feat: add a whole-run spend audit to the postmortem - #133
Conversation
QA drive — evidenceRan after the review loop closed, against the final commit ( ✅ Passed automatedManual-test item —
|
type: wrap-up-report
|
| Pass | Found | Nature |
|---|---|---|
| 1 | 2 Must Fix, 1 Should Fix | a direct contradiction: the untouched "pure spend" line vs. the new never-by-citation rule; and two of three main-loop stats named but not mechanically defined |
| 2 | 2 Must Fix | stats still not uniquely computable — which tool uses count as a "read"; and cache-creation priced against two possible TTL rates (1.25× vs 2×) |
| 3 | 1 Must Fix | the deterministic 5-minute fallback could publish a silently understated cost without disclosing it was a lower bound |
All six findings were fixed across four implementer rounds. The pass-1 Should Fix (verdict boundaries under-defined) was folded in rather than deferred, and pass 3 confirmed the three verdicts now partition cleanly. No Should Fix or Nice to Have survived, so there are no inline PR comments — the loop consumed everything it raised.
QA pass: 1 of 6 manual items executed (the command-shaped sync test, passed), 5 left to the human because each requires invoking /postmortem in an agent session, which a sub-agent may not do. 0 QA findings — the drive surfaced no bug.
One sequencing caveat, flagged deliberately: the pass-3 Must Fix was fixed in d454551 after the review loop hit its cap, so that two-line hunk was checked by the QA verifier against the final state but never by a diff reviewer. It is a disclosure clause with no behavioral reach, and it is called out in the PR's Residual risks so the human reviewer knows which hunk carries less review than the rest.
Human action required
- ⛔ Blocks verification / QA (prerequisite): none.
- ⛔ You must do (deploy / external): none. These are markdown instruction files; they reach consumer repos through each repo's own
update-skillssync PR, which this run deliberately did not trigger (AGENTS.mdBoundaries). - ✅ Done for you (applied in-run): nothing required applying — no schema, env, infra, dependency, or backfill in the diff. The diagram was rendered and hosted on the rolling
qa-assetsprerelease; the artifact bundle was re-uploaded at plan-complete and wrap-up.
The only thing genuinely waiting on you is PR review, plus the five manual items if you want them exercised — and this run is itself a good subject for them.
Residual risks / follow-ups
- The pass-3 fix carries less review than the rest of the diff (see Review outcome). Two lines, disclosure-only.
- Instruction adherence isn't mechanically enforceable — nothing asserts at runtime that a postmortem actually assigned verdicts. The guarantee is the step-2 success criteria, exactly as for every other criterion in the skill.
- Verdict judgment remains subjective at the margin. The three verdicts now partition explicitly and the fully-consumed-but-overpriced case is assigned to
marginal, but boundaries are still judgment. The de-escalation gate means a mistakenwastecannot by itself strip verification —zones.mdstill requires yield evidence, one change at a time. - Watch the audit's own cost. Three review passes each pushed for more determinism in the stat definitions; the final reviewer judged the result "scannable and not excessively bloated," but this is the section to watch if postmortems start feeling heavy. The new audit will measure itself.
- Follow-up (existing path, no new work item): adoption of any audit-driven proposals runs through
/postmortem-loop, as the item specified.
Dial record
zone: 2 # from the item; user-confirmed de-escalation from zones.md's nominal 1
lanes: single-codex # item's explicit review_lanes override
passes: {plan: 1/1, post_pr: 3/3}
findings: # total findings raised (Must Fix in parentheses)
plan: {pass1: {codex: 2 (0), claude: n/a}, later: {codex: 0, claude: n/a}}
post_pr: {pass1: {codex: 3 (2), claude: n/a}, later: {codex: 3 (3), claude: n/a}}
verifiers: {frontend: disabled_by_item, qa_pass: ran}
qa_findings: 0
wall_clock: 0:30 # scripted from session transcript JSONL: 03:21:39Z -> 03:51:05Z
deviations: "escalated 2->1: pass-1 diff review surfaced a direct contradiction between a new instruction and an existing one — a defect class that recurs in re-wording, so the fix rounds earned confirm passes"
pr_size: {files_changed: 3, additions: 58, deletions: 2}
tokens:
codex:
total: 334602
by_role:
implementer: 123548 # 4 dispatches, cumulative: 51060 / +36403 / +23126 / +12959
code_reviewer: 101337 # 3 dispatches: 24656 / 42282 / 34399
backend_verifier: 65880 # 2 dispatches: 30826 / 35054
plan_reviewer: 43837 # 1 dispatch
claude_subagents: 0 # none dispatched — every sub-agent this run ran on Codex
overseer: 10403014 # main session JSONL, message.id-deduped: out 101140,
# cache-read 10089620, cache-write 212115, input 139
total: 10737616
spend_ratio: 178960.3 # 10737616 / (58 + 2)
agents:
- {role: plan-reviewer, model: gpt-5.6-sol, effort: low, dispatches: 1, duration: ~4m, tokens: 43837}
- {role: implementer, model: gpt-5.6-sol, effort: medium, dispatches: 4, duration: ~11m, tokens: 123548}
- {role: code-reviewer, model: gpt-5.6-sol, effort: low, dispatches: 3, duration: ~12m, tokens: 101337}
- {role: backend-verifier, model: gpt-5.6-sol, effort: low, dispatches: 2, duration: ~7m, tokens: 65880}
- {role: overseer, model: claude-opus-5, effort: n/a, dispatches: n/a, duration: 0:30, tokens: 10403014}Note on spend_ratio: the number is real but reads oddly because cache reads are 94% of the overseer's token volume, and a 60-line diff is a tiny denominator. Priced rather than counted, the picture inverts — cache reads bill at 0.1×, so the overseer's cost is dominated by its 101k output tokens, not its 10M cache reads. This is exactly the kind of distortion the change this PR ships exists to surface: a raw-token ratio on a docs change is a misleading efficiency signal, and the new audit's cache-read-volume stat is what would flag it.
tbrownio
added
the
awaiting-human-review
Postmortem — postmortem-spend-audit (ops-only)Postmortem —
|
| Step / dispatch | Start | End | Dur | Tokens | Est. cost | Verdict | Note |
|---|---|---|---|---|---|---|---|
| Step 0 preflight + load | 20:21 | 20:23 | ~2m | overseer | — | useful | bundle + issue fetched, branch state checked; no gaps found |
| Step 1 plan authoring | 20:23 | 20:24 | ~2m | overseer | — | useful | plan.md consumed by every later dispatch |
| plan-reviewer ×1 | 20:24:32 | 20:25:25 | 0m53s | 43,837 | unknown | useful | 0 MF, 2 SF — both consumed into plan.md before implement |
| implementer #1 | 20:26:38 | 20:28:34 | 1m56s | 51,060 | unknown | useful | produced the 35-line base diff |
| backend-verifier #1 | 20:29:20 | 20:29:56 | 0m36s | 30,826 | unknown | marginal | see below — necessary question, fully superseded answer |
| Step 4 PR open | 20:30 | 20:31 | ~1m | overseer | — | useful | PR #133 |
| code-reviewer #1 | 20:32:09 | 20:32:57 | 0m48s | 24,656 | unknown | useful | 2 MF + 1 SF, all three fixed |
| diagram authoring | 20:33 | 20:41 | ~8m | overseer | ~$2.20 | useful | 2 render-fix cycles; embedded in the PR body |
| implementer #2 (fix) | 20:36:59 | 20:39:13 | 2m14s | +36,403 | unknown | useful | resolved both MF + the SF |
| code-reviewer #2 | 20:41:01 | 20:42:19 | 1m18s | 42,282 | unknown | useful | 2 MF, both fixed |
| implementer #3 (fix) | 20:42:53 | 20:44:26 | 1m33s | +23,126 | unknown | useful | 4 lines, both MF closed |
| code-reviewer #3 | 20:44:59 | 20:45:40 | 0m41s | 34,399 | unknown | useful | 1 MF, fixed |
| implementer #4 (fix) | 20:46:09 | 20:46:34 | 0m25s | +12,959 | unknown | useful | 2 lines, closed the last MF |
| backend-verifier #2 (QA) | 20:47:30 | 20:49:08 | 1m38s | 35,054 | unknown | useful | final-state proof + executed the sync manual test |
| Step 6 wrap-up | 20:49 | 20:52 | ~3m | overseer | — | useful | wrap-up, bundle, label, notify |
Aggregates. Dispatch wall-clock totals 11m42s across 10 dispatches — 37% of the run. The remaining ~63% is the overseer's own turns, of which the single largest block is the ~8 minutes of diagram authoring. Summed overseer turnaround between dispatches: ~19.5 min. Human-idle total: 0.1 min.
Spend audit
Production work — judged by downstream consumption. Every production dispatch's output was consumed: all four implementer rounds landed in the shipped diff, the plan was read by every subsequent dispatch, and the diagram is embedded in the PR body. Nothing produced went unconsumed. All useful.
Assurance work — judged by whether its question needed answering at this zone. Nine of ten dispatches are useful by this criterion. The one exception:
backend-verifier #1— marginal. Its question ("do AC1–AC5 hold?") genuinely needed answering before opening the PR — Step 3 requires it and a PR should not open unverified — so it is not waste. But its answer was fully superseded: three fix rounds landed afterward, andbackend-verifier #2re-proved the identical five criteria by identical means. The spend bought roughly 20 minutes of confidence. Note the cause carefully: the supersession was produced by the review loop finding real Must Fixes, not by a pipeline defect. This is the ordinary price of a productive review loop, which is why it ismarginaland notwaste, and why it supports no proposal.
Review-pass yield — the notable number. All three post-PR review passes found a Must Fix (2, 2, 1), and every one was real and fixed. The loop did not converge to zero; it hit its cap with a finding still outstanding. Under the criterion this PR ships, not one review pass was spend to question. This is direct evidence against de-escalating review effort on instruction-text changes, and it retroactively validates the one-notch escalation (2→1) the Overseer took after pass 1.
Main loop — mechanical stats.
- Duplicate reads: 1 path read twice — the rendered diagram PNG, read before and after the fix cycle. Deliberate and correct; 6
Readcalls total across 5 distinct paths. No waste signal. - Top-spend turns (est., 5-minute cache-write rate assumed → lower bound):
$1.154(03:39:28, 19,296 output) and$1.067(03:35:57, 17,480 output) — the two large diagram-JSON authoring turns. Then$0.498,$0.468,$0.437. The two diagram turns are ~11% of estimated main-loop cost. - Cache-read volume: 11,761,480 tokens — 98.1% of input-side tokens (11,985,361 total). At 0.1× input that is ~$11.76 of an estimated $20.02 main-loop lower bound.
The headline spend finding: the overseer main loop is ~97% of the run's total token volume (≈11.99M vs 334,602 across all ten Codex dispatches). Counted in raw tokens the sub-agents look almost free; priced, the ratio narrows sharply because cache reads bill at 0.1×. The wrap-up's spend_ratio: 178960.3 is technically correct and practically misleading — a raw-token ratio over a 60-line docs diff measures conversation length, not efficiency. This is precisely the distortion the change under review exists to surface, and it is the first thing the new cache-read-volume stat flags.
Follow-the-evidence: the two mechanical flags (the diagram turns, the cache-read share) were each inspected in their own transcript window only. No full-transcript semantic sweep was run.
What we asked for
Every postmortem should judge how a run spent its tokens, not just its time — a spend audit over everything that ran, with each timing-table row carrying a useful | marginal | waste verdict, production work judged by downstream consumption and assurance work by necessity-at-zone, plus scripted mechanical main-loop stats. Three additive markdown edits, report-only, repo-agnostic.
Outcome vs intended
Deferred — the human has not reviewed PR #133 yet. This is an operations-only postmortem; the outcome half runs on a later /postmortem invocation.
Mechanically, all five acceptance criteria were independently verified against the final commit, and this very document is the first artifact produced under the new instructions — the spend audit section above was written by following the rules the PR adds.
What to change so it doesn't recur
Two proposals, both from operational findings. Neither is applied.
1. claude/skills/do/SKILL.md — say what happens when the review cap is reached with a Must Fix outstanding
The run hit this case directly and the instructions do not cover it. Step 5 says the cap is reached "with survivors flagged in the wrap-up", which reads naturally for Should Fixes — but a Must Fix is by definition blocking, so "flag it and ship" is wrong, while "keep reviewing" violates the cap. The Overseer had to invent a route: fix it, route the final-state check to the QA verifier instead of a fourth review pass, and flag the reduced review coverage in the PR's Residual risks. That worked, but it was improvised, and a different Overseer would improvise differently.
Proposed edit — in Step 5, after the sentence "A pass with zero Must Fix from every lane ends the loop":
When the cap is reached with a Must Fix still open, the finding is still fixed — a cap limits review passes, never fixes. Land the fix, route its check to the verify stage rather than a further review pass, and record in the PR's Residual risks exactly which hunk shipped without diff review. Never leave a Must Fix unfixed because the cap was reached.
2. references/zones.md — note that instruction-text changes ratchet rather than converge
Three review passes each found a real Must Fix, and each finding was of the same shape: the previous fix was not yet deterministic enough. Reviewing prose that instructs an agent has a different convergence profile from reviewing code — there is no compiler to bottom out the argument, so a reviewer can always demand more precision. The loop terminated on the cap, not on agreement. Zone selection currently reasons about blast radius and reversibility, which correctly rated this change cheap; nothing tells the Overseer that the review on such a change may not converge.
Proposed edit — add to the zone-selection guidance:
Instruction-text changes (skills, agents, references) ratchet rather than converge: each pass can legitimately demand more determinism from prose the previous pass accepted, so the loop may terminate on the cap rather than on agreement. Judge exit on whether the remaining findings are material, not on reaching zero — and treat a pass that found nothing as the real convergence signal.
Explicitly not proposed: nothing about backend-verifier #1's superseded run. Its verdict is marginal, but the cause was a productive review loop, and a rule that skipped pre-PR verification to save that spend would trade real safety for a small saving. This is the case the audit's assurance criterion exists to protect, and the audit protected it.
Dial record & right-sizing
zone: 2 # item's, user-confirmed down from zones.md's nominal 1
effective_zone: 1 # Overseer escalated one notch after review pass 1
lanes: single-codex # item's explicit review_lanes override
passes: {plan: 1/1, post_pr: 3/3}
findings:
plan: {pass1: {codex: 2 (0 MF), claude: n/a}, later: {codex: 0, claude: n/a}}
post_pr: {pass1: {codex: 3 (2 MF), claude: n/a}, later: {codex: 3 (3 MF), claude: n/a}}
verifiers: {frontend: disabled_by_item, qa_pass: ran}
qa_findings: 0
wall_clock: 0:31
deviations: "escalated 2->1 after pass 1 surfaced a direct instruction contradiction"
pr_size: {files_changed: 3, additions: 58, deletions: 2}
tokens:
codex: {total: 334602, by_role: {implementer: 123548, code_reviewer: 101337,
backend_verifier: 65880, plan_reviewer: 43837}}
claude_subagents: 0
overseer: 11985361 # input-side; 98.1% cache reads. Est. $20.02 (lower bound)
total: 12319963
spend_ratio: 205332.7 # raw-token ratio over a 60-line diff — see the caveat above
agents:
- {role: plan-reviewer, model: gpt-5.6-sol, effort: low, dispatches: 1, duration: 0m53s, tokens: 43837}
- {role: implementer, model: gpt-5.6-sol, effort: medium, dispatches: 4, duration: 6m08s, tokens: 123548}
- {role: code-reviewer, model: gpt-5.6-sol, effort: low, dispatches: 3, duration: 2m47s, tokens: 101337}
- {role: backend-verifier, model: gpt-5.6-sol, effort: low, dispatches: 2, duration: 2m14s, tokens: 65880}
- {role: overseer, model: claude-opus-5, effort: n/a, dispatches: n/a, duration: 0:31, tokens: 11985361}Judgment: review effort was right-sized — arguably still underdone. Every one of the three post-PR passes found a real Must Fix, and the loop ended on its cap with a finding outstanding rather than on agreement. Under the criterion this PR ships, no pass was spend to question. The single dial that would have changed the outcome is the pass cap: at the item's captured zone 2 the cap was 1, which would have shipped a diff containing a direct instruction contradiction. The Overseer's one-notch escalation to zone 1 is what bought passes 2 and 3, and both earned their cost.
Acceptance
Still awaiting human review — operations-only postmortem. The PR is labeled awaiting-human-review; commits after that label's timestamp are the post-review rework metric.
System changes
- Postmortem comment on the anchor PR: recorded below on publish.
- Postmortem comment on the work item (issue feat: Postmortem spend audit: token-efficiency review of everything that ran #132): recorded below on publish.
- Approval verdicts on the two proposed changes: pending — an auto-run postmortem never waits for them. Sweep them with
/postmortem-loop.
Postmortems priced every run but only judged the spend of review passes. Research nobody consumed, a verifier re-proving the proven, and a bloated main loop all slipped through. Three additive instruction edits give the audit a computation rule, a place in the report, and success criteria plus a proposal target: - references/run-operations-analysis.md — every timing-table row gets a useful | marginal | waste verdict. Production work is judged by downstream consumption; assurance work by whether its question needed answering at the zone, never by citation, so a correct negative result is not waste. Main-loop coverage is three scripted mechanical stats computed without loading transcript content, with semantic digs limited to flagged windows so the postmortem does not itself become the wasteful run. - claude/skills/postmortem/references/postmortem.md — the body must name wasteful spend across all components and the efficiency changes it supports. - claude/skills/postmortem/SKILL.md — step 2 success criteria require the per-dispatch verdict; step 6 accepts token-efficiency proposals. The postmortem stays report-only, and de-escalation proposals the audit feeds still require yield evidence per zones.md. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Review pass 1 found two blocking defects and one gap: - The dial-record line still asserted flatly that "a pass that found nothing was pure spend", directly contradicting the new rule that assurance work is judged by whether its question needed answering at the zone. Two agents could reach opposite verdicts from identical evidence. Both occurrences are now qualified with the zone-necessity criterion. - Two of the three main-loop stats were named but never mechanically defined, so the audit was not reliably scriptable: path normalization, the token classes constituting turn spend, the output cardinality, and the reported unit are now all specified. - The three verdicts now read as mutually exclusive outcomes with stated boundaries, including the fully-consumed-but-disproportionate case. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Review pass 2 found two stats still not uniquely computable: - "duplicate reads" said which paths to normalize but not which tool uses count as a read, so a script counting only the file-reading tool would disagree with one that also counted search paths or shell-embedded ones. Now scoped to the file-reading tool's own path input. - top-spend turn cost priced cache-creation tokens at "their class-specific rate", but model-prices.md carries two cache-write rates (1.25x at 5m TTL, 2x at 1h). Identical turns could be priced 60% apart, reordering the top five. Now uses the recorded TTL class, defaulting to the 5-minute rate when only an aggregate survives. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The 5-minute cache-write fallback made the cost deterministic but not honest: model-prices.md prices a 1-hour-TTL write at 2x input against 1.25x at 5 minutes, so a turn with substantial aggregate cache creation could be reported with a definitive-looking cost understated by ~60% on its cache-write component — and ranked below a turn that cost more. The total now travels with the assumption that produced it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Three review passes ratcheted this text from an idea into a specification: path normalization, token-class arithmetic, ranking cardinality, TTL fallbacks, lower-bound labelling. Each finding was individually correct and the result was collectively unreadable — 40 lines an agent must wade through under context pressure, in a document whose whole point is that the postmortem must not become the wasteful run. Cut to 16 lines. Every acceptance criterion still holds: the per-row useful | marginal | waste verdict, production judged by downstream consumption, assurance by necessity-at-zone and never by citation, the three main-loop stats computed without loading transcript content, the follow-the-evidence limit, and the zones.md yield gate. What went is the scripting minutiae, which belongs to whoever writes the script — the reference already carries one. Also removed a duplicated requirement in the report template: the whole-run spend instruction appeared in both the Run operations block and the dial record. It now appears once and the dial record refers to it. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The reference script hand-rolled a four-way user-turn classification and computed nothing but span and stalls. It now also derives the three main-loop stats, in roughly the same number of lines: text extraction is shared across both content shapes, and usage/read collection folds into the single parse pass already happening. Reporting is now conditional. The script prints a stat only when it trips a threshold — a file read 3+ times, a turn over ~15% of run output, cache reads over ~95% of input — so a clean run reports one line saying nothing flagged instead of a stat dump nobody acts on. Silence is the signal that there is nothing to dig into, which is what the follow-the-evidence rule wanted in the first place. Using output tokens as the spend proxy for "top turn" also drops the cache-write TTL ambiguity entirely: output dominates cost when cache reads bill at 0.1x, so the extra arithmetic bought precision nobody needed. Verified against two real session transcripts: correctly flagged two genuine duplicate reads while leaving the untripped stats silent. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
8736995 to
eff8439
Compare
1 participant
Summary
What — Every postmortem now judges how the run spent its tokens, not only its wall-clock: a spend audit over everything that ran, giving each timing-table row a
useful | marginal | wasteverdict.Why — The postmortem already priced every run (the per-step timing table carries tokens and est. cost per row), but its judgment of that spend was scoped to review passes alone — "a pass that found nothing was pure spend." Nothing swept the rest: research dispatches, planning, implementer retries, verifiers, and the overseer's own turns. A research report nobody consumed, or a verifier re-proving the proven, slipped through unremarked, and there was no home for token-efficiency improvement proposals.
How — Additive instruction edits to the three surfaces where the existing spend judgment already lives:
references/run-operations-analysis.mdclaude/skills/postmortem/references/postmortem.mdclaude/skills/postmortem/SKILL.mdTwo design choices carry the weight:
Done means:a postmortem run after this change gives every dispatch and pipeline step auseful | marginal | wasteverdict with evidence, computes the three main-loop stats, and names the token-efficiency changes the wasteful spend supports — while staying report-only and never becoming a gate.Visual overview
Before, six kinds of timing-table row reach the postmortem and only one — review passes — passes through any spend judgment at all; the other five flow straight through unjudged, and the main loop is never measured. After, the same rows all pass through the audit, the main loop arrives with three scripted stats, and every row leaves with a verdict. The gate didn't move — it widened.
Note the criterion split at the bottom of the diagram: it is the reason this isn't simply "no citation = waste." Assurance work leaves no citation by design, so that rule would have condemned exactly the verification the pipeline depends on.
No screenshots: this change has no user-visible surface — the only consumers of these files are agents reading instructions.
.excalidrawsource in./tmp/postmortem-spend-audit/refs/.User journeys
Linear, single-actor, no branches. The postmortem author's journey: an agent runs
/postmortemon a completed/dorun → step 2 scripts the transcripts as before, and now also assigns each timing-table row auseful | marginal | wasteverdict and computes the three main-loop stats → the Run operations section of the written report names the wasteful and marginal spend with its evidence → step 6 proposes the token-efficiency changes that spend supports, alongside the stall-removing operational change. Nothing is applied; nothing gates.There is no runtime and no user-facing surface, so there is no branch to map.
Verification
Verified by inspection by independent Codex
backend-verifierdispatches (docs-only change — no executable surface; none of the six rubrics in.references/rubrics/covers a docs change, so the ACs alone gate). The final-state verification ran after the last fix landed; full quoted evidence in the QA comment below.references/run-operations-analysis.md:87-96carries all four elements: oneuseful | marginal | wasteverdict per row; production judged by downstream consumption — "name what consumed the output, or that nothing did"; assurance judged by whether its question "needed answering at this zone, never by citation"; and "A de-escalation proposal this feeds still needs yield evidence per.references/zones.md, one change at a time.":97-105— "script three stats from the already-parsed JSONL without loading transcript content — duplicate file reads, top-spend turns, and cache-read volume"; "Report them only when one trips a threshold… a clean run gets one line saying nothing flagged, never a stat dump"; then "follow the evidence — judge main-loop behavior only inside a window a stat flagged, never by sweeping the transcript." The reference script at:106-148implements all three and prints only the tripped ones.claude/skills/postmortem/references/postmortem.md:40-41— "name the wasteful and marginal spend across ALL components — dispatches, steps, and the main loop, not just review passes — and the token-efficiency change(s) it supports."claude/skills/postmortem/SKILL.md:80-81— "Every dispatch and pipeline step carries auseful | marginal | wasteverdict with its evidence, and the main-loop mechanical stats are computed.":157— "Token-efficiency findings from the spend audit are equally valid proposal targets.".references/zones.md), and no consumer-specific name, path, or ID appears. Thedcouple/orchestramention inSKILL.mdis pre-existing, not added here.Substance —
3 files changed, 71 insertions(+), 37 deletions(-)— the deletions are the old reference script and the over-specified first draft, both replaced. Two deletions are deliberate narrowings: both pre-existing "a pass that found nothing was pure spend" assertions were narrowed to the zone-necessity criterion, because leaving them intact created a direct contradiction with the new rule (see review history below). The verifier confirmed both amendments keep the overdone/right-sized/underdone judgment, the dial selection, the token-per-pass signal, and the zones-table tuning intact.git diff --checkclean.Report-only confirmed —
SKILL.md:28-30("This skill changes nothing… presented for the human to approve, not applied") and:159-165("Do not apply any of them"; "this step is a report, never a gate") verified uncontradicted by the new text.Gates cleared — plan review 1/1: 0 Must Fix, Approve (2 Should Fixes on plan repo-truth accuracy, folded in and independently re-verified). Diff review ran 3 passes after a one-notch zone escalation (2→1) taken because pass 1 surfaced a direct instruction contradiction; passes 1–3 found 2, 2, and 1 Must Fix respectively, all fixed.
Then the text was cut roughly in half (the spend-audit bullet went 40 lines → 17) — see the last three commits. Each review finding was individually correct, but together they had ratcheted a rule into a specification: path normalization, token-class arithmetic, ranking cardinality, TTL fallbacks, lower-bound labelling. That is a real defect in a document whose own thesis is that the postmortem must not become the wasteful run. The scripting minutiae came out; every AC element stayed, re-checked line by line above. A duplicated requirement in the report template (the whole-run instruction appeared in both the Run operations block and the dial record) also went. Build gate: not applicable — markdown only, no CI workflows in this repo,
daemon/untouched.Manual tests
The real proof is running the changed instructions and seeing the audit appear. Highest-value first:
Important (behavior of the instructions):
/postmortemon any completed/dorun → the written Run operations section assigns every dispatch and pipeline step auseful | marginal | wasteverdict, each with its evidence. — left to human: requires invoking the postmortem skill in an agent session, which a sub-agent may not doNice (hygiene):
scripts/sync.sh <consumer-repo>still syncs the three edited files cleanly, and the synced text contains no consumer-specific name or path.Areas not affected: the timing table's existing columns, the dial-record contract,
references/zones.md, the wrap-up format, the timeline HTML template and its render script, thedaemon/package, and every other skill.A note for whoever runs the five human items: this PR's own
/dorun is itself a good subject — it has research, planning, four implementer rounds, three review passes, and two verifier passes, including assurance work that correctly found things and a final pass whose finding was real. The ops-only postmortem posted on this PR is the first artifact produced under the new instructions.QA results
1 of 6 manual items executed, 5 left to the human. The one command-shaped item passed:
scripts/sync.shinto a throwawaygit inittarget synced all three edited files byte-identical, a second run produced zero further change (idempotent as documented), and the temp target was deleted. The five remaining items each require invoking/postmortemin an agent session — something a sub-agent is forbidden to do — so they are genuinely human work, not skipped work. All five acceptance criteria were re-proven against the then-final commit, and re-checked by hand after each of the three simplification commits that followed. No bug was found by the QA pass. Proof in the comment below.Deploy notes
None. The deploy-notes scan over the run's diff found no schema/migration, env var/secret, infra/CI, new dependency, or one-time script/backfill. These are markdown instruction files consumed by agents; they take effect for consumer repos on their next
update-skillssync, which is that repo's own PR flow and is deliberately not triggered here (perAGENTS.mdBoundaries).Residual risks
d454551and re-verified by the QA verifier against the final state, but it was not seen by a diff reviewer. It is a two-line disclosure clause with no behavioral reach; flagging it so the human reviewer knows exactly which hunk carries less review than the rest.marginal— but the boundaries are still judgment. The de-escalation gate means a mistakenwasteverdict cannot by itself strip verification:zones.mdstill requires yield evidence, one change at a time.Closes #132