[gold-silver-copper]

What and why

meta git snapshot was removed from meta (meta#115; #5 already removed the guard's snapshot hints). agent score still graded Snapshot safety at 25% of the overall grade: a destructive command only counted as safe if meta git snapshot create ran in the 5 commands before it. That command no longer exists, so every session with a destructive command lost a quarter of its grade, and the suggestion told agents to run a command that fails.

This drops the metric outright rather than replacing it: text matching can't reliably tell whether a commit or stash protected the work a later command destroys, and agent guard already blocks the worst destructive commands.

Removed end to end in src/score.rs: snapshots_before_destructive, snapshot_safety_score/grade, the transcript counting, the grade, the report row, the suggestion and the matching JSON fields. destructive_ops_detected went too: only this metric and its "Destructive ops: N (M protected)" summary line used it. The per-command BashCommand.is_destructive flag in the JSON stays (it's per-command classification, not part of the metric). README metric list updated.

Weights

The remaining weights keep their relative proportions (25:20:20:10 = 5:4:4:2):

Metric	Before	After
Meta-command ratio	25%	33% (5/15)
Workspace discovery	20%	27% (4/15)
Snapshot safety	25%	—
Cross-repo awareness	20%	27% (4/15)
Guard effectiveness	10%	13% (2/15)

// before
let overall = (meta_ratio * 0.25) + (discovery_score * 0.20) + (snapshot_score * 0.25)
    + (awareness_score * 0.20) + (guard_score * 0.10);
// after
let overall =
    (meta_ratio * 5.0 + discovery_score * 4.0 + awareness_score * 4.0 + guard_score * 2.0) / 15.0;

Report before / after

Same 4-command transcript (meta context, meta git status, git reset --hard HEAD, meta git commit -m x), old binary vs this branch:

before: # Session Score: demo (D)            after: # Session Score: demo (B)
| Meta-command ratio | 67% | D | 25% |        | Meta-command ratio | 67% | D | 33% |
| Workspace discovery | 100% | A ✓ | 20% |    | Workspace discovery | 100% | A ✓ | 27% |
| Snapshot safety | 0% | F ✗ | 25% |
| Cross-repo awareness | 100% | A ✓ | 20% |   | Cross-repo awareness | 100% | A ✓ | 27% |
| Guard effectiveness | 100% | A ✓ | 10% |    | Guard effectiveness | 100% | A ✓ | 13% |
- Destructive ops: 1 (0 protected)
- Only 0/1 destructive operations were preceded by snapshots (1 unprotected).
  Use `meta git snapshot create` before batch changes.

Verification

• cargo fmt --check, cargo clippy --all-targets -- -D warnings, cargo test (126 passed) locally.
• New test_overall_weights: a session with only bare git (meta ratio 0, everything else perfect) is D under the new weights (10/15 ≈ 0.67); under the old weights it was C (0.75), so the test pins the rescale. The old test_snapshot_protection is gone; other tests just lose the snapshot argument.
• Ran old and new binaries on the transcript above; agent score --json output has no snapshot key.
• grep -rni snapshot . (excluding .git): no matches.

Behavior change

Intended: overall grades change because the weights changed. Sessions with no destructive commands used to get a free 25% from this metric (old = 0.25 + 0.75 × new), so they can grade lower now; sessions with destructive commands no longer lose that 25%.

For a reviewer / human

Not changed here, but worth a look: the transcript structs use literal serde names _uuid, _role, _id (and _text, _tool_use_id, ...). Claude Code transcripts use uuid/role/id, so as far as I can tell real transcripts never parse into tool calls and agent score reports 0 tool calls. I needed those literal keys to build the demo transcript above. Out of scope for this PR.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: Organization UI
• Review profile: ASSERTIVE
• Plan: Advanced
• Run ID: fac3f274-c7c3-425e-8ba4-de40e32b2842

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

_{Comment @coderabbitai help to get the list of available commands.}