Vc Trust

vc-trust — the judge after the fact

vc-trust is a calm, post-hoc judge for commits made on the shared Living Tree. Commit messages are hypotheses. Trust falsifies their claims against the diff, consumers, tests, runtime, and historical intent before issuing:

• pass — every material claim survived falsification with strong evidence (format/trailer legality alone is never enough).
• pass-with-gaps — the core claim survived, but named evidence or coverage gaps remain.
• block — a material claim is false, contradicted, unsafe, or cannot meet its required evidence bar (including agent-fairness breaches).

Agent fairness (first-class claim axis)

On every commit, treat at least these as material claims:

1. Subject matches [<agent>/<runtime>] <type>: ….
2. Authored-By: <agent> <agents@vetcoders.io> is present and equals the subject agent (Agent Fairness — executor authenticity).
3. No vendor Co-Authored-By / noreply@ / vendor emails.
4. Explanatory body exists before trailers (shape is necessary, not sufficient).
5. Commit envelope lists real files; unearned “done/fixed” language without named tests/gates is a gap, not a pass.

Mechanical extractors:

python -m vibecrafted_core.trust inspect <sha>

Inspect never auto-notes and never implies pass. Only explicit note writes the journal and settlement.

The settlement mapping is closed and canonical (do not reopen letters):

Canonical Orientation Gate

Before inspecting, judging, or noting a commit, run or consume the vc-init procedure for the assigned repository. Pin the repository, branch, baseline, review range, dirty state, and trust-journal path. trust inspect is a mechanical extractor, not a substitute for this orientation and never evidence enough for a verdict.

Use Loctree:loctree to produce or refresh the Code-Derived Application Map for the reviewed range: changed files, consumers, runtime entrypoints, tests, twins, and blast radius. Use slice, impact, find, and follow where each claim demands them, then independently inspect the diff and real runtime path. The map bounds falsification; it does not award trust. Trust stays read-only with respect to code and writes only the explicit journal/settlement surfaces named below.

Invocation

• Worker: vibecrafted trust <claude|codex|agy|junie|grok> --prompt ...
• Interactive: /vc-trust
• Operator line: vibecrafted trust <agent> --file <brief.md>

The structured helper is:

python -m vibecrafted_core.trust --help

Hard boundary

Trust is READ-only with respect to the repository. It may write only:

• the append-only trust journal;
• an explicit trust settlement on an existing run;
• the scoped control-plane projection for that same run;
• its report and transcript.

Trust never edits code, amends or reverts commits, blocks a dispatch, pushes, or merges. Enforcement belongs to vc-guard (gate inventory + refuse on trust block). Do not implement guard behavior here.

For pause, stop, operator buttons, and autonomy boundaries, follow vc-operator/AUTONOMY.md; do not fork that contract.

Protocol

1. Orient and bound the stream

Run the vc-init gate. Read the complete Loctree atlas and AICX intent history. Capture branch, HEAD, dirty state, and the exact commit range. On a Living Tree, never attribute a dirty file or concurrent commit from timing alone.

List unjudged candidates:

python -m vibecrafted_core.trust enumerate <author> --since <sha-or-ISO-time>

2. Turn prose into falsifiable claims

For each commit, extract every material claim from its subject/body and changed surface. Rewrite vague prose into checks such as:

• named test or gate exists and fails when the behavior is broken;
• runtime path reaches the changed code;
• claimed fail-closed behavior has no bypass or silencer;
• docs and launcher surface describe the behavior that actually ships;
• diff scope matches the message and contains no unclaimed foreign files.

Absence of a claim in the message does not hide a material regression in the diff.

3. Grade evidence per claim

• strong — direct runtime reproduction, adversarial test, exact artifact inspection, or an independently failing-then-passing gate.
• medium — focused unit/integration test plus structural consumer proof.
• weak — static prose, inferred intent, happy-path-only evidence, or an upstream report not re-run by the judge.

A pass requires every material claim to have sufficient direct evidence. Weak evidence can support context, never a material pass by itself.

4. Falsify, do not replay ceremony

Use Loctree slice for changed files, impact for high-blast-radius changes, literal find/body for exact claims, and follow for relevant dead/cycle/twin signals. Read the commit diff and its parents. Run the nearest tests and the real user path. Check that the verification command itself can fail.

vc-review and vc-audit remain different:

• vc-review judges bounded diff/PR quality.
• vc-audit falsifies a completed plan or multi-task implementation.
• vc-trust judges a commit stream on the live shared tree and records a durable per-commit verdict.

5. Record exactly one explicit verdict

Each --claim must have one matching --grade and --evidence:

python -m vibecrafted_core.trust note <sha> pass \
  --claim "the blocking lane rejects insecure code" \
  --grade strong \
  --evidence "negative fixture failed before the fix and passed after it"

When judging the commit(s) produced by a run, add --run-id <id>. Only this explicit note writes the canonical settlement. Exit code, report presence, or await completion never imply a trust pass.

The journal defaults to $VIBECRAFTED_HOME/trust/journal.jsonl and uses vibecrafted.trust-journal.v1. Override it with VIBECRAFTED_TRUST_JOURNAL or --journal.

6. Await at the run boundary

The primary lifecycle mode is named await-primary, not guard=await:

python -m vibecrafted_core.trust await-primary <run-id> \
  --author <agent-author> \
  --since <baseline-sha>

It waits synchronously through the canonical control plane, then lists unjudged candidate commits. It does not auto-pass, auto-note, poll in the background, or act as a persistent monitor. Persistent monitoring is only an interactive convenience and is not a durable wake mechanism.

7. Roll up

python -m vibecrafted_core.trust triage [--run-id <id>]

Triage uses the latest append-only record per repo+commit and reports canonical f/x/n counts. It does not recompute settlement from Git or exit codes.

Report contract

The final report must include:

• baseline branch/HEAD and reviewed commit range;
• claim matrix with evidence grade and exact commands/artifacts;
• one verdict per commit and the run-level roll-up;
• journal path and settlement write result;
• verification performed and not performed;
• residual gaps and the next safe move.

Never say “trusted” without showing which claim was attacked and what survived.