixed (A13 / #850).

Unscoped multi-project search refuses fast (~1–2s). Workspace-root safety refuse uses workspace_root_refused; the generic >1500-file defaulted-PATH ceiling is a separate gate. Escape hatches: explicit PATH, --max-depth, or --allow-broad-generated-scan — --glob/--type alone do NOT bypass the defaulted-PATH ceiling. Prefer per-repo for deep --type ts.

tg codemap still TIMEOUT on WSL (90s). Importers/callers-at-root may be deadline-partial.

CLI traps: tg classify has no --json; scan ruleset names from tg rulesets; session context-render needs absolute session-root PATH; reserved+flag refuses (A90); bare unknown tokens may still search.

GPU (experimental) — verified on v1.91.0, no change through v1.93.2

SUPERSEDED stamp (2026-08-11, append-only): unchanged through v1.110.14 -- GPU remains experimental/unpromoted (no crossover proven; see tensor-grep-gpu + docs/gpu_crossover.md). Re-verify every release.

Hardware visible (2× ~12GB). Build without CUDA: calibrate FAIL; search GPU → CPU fallback; doctor search_ready=false. v1.93.0 (A11) fixed the WSL bare-shim cross-domain misclassification that produced a bogus path_not_found; full detail in tensor-grep-gpu.

Provider Modes

• Default: native
• Optional:
  • tg defs REPO_PATH SYMBOL --provider lsp
  • tg blast-radius REPO_PATH SYMBOL --provider hybrid

Use lsp or hybrid only if native lookup seems ambiguous or incomplete.

Patch Guidance

• Prefer editing files directly if your tools allow it.
• If you must emit a patch, make it a git-style unified diff with diff --git headers and enough context lines to apply cleanly.
• Avoid unrelated files, caches, summaries, or prose.

Reference

See REFERENCE.md for current command patterns and examples.