# Tensor Grep Prepare

> Use when an agent needs one-call edit readiness before changing code — tg prepare returns primary target, confidence, callers/blast-radius floor, validation commands, and claim/evidence coordination hooks. Prefer over the multi-step orient→search→agent→route-test→callers→evidence→ledger loop for routine edits. Pair with tg install-dense once for semantic find quality.

- Skill: `oimiragieo/tensor-grep-prepare` (Agent Skill)
- Install (CLI): `npx skillmds@latest add oimiragieo/tensor-grep-prepare`
- Raw SKILL.md: https://api.skillmd.com/api/skills/oimiragieo/tensor-grep-prepare/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: oimiragieo (https://skillmd.com/u/oimiragieo)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/oimiragieo/tensor-grep-prepare

---


# tensor-grep prepare (one-call edit readiness)

Verified against **tg 1.110.14** — honesty note (2026-08-13): this stamp was bumped from 1.110.13
by the 2026-08-11 drift sweep (`docs/audits/2026-08-11-skill-audit-findings.md` item 6), but the
dogfood table below records **no dated 1.110.14 row** — its newest rows are **1.110.13**, and the
~5s saddle / ~22s `--out`/`--claim` numbers this header previously presented as a LIVE 1.110.14 run
are byte-for-byte the 1.110.13 receipts. The 1.110.14 stamp therefore currently rests on the
1.110.13 receipt (prior 1.110.12 CUJ still representative); a future pass must either add a dated
1.110.14 dogfood row or keep citing 1.110.13, as this header now does.

Note: anonymous `--claim` `agent_id_hint` remains strong (“NOT attributable…” until
`TG_LEDGER_AGENT_ID` / `TG_EVIDENCE_AGENT_ID` is set).

## When to use

| Need | Command |
| --- | --- |
| Single call before editing | `tg prepare REPO/src "task" --json` |
| Also submit advisory ledger claim | `tg prepare REPO/src "task" --claim --json` |
| Persist the capsule for evidence emit | `tg prepare REPO/src "task" --out capsule.json --json` |
| Whole-repo / large root | `tg prepare REPO "task" --deadline 20 --json` (expect partial) |

Prefer **`REPO/src`**. Default deadline is 60s (like agent cold path); pass `--no-deadline` only when intentional.

## What you get

- `primary_target` + `confidence` + `ask_user_before_editing`
- `validation_commands`
- `blast_radius_floor` (callers_count, top_callers, trust summary)
- `coordination.claim` — args ready; `submitted=true` only with `--claim`; an anonymous claim
  (neither `TG_LEDGER_AGENT_ID` nor `TG_EVIDENCE_AGENT_ID` set) additionally stamps
  `coordination.claim.agent_id_hint` with the STRONG hint — "this claim is filed as 'anonymous'
  and is NOT attributable to you. Set TG_LEDGER_AGENT_ID (or TG_EVIDENCE_AGENT_ID) to a stable
  per-agent value so other agents can see WHO holds it." — plus the machine-branchable
  `agent_id_is_anonymous=true` sibling, so the caller knows why `agent_id` reads `anonymous`.
  (The earlier weak hint "set TG_LEDGER_AGENT_ID for a stable identity" is RETIRED — superseded
  per the "Strengthened from" comment in `src/tensor_grep/cli/prepare_service.py`; grep
  "Strengthened from" to re-verify. Do not re-quote the weak form.)
- `coordination.evidence` — argv/note to emit a receipt without guessing flags

Dogfood:

| Case | Result |
| --- | --- |
| `prepare …` saddle (**1.110.13**) | PASS ~5s |
| `prepare … --out/--claim` tg `src` (**1.110.13**) | PASS ~22s |
| `prepare …` saddle (**1.110.12**) | PASS ~6.5s |
| `prepare … --out/--claim` tg `src` (**1.110.12**) | PASS ~20s |
| `prepare …` saddle (**1.110.10**) | PASS ~6s |
| `prepare … --out/--claim` tg `src` (**1.110.10**) | PASS ~14s |
| `prepare … --out` (**1.101.31**) | PASS ~8s — overall 0.9, callers_count=1, file ~8KB |
| `prepare … --claim` (**1.101.31**) | PASS ~13s — anonymous + strong hint; env id clears hint |
| `prepare … --out` (**1.101.22**) | PASS ~7s — overall 0.9, callers_count=1, file ~8KB |
| `prepare … --claim` (**1.101.22**) | PASS ~9s — stronger anonymous `agent_id_hint`; env id clears hint |
| `prepare … --out` (**1.101.19**) | PASS ~7s — overall 0.9, callers_count=1, file ~8KB |
| `prepare … --claim` (**1.101.19**) | PASS ~8s — anonymous + hint; env id clears hint |
| `prepare … --out` (**1.101.17**) | PASS ~7s — overall 0.9, callers_count=1, file ~8KB |
| `prepare … --claim` (**1.101.17**) | PASS ~8s — anonymous + hint; `TG_LEDGER_AGENT_ID` clears hint |
| `prepare … --out` (**1.101.9**) | PASS ~6s — overall 0.9, callers_count=1, file ~8KB |
| `prepare … --claim` (**1.101.9**) | PASS ~8s — submitted=true + `agent_id_hint` when anonymous |
| `prepare core/hooks matches_trigger` (**1.101.7**) | PASS ~13s — overall 0.9, callers_count=1, claim.submitted=false |
| `prepare … --out` / `--claim` (**1.101.7**) | PASS — file ~8KB; submitted=true + `agent_id_hint` when anonymous |
| `prepare core/hooks matches_trigger` (1.92.1) | PASS ~8s — overall 0.9, callers_count=1, claim.submitted=false |
| `prepare … --claim` (1.92.1) | PASS ~9s — claim.submitted=true (`agent_id` anonymous unless env set) |
| `prepare … --out capsule.json` (1.93.0) | PASS — file written byte-identical to stdout JSON; symlink/dangling-symlink/dir destination refused |
| `prepare tensor-grep/src` (1.91.0) | PASS ~27s — callers_count=2, claim.submitted=false |
| `prepare tensor-grep --deadline 20` (1.91.0) | INCOMPLETE — partial/deadline (same class as whole-repo agent) |

**`--out FILE` (v1.93.0, A12(d)) replaces the old manual-redirect workflow.** Chain straight into
evidence emit, no separate save step:

```bash
tg prepare REPO/src "task" --out capsule.json --json
tg evidence emit REPO --capsule capsule.json --query "task" --json --agent-id "$AGENT_ID" --out receipt.json
```

Without `--out`, prepare still does **not** write the capsule file on its own — save stdout instead
(`tg prepare ... --json > capsule.json`), or use `coordination.evidence.argv`.

**Ledger cross-ref:** `--claim`'s behavior is affected by the #706 PATH-canonicalization fix (see
`tensor-grep-ledger`) — a claim submitted here is now visible from `tg ledger list` at any subtree PATH
under the same repo, not just the exact PATH `prepare` was run against.

## Hard stops

1. `ask_user_before_editing.required`
2. `partial` / `partial_reason`, `scan_limit.possibly_truncated`,
   `blast_radius_floor.possibly_incomplete`, exit `2` — do not claim full coverage. (These are
   prepare's REAL incompleteness signals — `partial_reason="deadline"` folds into the top-level
   exit-2 honesty gate; a plain `max_callers` output cap stays exit 0. Note prepare emits NO
   `result_incomplete` field: `grep -c result_incomplete src/tensor_grep/cli/prepare_service.py`
   is 0, verified 2026-08-13 — an earlier revision of this hard stop listed it in error.)
3. Empty `blast_radius_floor` with `source=no_primary_symbol` — narrow PATH or raise budget

## Dense semantic companion

```bash
tg install-dense --json   # once per machine; never auto-runs
tg find "intent" REPO/src --deadline 20 --json   # dense leg after install
```

Dogfood: `install-dense` PASS ~4s (uv-tool + potion-code-16M); post-install `tg find` no longer reports BM25-only `rank_fallback_reason`.

## Related

- `tensor-grep`, `tensor-grep-ledger`, `tensor-grep-find-and-route`, `tensor-grep-enterprise-agent`