agentleFS
okeefecookeefeco/pyeye-mcp/.claude/skills/pyeye-verify/SKILL.md

pyeye-verify

Skill0 starsChanged yesterdayMITView on GitHub

Use when verifying pyeye for regressions or evaluating how well it serves an AI agent across scenarios — "verify pyeye", "check pyeye for regressions", "dogfood pyeye against django/jaraco", "did pyeye regress", "score pyeye agent consumption". Runs a scenario's probe suite, diffs against its recorded baseline, and scores an agent-consumption rubric. Uses the pyeye-scenarios skill for the scenario data. Does NOT trigger for ordinary code navigation (use python-explore) or for the pytest unit suite.

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 1 is distinctive, and 2 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 2

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Run it from the checkout whose build you are verifying.

    line 96
  2. Very specific:

    Don't re-implement this check in pytest.

    line 109
  3. Distinctive:

    If a scenario reveals a miss that the fixtures don't cover, add a fixture case to #614's test rather than a second copy of the check.

    line 112

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

pyeye Verify

A repeatable, scenario-independent procedure for two questions:

  1. Did pyeye regress? — diff live probe output against a scenario's recorded baseline.
  2. How well does pyeye serve an AI agent? — score the shape of the output against the agent-consumption rubric, not just whether it's correct.

The scenarios, pins, probe suites, and recorded baselines are the data, held in [[pyeye-scenarios]]. This skill is the procedure that runs against any of them.

Announce at start: "Using the pyeye-verify skill to verify <scenario> on <build>."

Why a skill harness, not pytest

This complements the pytest suite — it doesn't replace it. These scenarios hit real third-party repos at pinned commits (too heavy for CI; never vendored), and the agent-consumption rubric is qualitative (altitude, honesty, response shape) — awkward to assert in pytest, natural for an agent rubric. Run on demand; not a CI gate. Distinct from python-explore (user-facing tool mechanics) and the built-in verify (run-the-app).

Procedure

digraph pyeye_verify {
    rankdir=TB;
    "Pick scenario (pyeye-scenarios)" [shape=box];
    "Provision / confirm at pinned commit" [shape=box];
    "Run the probe suite" [shape=box];
    "Run the round-trip probe" [shape=box];
    "Diff vs recorded baseline" [shape=box];
    "Divergence?" [shape=diamond];
    "Classify: fixed | regressed | env" [shape=box];
    "Score agent-consumption rubric" [shape=box];
    "Report both verdicts" [shape=box];

    "Pick scenario (pyeye-scenarios)" -> "Provision / confirm at pinned commit";
    "Provision / confirm at pinned commit" -> "Run the probe suite";
    "Run the probe suite" -> "Run the round-trip probe";
    "Run the round-trip probe" -> "Diff vs recorded baseline";
    "Diff vs recorded baseline" -> "Divergence?";
    "Divergence?" -> "Classify: fixed | regressed | env" [label="yes"];
    "Divergence?" -> "Score agent-consumption rubric" [label="no"];
    "Classify: fixed | regressed | env" -> "Score agent-consumption rubric";
    "Score agent-consumption rubric" -> "Report both verdicts";
}
  1. Pick a scenario from [[pyeye-scenarios]] and provision/confirm it at its pinned commit (the scenarios skill has the commands). Note the build under test — the global server, or a delivery worktree's server — because baselines are build-tagged.
  2. Run the probe suite verbatim, with the scenario's project_path.
  3. Run the round-trip probe (section 1b) on the scenario's round-trip module.
  4. Diff vs the recorded baseline (next section). That includes the round-trip counts.
  5. Score the agent-consumption rubric (section after).
  6. Report both verdicts in the format below.

1. Regression check

Compare each probe's live result to the recorded baseline row. Every divergence is signal — classify it, never wave it away:

ClassificationMeaningAction
Fixeda ❌ baseline row (known gap, tied to an issue) now passesUpdate the baseline in [[pyeye-scenarios]], cite the issue/PR that fixed it, note the build
Regresseda ✅ baseline row now failsFile an issue with the probe, expected vs actual, and the build; this is the harness's primary job
Environmentdifference explained by the build under test, not behaviorRecord against that build, tag the build; do NOT treat as a regression

The classic environment case — now a confirmed before/after: the namespace-jaraco edge-scope + cold-start rows returned external / variable / empty on the stale global server (2026-06-21), then flipped to project-scoped handles once the server included the fixes for #423/#444/#454 (2026-06-22). Same probes, different build — expected, not a regression.

A baseline row tied to an open issue is an acceptance check. When that issue's fix lands, this is where you confirm it on real code — exactly what happened for #444/#454 in the namespace-jaraco baseline.

1b. Round-trip probe

The hand-picked probes check single handles. The round-trip probe checks that every handle pyeye emits can be fed back in. It takes every stub from one expand(<module>, "imports") and runs each through inspect and resolve (#616).

uv run python scripts/roundtrip_probe.py <project_path> <round-trip module> 2>/dev/null
# add --json for the full per-handle lists
  • Run it from the checkout whose build you are verifying. It imports that checkout's pyeye directly; the MCP server is not involved. The first run builds the index, so allow about a minute.
  • It reports stubs and the number that don't round-trip, split into inspect miss, resolve miss, disagree (one tool finds it, the other doesn't) and handle mismatch (both find it under different handles).
  • Compare these to the scenario's round-trip baseline row in [[pyeye-scenarios]]. Classify a change like any other divergence: fewer is fixed (cite the PR, usually the #629 consolidation work such as #623), more is regressed (file an issue), and a count that moves with the build is environment.
  • A disagree or handle mismatch entry also fails the rubric's Cross-repo correctness dimension. Cite it there.
  • Don't re-implement this check in pytest. test_emitted_handles_round_trip (#614) is the in-suite sweep over fixtures and src/pyeye, with an allow-list. This probe is its counterpart on real repos at pinned commits. If a scenario reveals a miss that the fixtures don't cover, add a fixture case to #614's test rather than a second copy of the check.

2. Agent-consumption rubric

Correctness is necessary but not sufficient. Score the shape of each response — this is pyeye's defensible axis (honesty + Python-depth). Each dimension is PASS / FAIL with the probe evidence.

DimensionPASS looks likeFAIL looks likeAnchor example (from baselines)
Absence vs zeroempty/measured-none distinct from not-measured; missing edge_counts key ≠ 0a 0 that actually means "didn't measure", or empty conflated with unsupporteddjango inspect omits callers from edge_counts (absent ≠ 0)
No content leakagepointers + structured facts; file:line to Readsource bodies / snippets in the responseevery probe returns handles + locations, no source
Honest limitsreverse-reference edges refuse with a reasonfaked/empty caller data, or grep/legacy-tool fallbackdjango expand(…, "callers") → deferred_reference_backend (#333)
Right altitudecheap by default; truncation flagged (truncated, max_nodes)firehose with no bound; silent truncationjaraco trace(imports) caps at max_nodes with truncated:true
Canonical handlesre-exports collapse to definition sitealias path returned as the identityresolve("django.db.models.Model") → …base.Model
Static-surface honestystatic result not claimed as runtime-exhaustivea static members count sold as the complete runtime surfacedjango members:68 excludes metaclass-injected _meta/objects
Cross-repo correctnessa symbol's scope/handle consistent across entry pointssame symbol labelled differently by different primitives✅ #454: resolve and the imports/callees edges now agree (project) for namespace siblings
Determinismsame query → same result across runsorder/contents vary run-to-runre-run a probe twice and compare

The cross-repo row is a worked example of this rubric earning its keep — it surfaced #454. A namespace sibling that was also pip-installed split scope across primitives: resolve("jaraco.context.ExceptionTrap") reported scope:"project" while the same handle inside expand/trace(...imports) reported scope:"external", because Jedi's follow_imports preferred the installed copy on the environment path. Issue #454 filed and fixed it (build_stub reconciles an edge target's scope to the project definition site when the canonical handle names a project module); the edges and resolve now agree. Kept here as the canonical example of the cross-primitive inconsistency this dimension catches.

Report format

## pyeye verify — <scenario> on <build> (<date>)

**Regressions:** none | <list: probe, expected, actual, classification>
**Round-trip:** <module>: <n>/<stubs> don't round-trip (baseline <m>/<stubs>)
**Fixed since baseline:** <list, with issue/PR>
**Rubric:** <n>/8 PASS
  - <dimension>: PASS/FAIL — <one line of probe evidence>
**Actions:** <baseline updates, issues filed>

Lead with the regression verdict (it's the gate). Keep rubric findings to one evidence line each — point at the probe, don't paste output.

Red flags — you're misusing this skill

  • Running probes without provisioning at the pinned commit — the baseline no longer applies.
  • Treating a build-explained difference as a regression (or vice-versa) — always tag the build under test.
  • Calling the rubric "all pass" without probe evidence per dimension.
  • Pasting source content to "show" a result — that itself fails the no-leakage dimension.
  • Quietly editing a baseline to make a diff disappear — a baseline change must cite the issue/PR and build that justifies it.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.