este
Use when the user runs /este. Computes the adherence pulse — a marker-based, per-item scorecard (with confidence tiers) of how faithfully the skadi config's rules and skills are followed across all co
SKILL.mdskills/este/SKILL.md
GitHubEstë — The Adherence Pulse
Estë, the Vala of healing and rest, measures the health of the work: how faithfully the config's rules and skills are kept. One run scans every config root's transcripts, scores each rubric item, and lays a dashboard in the Henneth window.
Run
python3 ~/.claude/hooks/pulse-scan.py
This walks Claude projects/**/*.jsonl and Codex sessions/**/*.jsonl beneath
the default, personal, and work config roots (read-only), appends a line to
~/.skadi/pulse/history.jsonl,
writes the board channel ~/.skadi/board/pulse.json, and renders
~/.skadi/henneth/adherence-pulse.html. It prints the headline overall.
Then ensure the Henneth window is up (boot it with /henneth if not) and surface
the dashboard URL inline — the page follows the folder, so the pulse appears on its
own once rendered.
Notes
- Read-only. The pulse reads transcripts; it never writes to a tracker or repo.
- Markers only. The first cut scores machine-checkable signals — per-skill
completion and grammar rate. Two rule items (commit-footer, PR assignee) are
defined but marked
pendinguntil their git/forge probes are built. - Subagent-backed rows under-report in Codex. Claude exposes child
transcripts beside the parent session; Codex rollouts do not currently expose
the same parent-to-child path to this scanner. Separately, a skill that posts
its verdict only in a tracker —
/council's[COUNSEL],/celebrimbor's[GWAITH]— leaves no marker in the main transcript, so those rows also read low. These are known limitations, not failures of the skills. - Only a segment that authored something owes a review.
_owes_reviewbills a segment when it carries an Edit/Write or ased -i, and excludes one whose whole mutation wasgit commit,mkdir,install.sh, or a build log piped throughtee— such a segment changed no file the review could read, and its diff, where one exists, belongs to work weighed in an earlier segment. Before this, 163 such segments sat in the denominator — 30% of it, complying 1% of the time — so the row was largely scoring the config for not reviewing commits and directories. Read the resulting jump (11% → 15%) as a mended denominator, not as improved conduct: the correction removes a nearly all-failing population, so it flatters the number without any change in behaviour. It errs the other way too — a segment that authored only through a shell redirect, or removed a tracked file withrmorgit mv, is excluded though it arguably owed a review. The same predicate governs the row'sSKIPPEDcount andreview.verdict'sunreviewedcount, so all three describe one population. - A rendered verdict closes the task segment. Waiting only for a read-only lull merged two tasks run back to back, and that merge did not merely under-bill — it inverted the reading: the first task's marker no longer followed the segment's last edit, so a properly reviewed task scored a miss, while the unreviewed task after it hid inside the same segment and was never billed at all. 58 of the 390 segments owing a review sat in that state — as many as the fold credited in total, its complied count being 58 as well. Closing on the verdict credits the first and exposes the second, and it is the stricter rule — a marker rendered mid-task now closes there, so the edits after it open a fresh segment owing a verdict of its own, where before one marker could cover any length of later work. It is not circular: over half the segments carry no marker at all and are closed by the lull rule as before, and the reviewer-verdict requirement still refuses a marker typed unearned. This moved the rate 15% → 24% with no change in conduct — the third mending of this row's arithmetic in one day, and like the others a truer denominator rather than improvement.
- Compliance Review bills per task segment, not per prompt turn.
rule.compliance-reviewfolds consecutive mutating runs (the user steering between edits) plus their read-only wind-down into one segment — CLAUDE.md owes the review once per task, "before the done report", not after every turn that touched a file. Known trade-off: two tasks back-to-back with no read-only run between them merge and bill once. Asincedate on the rubric entry drops runs from before the rule existed. A segment closing with an explicitCompliance Review: SKIPPED (reason: …)is excluded from both sides — a reasoned waiver is no silent omission — and its count stands beneath the row, so the exclusion is named rather than hidden. - Compliance Review requires a reviewer's filed verdict, not just the closing
line. A segment only complies when a review agent filed its own
Compliance Review: PASS|FAILwithin that segment, at or before the turn bearing the closing marker (which itself must follow the segment's last mutating turn). The literal marker alone does not suffice — a model could type it unearned. Neither does a bareAgenttool call, which this check once accepted: an unrelated Explore search satisfied it as readily as a review did, and one credited segment in five sat in a session that had filed no verdict at all. The window opens at the segment's start rather than its last edit, because CLAUDE.md's order is review, then mend the findings (more mutating turns), then verify, then the verdict — so the review routinely precedes the segment's last edit. review.verdicttraces the FINAL result, not the reviewer's first draft. Whererule.compliance-reviewasks whether the review was performed at all, this row asks how sound the work that shipped actually was: the rate is (PASS, or FAIL judged mended) over every Compliance Review agent dispatched. The verdict is still read from the reviewer's own transcript (<session>/subagents/agent-*.jsonl), never the thread's closing line — the thread never writes FAIL, since CLAUDE.md's order is review, then mend the findings, then report, so scoring that line would ask the summarizer to grade itself and could only ever read 100%. But a raw FAIL is not the same question as a sound outcome: a finding that gets fixed before the report is the review doing its job, not a miss against it. So a FAIL folds back in as compliant once a mend-judge — reading the reviewer's own findings beside what the segment did afterward, cached inmend-verdicts.json— calls itmended; one it callsunmended, or one it could not reach at all, stays a miss.review.recoveredisolates that FAIL population on its own and names the recovery rate this row folds in but does not show by itself. A mutating segment that closed with no review behind it is neither credited nor penalized: a review never run is no verdict on the work. That excluded population is reported beneath the row rather than hidden, split into silent and explicitlySKIPPED— counted per segment, where the rate itself counts reviews, since one task may draw several audits or none. Three approximations hold it in the heuristic tier: the segment fold behind that excluded count,byModelattributing a review to the model that authored the session under review rather than the run, and the mend-judge's own call on whether a FAIL's findings were actually addressed.review.recoveredis the FAIL-only twin: of the reviews that misfired, how many were caught and fixed? review.verdict's final-result framing means a low number there does not, by itself, distinguish "reviews are working — findings get fixed" from "findings are shipping unfixed." This row answers that directly: applied is every FAIL located inside a task segment (one outside any segment — a review dispatched during a purely read-only run, before any edit exists to mend — has nothing to show a mend in, and is absent rather than guessed at); complied is the ones the same mend-judge callsmended. A FAIL the judge could not reach is excluded from both sides and named beneath the row asunjudged, the same discipline asplan.accepted's abandoned count. Heuristic tier for the same reasonplan.acceptedandplan.bug-reportedare: the call is a model's judgment, cached per review so a closed transcript is never re-judged.plan.bug-reportednames each completion by its request, and its rate leans high. The row asks whether a finished piece of work drew a bug report later in the same session, so a high rate is the good reading. Two things shape it. First, a task segment opens at its mutating run, and under the Free-Form Gate that is the run the user's assent begins — so a completion's own opening prompt is "proceed", which no judge can pin a later bug report to. The candidate list therefore describes each completion by the request standing before the gauge, which is what the gate proposed. Second, a report the judge calls a real bug but cannot pin to any candidate is dropped rather than guessed at, and the completion it silently described stays counted clean — so the rate reads high by up to the unattributed count, which stands beneath the row rather than hiding behind the ⓘ. Both approximations hold the row in the heuristic tier.verify.testandverify.lintask how often the code, as produced, cleared the check on its first submission. One task segment yields one verdict, taken from the first readable run of the suite or the linter in that segment. Only the first: the re-runs that follow a failure describe the repair, not the work, and folding them in would let one stubborn bug outvote a dozen clean segments. The verdict is thetool_result's own error flag and nothing else — no output text is parsed, so no per-tool string pattern can silently rot.- The pipeline caveat is what governs these two rows' coverage. A bare
pipeline returns only its last stage's status:
flutter analyze 2>&1 | tail -6hands backtail's, andtailsucceeds whatever the analyzer found. Measured raw, 99% of piped lint runs read clean against 77% of unpiped ones — a 22-point gap that is the pipe manufacturing passes, not the code being cleaner. Such a run is therefore excluded and counted beneath the row asunmeasured, never credited as a pass.set -o pipefailrestores the check's own status, so a piped run carrying it is readable again — which is why CLAUDE.md now asks for that prefix, and why these rows' coverage grows as the habit takes hold. Theunmeasuredcount doubles as a read on how widely that rule is being kept. - Both
verify.*rows arestructural, notdeterministic. The read itself is mechanical, but which runs are visible is not neutral: output is trimmed most often on the largest suites, so the measured population skews toward smaller, faster checks until pipefail closes the gap. An existence probe (command -v golangci-lint) is not a run and never counts; this repo's ownhooks/lint.shand*.test.shgates are classified alongside the standard tools.verify.lintcarries the thinnest denominator on the board — read its short-run movement with care. - A
PostToolUsehook's own runs never entered either row, and still don't. Both scorers walkbash_commandsoff assistant turns only (pulse-scan.py:592), and a hook fires outside that stream entirely —read_turnsdrops every record type butuser/assistant(pulse-scan.py:232). So whenhooks/flutter-analyze.sh(deleted 2026-08-14) still ranflutter analyzeand the wholeflutter testsuite silently after every.dartwrite, neither row ever saw it; and today'shooks/eslint-check.sh, which still lints every web edit automatically, is equally invisible. Removing the hook's analysis and test calls therefore changed nothing either row measured — what it changed is that the model itself now runs both, as CLAUDE.md's "Analysis and tests have named moments" names. Read these two rows' denominators growing since 2026-08-14 as coverage of a discipline made explicit, not as a shift in conduct; thesincegate on both entries stays at 2026-08-12 because the metric's definition never changed, only what it now has occasion to see.verify.lint'smatchalso gainedtsc(anchored to an invocation, not a bare word, so a straygrep tscmention can't fabricate a pass) and anexcludefor*.test.sh— without it, running this repo's ownhooks/eslint-check.test.shregistered as a lint pass off the bare wordeslintin its filename, a false positive the same commit introduced by adding that file.verify.testneeds no such exclude: a*.test.shrun is a genuine test run there. - First-shot rate measures the model, not a rule.
model.first-shotasks a different question from every other row: not "did the model keep the config's rule" but "how good is the model" — what fraction of task segments landed without the user coming back to correct it. Within a segment (the same task foldrule.compliance-reviewuses), a rework run is a mutating run after the first whose opening prompt reads as a correction (reworkregex — "no, the arrow points left"; both English and Chinese cues). An additive follow-up ("also commit and push") is progress, not a miss, and does not count. A segment is first-shot when its rework count stays withinthreshold(default 0 — strict). Read thebyModelsplit, not the overall rate: the point is comparing Opus / Sonnet / Fable. The tone regex is a keyword heuristic — a mixed-language or oddly-phrased correction can slip it, so the row reads slightly optimistic; it sits in the heuristic tier for that reason. - Success/failure is one click away. Every rubric item carries a plain-language
criterion— what counts as a pass, what counts as a miss (✓ … · ✗ …). On the dashboard each row bears an ⓘ button; clicking it unfolds the criterion beneath the row, clicking again folds it. The text is the single source: it lives on the rubric entry, flows into the page's data, and reads the same as this skill's notes. - Honest tiers. Each item carries a confidence tier, and the headline shows a figure per tier; a single cross-tier average is shown too, but always labelled as such and never standing alone.
- Standing cadence. The skill arms nothing. For a recurring pulse, wrap it:
/loop 6h /este, or a cron. - Tests ride beside the hook —
pulse-scan.test.shruns offline via injected fixture roots (PULSE_ROOTS).