fix(review-telemetry): record the run's models and skills version

[ultimatile]

Summary

review-telemetry records which gate surfaced which finding, but nothing about the model that produced it. No field distinguishes two runs whose gates ran under different models, so a per-gate comparison across the log pools them without saying so.

Schema 4 names the model and skills state each run ran under.

Changes

Records carry schema: 4 with three run-level fields — skills_version, fixer_model, fixer_effort — and one per-gate field, gates[].reviewer_model.

Each new field names the state the run ran under rather than the state at record time. Where the recording session cannot vouch that the two agree — the model changed mid-run, the skills tree moved after the gates read it, the record is written from a later session — the field falls to the path the record already has for a value nobody can attest: ask the user, and failing that write null and name the reason in gaps.

gates[].reviewer_model is always written and takes one of three values, first branch that applies: "unobservable" when some iteration ran under a reviewer that exposes its model to nobody; null when some iteration's model is not known; otherwise every iteration's model, deduplicated and joined by +, so a gate split across two models stays out of either model's bucket rather than crediting one. Absence carries no meaning, which keeps a forgotten field from reading as a positive claim about the model.

The schema note now tells a reader to gate any query over the four fields on select(.schema >= 4), alongside the same instruction for the fields earlier bumps added. No aggregation one-liner for them joins the existing ones under Reading the log: the skill defines the record, and the queries are the analyst's.

Impact

Records written before this change stay readable: the four fields are additions, and no existing aggregation one-liner reads any of them.

An analysis session reading reviewer_model finds it on every gate object, so a bare field read needs no presence test.

Verification

git describe --tags --always --dirty, run through the path skills_version names, returned 2026.8.5-1-gc2f48c4 — the <tag>-<count>-g<sha> form the record-shape example carries. $CLAUDE_EFFORT, the variable fixer_effort names, expanded to high.

The record-shape example parses: extracting the fenced block and piping it to jq -e . exited 0. mdformat --number --compact-tables --check exited 0 on skills/review-telemetry/SKILL.md.

Notes

skills_version is produced by a git command against the skill's own directory, so it resolves only where that directory is a checkout of this repository. Under the claude plugin install route the README documents, that directory is not a checkout, the command fails, and the field records null with the reason named in gaps.

[Copilot]

Pull request overview

Updates the review-telemetry skill’s JSONL record schema to capture the model/effort context each run and gate executed under, enabling accurate comparisons across telemetry collected under different model or skills states.

Changes:

• Bump the telemetry record schema from 3 → 4 and add run-level fields: skills_version, fixer_model, fixer_effort.
• Add per-gate gates[].reviewer_model with normalization rules and update the schema-gating note accordingly.
• Bump plugin marketplace version to 2026.8.6.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
skills/review-telemetry/SKILL.md	Extends telemetry schema and documentation to record run/gate model context and normalization rules.
.claude-plugin/marketplace.json	Version bump reflecting the skill behavior/schema update.

Suppressed comments (1)

skills/review-telemetry/SKILL.md:99

• This says gates[].reviewer_model is “always written”, but that’s only true starting with schema 4. Adding the schema qualifier here prevents confusion when reading older records (where the field is absent).

- `gates[].reviewer_model` is always written, and takes the first branch that applies. **`"unobservable"`** — some iteration ran under a reviewer that exposes its model to nobody, `copilot-pr` being the case that reaches it. **`null`** — otherwise some iteration's model is not known. **A string** — otherwise: every iteration's model, deduplicated and joined by `+` in lexicographic order, so a split gate writes one token and stays out of either model's bucket rather than crediting one, and a gate that ran wholly under `fixer_model` repeats that value. Whatever the branch, `gaps` names every iteration whose model was knowable and went uncaptured.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.