You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
One comment per event, as it happens: its date, a duration or a link. Nothing here comes from outside this repository except the maintainer's decisions, recorded as such.
What is measured
Criterion
Target
Tech lead — a compliant repository (from uv tool install to nstack doctor with no gap)
≤ 30 min
Tech lead — framed, first cycle started
≤ 2 h
Developer joining — sync events (a question the repository could not answer)
0, each named
AI agent — drift probes refused as expected
10 of 10, P4 observed
Second team — merged using only the contract and the repository
0 sync events, 0 waits
Evidence — review time per sheet pull request
≤ 15 min, median
Cycle — every delivery pull request tied to a deliverable or out-of-cycle; no extension
Task 1 — repository, teams, App (2026-09-17): the public repository created empty; teams catalog and loans created, write access given to maintainers, catalog and loans (the maintainer's statement: the App cannot read team permissions); the repository added to the napkinstack-agent installation — verified: the installation lists engineering-os and tool-library; allowed actions include astral-sh/setup-uv — verified by the first CI run.
16:49:59 the README's sentence, committed by the maintainer.
16:52 before the push: 2 commits, authors and committers napkinstack-admin only; gitleaks over the history: no leak. main pushed by the agent under its App.
16:53 Governance on the push: Hooks and secrets and Fitness functions green; Modules green. PR scope and review budget runs on pull requests only.
Next: the checklist of the README's "To do once on GitHub", then nstack doctor.
Task 2 — first nstack doctor, run by the maintainer with a short-lived token of their own (Administration: read), before the checklist:
Workstation L1–L6: all OK.
GitHub: 8 gaps, 0 not verified — G1–G4 and G12 (no ruleset yet), G5–G6 (Advanced Security), G11 (labels); G7–G10 OK, set at the organisation level. Every gap names its place in the settings and the action.
Finding F1 (minor, message): with no ruleset on main, G12 fails with "remove every bypass actor" — there is no actor to remove; the action that applies is G1's, "create the ruleset". Source: _no_bypass in doctor.py returns false when no ruleset applies, and the message is G12's only one. P6 asks for the action that fits the case.
Task 2 — the checklist, G11: the three labels cross-module, over-budget and out-of-cycle were created by the agent under its App, at the maintainer's request, with the colours and descriptions engineering-os uses. Counted as help from the agent: a tech lead alone would have created them by hand in Issues → Labels, three forms.
Finding F2 (in NapkinStack's own repository):engineering-os has cross-module and over-budget but notout-of-cycle, required by G11 since M8d — its own doctor would report the gap. To fix there, not here.
Finding F3 (friction at setup, G4): the checklist asks the tech lead to require four checks in the ruleset, but GitHub only offers a check it has already seen. Test sheet and cycle and PR scope and review budget run on pull requests only, so on a repository that has none they cannot be selected — the maintainer hit exactly that. What unblocked it: opening a first pull request (#2) so the checks run once. The README and doctor's G4 action say nothing about it.
Task 2 — the first pull request (#2): README: what this repository is for, by the agent under its App, assigned to the maintainer. The four checks ran and passed: Fitness functions, Hooks and secrets, PR scope and review budget, Test sheet and cycle, plus the Modules workflow (no module touched).
Also observed, in a real project this time: a pull request that changes no module beyond its description asks for no test sheet and no cycle — NapkinStack v0.3.1's D24 fix behaving as intended.
Task 2 — done.nstack doctor, run by the maintainer with their own short-lived token: 0 gap, 0 not verified. Verified from the agent's side, through the App, which has no Administration permission: the rules applying to main require a pull request, 1 approval, a code owner review, no force-push, no deletion, and the four checks Fitness functions, PR scope and review budget, Hooks and secrets, Test sheet and cycle. The bypass list and Advanced Security are readable only with the maintainer's token: their doctor run is the evidence.
Duration — PDR-0001's criterion (a compliant repository in under 30 minutes): met.
Measure
Value
Start (date, before uv tool install)
16:45
End (date, after the last doctor)
21:20
Effective time, breaks removed, as the maintainer counts it
~25 min
The wall clock is not the working time: the session was interrupted several times. The figure that counts is the maintainer's own, 25 minutes.
Conditions, so the figure is not read for more than it is worth:
The agent did two things a tech lead alone would have done by hand: creating the three labels (G11), and opening the first pull request to make the checks appear (F3).
The maintainer already knew the checklist, having written it — a newcomer would read more.
The organisation's settings (G7–G10) were already in place from engineering-os: a first project in a fresh organisation would add that work.
The first pull request (#2) is merged, approved by the maintainer, merged by the agent under its App: the barrier held from day one — the pull request stayed BLOCKED until the approval.
The decider wrote the idea outside the repository, in French. Decision: the project's language is English; the idea was translated faithfully and the translation is what nstack discover kept, in docs/project/inputs/, with a line saying it is a translation. The French original stays outside the repository.
nstack discover on branch discovery/tool-library: the idea kept, docs/project/discovery.md created, and the command printed what to do next in the agent. It called no model — as PDR-0002's extension requires.
The framer's brief is published before the session starts: issue Discovery — the framer's brief (stages 1 to 4) #3. It is the whole of what that session receives; anything else it must find in this repository. The decider's answers will be posted there too, verbatim.
The framer runs stages 1 to 4 only. Stage 5 goes to a session that has not written them.
Task 3 — discovery: done, decided go (pull request #5, merged 2026-09-18 by the agent under its App, after the decider's approval).
What
Result
Sessions
Three, isolated: framer (stages 1–4, brief in #3), challenger (stage 5, brief in #4, did not write 1–4), decider (stage 6). Every brief published before its session started; the decider's answers posted verbatim
Rounds
One. No clarify
Stage 2
19 sourced rows; four pages unreachable, recorded as such
Stage 5
10 pre-mortem endings, 4 risks with a cheap test and a proposed threshold each, 11 counter-evidence rows, 13 flaws each quoting the line it objects to
Decision
go — an acknowledged rehearsal, criterion A only; criterion B (twenty real loans) dropped, not deferred; no test kept as a spike
Checks
nstack fitness and nstack plan pass; C7 mutation-tested on a throwaway copy: a go with no decided_on fails, a go with no challenger fails
What the framework earned here. The challenger found in one session what stage 2 had missed and what decides the product: entrevoisins.eu ships the same product free in French today; smiile.com, the French reference for free neighbourhood mutual aid, now serves a domain-sale page after 500,000 neighbours; Mutum was liquidated with 113,000 objects listed. It also showed that the success criterion as written could not be passed or failed and had no owner — the ending "framework declared, product never measured" was the likeliest one on the document's own evidence. The decider removed the criterion rather than arrange that ending. The rule that produced this is PDR-0002's extension — a challenger in another session — and it paid for itself on the first use.
What it leaves for the framing, and criterion A depends on it: with no deployment and no real user, what counts as a verified deliverable? PDR-0003 asks for test sheets run by a verifier, with evidence; that evidence can no longer come from a neighbourhood. If the charter does not settle it, a cycle can close green while proving only that the paperwork moved — the same structure the challenger found on criterion B, displaced onto criterion A. Eight places in sections 1 to 5 still speak of the dropped criterion B; those sections stay as the round-1 record, so the charter has to settle them.
Sync events so far: 0. The framer and the challenger asked nothing outside this repository; the decider's answers went through the orchestrating session and are posted in #3 and #4.
Task 4 — framing: the brief is published, the session is not started
2026-09-18. The framer's brief is issue #6, published before the session that will follow it exists — the same isolation rule the discovery used, where it proved its worth.
The timer has not started. It starts when the framer session opens and stops at the merge of the accepted charter and cycle, per the plan. The decider is not available for the interview yet, and starting the session without them would spend the two-hour budget on waiting.
What the brief fixes, so that it cannot drift once the session is running:
the question the framing must settle — with no deployment and no real user, what counts as a verified deliverable — put to the decider rather than invented;
what the discovery already decided and the framer may not reopen: criterion B dropped rather than deferred, the product objections accepted as true, no test kept as a spike, sections 1 to 5 left as the round-1 record;
the constraints this validation requires of the charter and cycle 1: two modules with their owners, a versioned contract provided by one and consumed by the other, a user-facing deliverable each, an appetite of one week, the second deliverable ready only after the first is merged, and a tooling profile that runs on GitHub-hosted runners.
Recorded for PDR-0004, accepted the same day: from this task on, every question put to the decider is logged with the stage of the project, the field it comes from, who it was put to, and whether a check could have answered it. The framer is not told about this. Its brief says nothing of it, so that observing the interview does not change it; the orchestrating session records from the outside, since every question passes through it anyway.
2026-09-18, 10:34 UTC. The framer session is open. It is a fresh session: it carries nothing of the orchestrating session's conversation, and receives only issue #6, published before it existed.
The timer runs from this moment to the merge of the accepted charter and cycle. Budget: two hours (PDR-0002).
What the session was given beyond its brief, recorded so the measurement can be read honestly: the plumbing of this validation, not of the framework — the bot's commit identity, the push form that works under the sandbox, the gh-agent wrapper, the pinned engine version, and the hosts each command needs. These are artefacts of how this agent is isolated (ADR-0004) and of the workstation's sandbox. A question about any of them would be noise in the sync-event count, not a finding about the framework.
What it was not given: anything about the project, the decision, the products the challenger found, or what the decider wants. All of it has to come from the repository or from a question.
Known gap, recorded before it can be claimed as a success. This repository pins _commit: v0.3.1, so its CI installs napkinstack==0.3.1. Rules H1 (D27) and the decider check (D28) were merged after that release and are not in it: a charter keeping the template's decider: "@<github-handle>" placeholder would pass this repository's CI today. The orchestrating session runs the newer checks against the branch as a second gate. This is what Task 10 will exercise for real.
Task 4 — batch 1: six sync events, and the first PDR-0004 measurement
2026-09-18, 10:41 UTC (7 minutes in). The framer read the brief, AGENTS.md, both playbooks, all 629 lines of the discovery, the templates, seven chapters of the handbook, the three workflows and the engine's own check code at v0.3.1 before asking anything. It wrote nothing and created no branch until its questions were answered. Questions and answers are on issue #6.
How the answers were produced, recorded so this can be read honestly later: the decider asked the orchestrating session, acting as product expert, to draft them; the decider ratified them and amended one. They are the decider's decisions. The six questions are genuine sync events either way — no session could have answered them from the repository — and are counted as such.
Sync events: 6
#
Question
Stage of the project
Field it comes from
Put to
Could the repository have answered it?
Q1
What counts as a verified deliverable, with no deployment
A charter, after a discovery whose evidence criterion was dropped
Verification
Decider
No. The handbook's verification rules assume a product someone uses; nothing covers the case
Q2
Which event ends the project
A charter being written
Product
Decider
No. The discovery gives criterion A's content, not the event that closes it
Q3
The stack, and the tool behind e2e
A charter with no tooling profile filled
Architecture and platform
Decider
Partly. Every constraint is in the repository — allowed actions, uv as the only prerequisite, phone-first, near-zero budget. The repository states outright that it supplies no stack, so the choice is the decider's
Q4
The criticality of the two modules
Before the modules are scaffolded
Governance and security
Decider
No — and worse: it answers twice. One rule raises criticality for a domain handling personal data; another forbids imposing critical ceremony on a small module. The framer had to arbitrate with no guidance
Q5
GDPR: publish a contact and answer a deletion request?
A charter carrying an open question
Legal
Decider
No. It is about the decider's own willingness
Q6
When cycle 1 starts
A cycle being written
Process
Decider
Yes, by convention. The playbook already rules that dates move in the acceptance pull request; "start is the day the cycle is accepted" would have removed the question entirely
What this first sample says. Of six questions, one (Q6) a convention already in the repository could have removed, and one (Q4) the repository answered twice in opposite directions — a category PDR-0004 had not anticipated: not a question the framework should have asked, but a contradiction it should not have left standing. The remaining four are genuine decisions that belong to a human, and every one of them arrives at a stage that predicts it: verification and product at the charter, architecture at the tooling profile, governance before scaffolding.
Defect candidate, for Task 9
The handbook contradicts itself on criticality: playbooks/framing.md raises it for a domain handling personal data, docs/os/07-governance.md §6 forbids the ceremony of a critical system on a small module. Both apply here and they point opposite ways. The framer resolved it correctly — the personal-data trigger is about real personal data, and nothing here is deployed — but it resolved it, rather than reading it.
Deviation from the M9 plan, recorded
The plan requires "team 2's first deliverable ready only once team 1's first is merged". The framer applied it faithfully and gated D3 (loans) on D2 (catalog's user-facing deliverable). The decider amended the constraint:D3 gates on D1, the versioned contract, instead.
The reason is the plan's own purpose. Gating the consumer on the producer's interface contradicts what cycle 1 exists to demonstrate — that two modules know each other only through their contract — and it concentrates the risk: if D2 slips inside a one-week appetite, D3 never starts and the two-team observation, the most valuable thing this validation can produce, is lost entirely. The dependency and the sequencing are still real; they now run through the contract, which is where the framework claims they should.
Task 4 — the clock, recorded honestly while it runs
2026-09-18, 11:46 UTC — 72 minutes since the framer session opened. The pull request (#7) has been waiting for the decider since 11:05. The plan measures the framing as wall clock from the session's start to the merge, so that wait is inside the measurement. It is recorded here rather than explained away at the end, and it follows the precedent of Task 2, where effective work and wall clock were logged separately with their conditions.
Framer, reading and first batch of questions
10:34 → 10:41 (7 min)
Decider's answers drafted, ratified, published
10:41 → 10:53 (12 min)
Framer writing both documents, checks, pull request
10:53 → 11:05 (12 min)
Orchestrator verifying the deliverable independently
11:05 → 11:15 (10 min)
Effective framing work
≈ 41 min
Design discussion unrelated to this framing (a PDR-0004 constraint)
11:15 → 11:35
Waiting for the decider, who is away from the workstation
11:35 → ongoing
What this says about PDR-0002's criterion, and it is not a detail. The criterion is "the framing in one session of two hours". Effective work is well inside it. Wall clock will not be, and that is not a defect of the framing — it is the criterion measuring the decider's availability as if it were the framework's cost. A tech lead evaluating NapkinStack will read "two hours" as two hours of my attention, not two hours of calendar. Candidate finding for Task 9: PDR-0002's criterion should be stated in effective work, with the wall clock recorded beside it. Task 2 already hit this and worked around it by logging both; twice is a pattern, not an accident.
A second observation, for PDR-0004. Part of the wall clock above is a design conversation that had nothing to do with framing this project. The orchestrating session could not have known that in advance, and neither could any timer — which is a reminder that a stage tells you what to ask, never how long a human will take to answer. Nothing in PDR-0004 should ever try to measure the second.
Task 4 — done. The framing is merged, and PDR-0002's criterion is missed on the measure the plan chose
Merged 2026-09-18 at 13:02 UTC — pull request #7, 73bb72b. The charter and cycle 1 are accepted. The timer ran from the framer session opening at 10:34 to that merge.
Wall clock, which PDR-0002's criterion measures
148 min — over the two-hour budget
Effective work — framer, answers, writing, verification, acceptance
≈ 53 min
Design discussion unrelated to this framing
≈ 20 min
Waiting for the decider, away from the workstation
≈ 75 min
The criterion is missed as written, and met on what it was trying to measure. This is the second time: Task 2 hit the same thing and logged both numbers with its conditions. Twice is a pattern. A tech lead reading "the framing in one session of two hours" understands two hours of my attention, not two hours of calendar — and calendar time is mostly the decider's availability, which is not a cost of the framework. Finding for Task 9: state the criterion in effective work, and record wall clock beside it.
What the framing produced
Three deliverables, one week, 2026-09-19 to 2026-09-26. D1 the versioned contract and the double built from its schema, D2 and D3 one user-facing outcome per module, catalog and loans, both standard and user_facing. D3 gates on D1 — the decider's amendment to this plan, so that the consumer waits on the contract rather than on the producer's interface.
The question the framing existed to settle is answered as charter constraint C1, in four clauses: the sheet is the criteria; the verifier is a separate session and no scenario may be human only while nothing is deployed; the evidence is machine-produced and re-runnable, never a screenshot pasted by hand; and the criteria are fixed at acceptance, because the sheet is the criteria means nothing if the criteria can move.
Six findings, from one framing session
PDR-0002's criterion measures wall clock, and therefore the decider's availability (above).
contracts/ and platform/ are modules for one check and invisible to two others.MODULE_DIRS includes them; module-checks.yml greps ^modules/ and pr_scope.sh greps ^(modules|services|apps|packages)/. A pull request touching contracts/ alone requires a test sheet and can produce no .evidence/ artefact — it would have no admissible evidence under C1. Found by the framer, verified by the orchestrating session.
The handbook contradicts itself on criticality.playbooks/framing.md raises it for a domain handling personal data; docs/os/07-governance.md §6 forbids the ceremony of a critical system on a small module. Both applied; the framer had to arbitrate with no guidance.
playbooks/framing.md says nothing about the technical framing. Zero occurrences of ADR, tooling, stack or architecture. Module decomposition is routed to docs/os/02-modules.md §3; nothing else is. The M9 plan had to require the tooling profile by hand, because the playbook does not.
The framing created a contradiction it could not avoid. The stack landed in the charter's constraints, while docs/tooling-profile.md states it is "the only place in the repository where a tool name appears". The framer's bounds forbade touching that file. The charter should point at the profile rather than duplicate it, and nothing says so.
A structuring decision can land with no ADR.nstack new-module prints a reminder to write a creation ADR; nothing requires one. Two modules and a versioned contract are exactly what docs/os/06-decisions.md submits to the Prior Art Gate.
Findings 4, 5 and 6 are the same defect seen three times: technical decisions taken during the framing have no artefact to land in, so they land in the product document. playbooks/framing.md is not being corrected now — it is the thing under measurement, and changing it mid-validation would destroy the measurement.
Finding — a human approval survives a force-push, and nothing in the checklist refuses it
Observed 2026-09-18, on pull request #9 of this repository.
The decider approved commit f40ae95. The branch then had to be rebased onto main — nstack new-module appends a line to .github/CODEOWNERS, so the two module-creation pull requests conflicted on that one file. After the rebase the head became 2859e7e, a commit the decider has never seen.
GitHub still reports, on the current head:
reviewDecision : APPROVED
mergeStateStatus : CLEAN
approved on : f40ae95 (the commit before the rebase)
The merge was refused by the orchestrating session, not by the forge. The only thing standing between an approval and content the approver never read is --match-head-commit, passed by the agent to its own merge command. That is the agent's discipline, not an enforced rule — and the agent is the party that pushed the new commits.
This matters because it is the one barrier ADR-0004 rests on. The decision took care that an agent can never approve; it did not take care that an agent can change what was approved after the fact. The barrier is weaker than the ADR claims, and the gap is closed today only by a convention the ADR does not name.
The setting exists on GitHub — "Dismiss stale pull request approvals when new commits are pushed" — and no rule of the checklist requires it: doctor runs G1 to G12, covering the pull request requirement, the approval count, code owner review, required checks, secret protection, vulnerability reporting, allowed actions, SHA pinning, workflow approval, workflow token, labels and the empty bypass list. Stale approvals are not among them.
Candidate G13, and a candidate defect against ADR-0004. Both for Task 9.
Second finding, from the same event: two teams conflict on their first commit
nstack new-module writes its owner line into .github/CODEOWNERS, at the same place for every module. Two teams creating their module in parallel therefore conflict before a single line of code exists — on the file that decides who approves what.
The conflict is trivial and resolves in seconds. What it costs is the promise: several teams work in parallel with no synchronisation meeting. The very first action of the second team needs the first team's work to be merged. A file per module, or an ordered insertion, would remove it.
The stale-approval hole is closed on both repositories — the checklist gap is not
2026-09-18. The maintainer enabled "Dismiss stale pull request approvals when new commits are pushed". Verified from the agent's session, through the API, on both rulesets:
The barrier ADR-0004 rests on is now held by the forge. Replayed against today's event: a rebase after approval would dismiss it, the pull request would return to REVIEW_REQUIRED, and the agent — which cannot approve — could not merge what it had just pushed.
require_last_push_approval was left false, deliberately. It was first proposed as complementary; it is redundant once approvals are dismissed on push. It would only cover a case this flow does not have: an approver pushing after their own approval.
What is not fixed, and stays a finding for Task 9: nothing in the framework requires this. nstack doctor runs G1 to G12 and none of them reads this setting, so a project generated tomorrow ships with the hole open and doctor reports it compliant. Candidate G13. The instance is closed; the defect is not.
Minor, same audit
The two repositories disagree on merge methods: engineering-os allows squash only, tool-library allows merge, squash and rebase. The flow squash-merges everywhere. Not a hole, but a divergence between the reference repository and the project it generates, and the checklist does not cover it either.
Merged 2026-09-18, pull request #11 → 08fb6cf. Deliverable D1: contracts/catalog-api/v1/openapi.yaml, hand-written, and a double built from that document alone.
What the plan asks to be measured
Sync events
0. The author session asked no question before writing: "nothing in the brief turned out to be ambiguous once the cycle, the charter, the module's AGENTS.md/MANIFEST.yaml and the engine's own rules were read." Neither verifier was blocked. The definition applied is M9's — a question a session could not answer from the repository and had to put to the maintainer. The budget overrun and the two findings reached the decider through the pull request, which is the review channel, not a synchronisation
Maintainer's review
10 min, beyond the verifier: none. PDR-0003 targets a median under 15 minutes; this pull request carried 706 changed lines
CI failures, and whether each was legitimate
2 failures, both legitimate, both the framework working.T2 no verifier named on the author's commit — by design, the author leaves the sheet empty. T4 scenarios verified on another commit than the head after the fix — the evidence was produced on the previous head. No spurious failure
Verification
Two sessions, neither having written what it judged. The second re-ran everything on the new head rather than trusting the first
What the barrier caught that a green sheet had not
The sheet was green and complete when the first verifier finished. The review that followed still found two real defects:
The documented way to start the double did not work — No module named contract_double, because the module declares no [build-system] and the scenarios passed PYTHONPATH to their subprocess themselves. It mattered because D3's premise is that loans builds against this double, and the README is where that session will look.
Two claims about the double were false: "point it at any OpenAPI document and it serves that one" and "it answers 501 rather than a wrong answer". Against a document declaring a non-string path parameter, it returns the declared 400 — a wrong answer, silently.
The decider's call: fix the first, and for the second withdraw the claims rather than build coercion — a promise known to be false costs more than a limitation that is declared. The author's fix removed the trap instead of documenting around it: the command in the README is now the one the scenarios themselves run, so the two cannot drift apart unnoticed.
Proof of posture, and it is mechanical rather than stated
After the fix, the sheet still carried 729cf67 in its three commit cells while the head was bd1227f, and T4 was red for exactly that. Had the author updated those cells, T4 would have gone green and the verification would have been a lie. Its red is the evidence that the author stayed inside their posture — no human had to check.
Framework findings from this task
T2 cannot tell a verifier from an author. The Verifier: line is free text; the check only refuses a placeholder. Its own message says "someone other than the author of the change" and nothing enforces it. A sheet filled by the authoring session would pass identically. Charter C1 clause 2 rests on convention alone. Same shape as the stale-approval hole: the barrier is a habit, not a constraint.
provides is never confronted with reality.boundaries.py never reads provides; nothing checks that provides[].path exists or holds a document. The rule that kept it empty until D1 was held by hand alone.
nstack pr-check and nstack pr-scope count modules differently and print contradictory lines in the same CI run on the same commit.
nstack e2e has no way to say a scenario is about the module's own envelope. Noted for what it is: not a defect, an absence observed while writing the first real sheet.
Task 7 — the drift probes: 6 of 8 refused, 2 accepted what should have been refused
P1 to P8 run on 2026-09-19, one pull request each, none merged. P9 and P10 are the maintainer's, P11 belongs to Task 8.
Deviation from the plan, recorded: the plan asks for an author session per probe. They were written by the orchestrating session instead. Eight sessions for eight deliberate one-line violations is disproportionate, and what is measured — does the check refuse? — does not depend on who wrote the line. Every probe was also reproduced locally on the project's own pinned engine, so each result below is a rule identifier, not a red cross.
#
Probe
Expected
Observed
P1
One change to two modules, no cross-module label
P1
✅ PR scope and review budget refuses
P2
loans imports catalog's implementation
B2
❌ nstack fitness exits 0. Accepted.
P3
loans uses the contract with consumes emptied
B1
❌ Every check green. Accepted.
P4
v1 broken in place, producer edited to match
to observe
⚠️ No fitness rule. Caught only by the module's own e2e
P5
A user-facing change with no test sheet
T1
✅ FAIL [T1] Test sheet missing
P6
A scenario passed with no evidence
T3
✅ FAIL [T3] passed without evidence and the commit verified
P7
A delivery pull request naming no deliverable
K3
✅ FAIL [K3] no deliverable named
P8
out-of-cycle with no justification
K4
✅ FAIL [K4] label out-of-cycle without its justification
The two failures are one defect, and it is the most serious this validation has produced
The boundary rules are blind to a contract-based dependency — which is the only dependency this framework allows between modules.
boundaries.py builds the real graph from import lines. A module that consumes an HTTP contract imports nothing of its producer, so the graph stays empty. Three consequences, all now observed rather than argued:
B1 cannot see an undeclared dependency (P3). Emptying consumes while the module reads the contract is accepted by every check. The declared graph can lie freely.
B2 is walked around by a path manipulation (P2). from catalog.listing import Tool preceded by sys.path.insert(0, "../catalog/src") is accepted: the marker B2 looks for (/src/) is on a line that carries no import keyword, so the detector never reads it.
B4 warns on a real dependency — already logged during D3, same root cause.
docs/os/02-modules.md and the fitness functions promise that the declared graph is compared with the real one. For the boundary this product exists to enforce, that comparison does not happen. Every NapkinStack project is affected, not this one.
P4 — what catches a broken contract, and what does not
Renaming a required field of v1 in place, with the producer edited to match, is caught by no fitness rule. nstack fitness exits 0. What caught it is nstack e2e catalog — D1's own scenarios, with KeyError: 'approximateLocation' — and only because those scenarios read the contract's own examples, which the probe left untouched. Had the examples been updated in the same pull request, nothing in CI would have refused it, and every consumer of v1 would have broken silently.
This confirms what the contract rules already mark "to automate": compatibility of a contract version is checked by nobody. The producer's own tests are not a substitute — they pass by construction once the producer follows its own break.
Probes kept open
P1, P4, P5, P6, P7 and P8 are closed: the refusal was observed and is logged above. P2 (#22) and P3 (#23) stay open as live evidence of the defect, until the decider says what to do with them.
The plan requires a decision here:"a probe that passes when a refusal was expected: a defect in the register, and the maintainer decides whether to stop the validation."
Task 8 — the cycle closed, completed, seven days early
Pull request #29 awaits the decider. The cycle's state once it merges: closed, outcome: completed, ended_on: 2026-09-19.
Deliverables accepted
3 of 3 — D1, D2, D3
Deferred
none
Dropped
none
The cycle's real end
2026-09-19, seven days before its end date
Appetite
one week. The work took one day
The playbook's first closure path applies — every deliverable accepted — so the circuit breaker was never reached.
P11, verified without being faked
The breaker needs a delivery pull request after the end date, and 2026-09-26 has not arrived. Moving the real cycle's dates to manufacture it would have corrupted the record of a validation whose whole value is that its records are true. The rule was reproduced on a throwaway copy of the tree instead, with the window moved into the past:
FAIL [K2] circuit breaker: 01-two-modules-one-contract.md ended on 2026-09-12,
with no automatic extension.
Deviation recorded: P11 is verified as a rule, not as a pull request in this project.
And the honest limitation: the circuit breaker was never exercised on the real cycle, because the work finished early. The most interesting case of Task 8 did not occur, and no amount of reasoning replaces having seen it fire on a real cycle. It stays unobserved until a cycle actually overruns.
What the closure records, and what it refuses to hide
Every clause of the charter's only success criterion holds, and none of its three misses occurred. Criterion A is met as it is written.
The closure also states, in the cycle file itself, what the green cannot say: Task 7 showed that two of the barriers the criterion counts on cannot refuse anything for a contract dependency — B1 for an undeclared one, B2 for an implementation import reached through a path manipulation. Neither was bypassed here. They were never able to refuse.
Both statements are true, they are recorded together, and the exit decision has to weigh them together.
Task 7 — P9 and P10, run by the maintainer as the App. Both refused by GitHub itself
Target: pull request #30, head 408be1b, CI green, deliberately unapproved, so that nothing but the rule under test could explain a refusal. Run from the session's shell prompt, because the agent's own attempt is stopped by its harness before the request leaves the machine (M8f) — and a refusal by the harness would have proved nothing about the forge.
P9 — merge with no approval, as the App:
PUT repos/NapkinStack/tool-library/pulls/30/merge
→ HTTP 405
Repository rule violations found
Waiting on code owner review from NapkinStack/maintainers.
P10 — the App approves its own pull request:
POST repos/NapkinStack/tool-library/pulls/30/reviews event=APPROVE
→ HTTP 422
Review Can not approve your own pull request
Both refusals come from GitHub, and P9's names the missing rule rather than failing vaguely. ADR-0004's barrier holds on the validation project, under the ruleset a generated project receives — not only on the repository that develops the framework, where it was first observed in M8f.
Pull request #30 closed, never merged, branch deleted.
Task 7 complete — 11 probes
Refused as expected
9 — P1, P4*, P5, P6, P7, P8, P9, P10, P11†
Accepted when a refusal was expected
2 — P2 and P3
* P4 was a "to observe": no fitness rule refuses a broken contract version. It was caught by the producer module's own e2e, and only because those scenarios read the contract's examples, which the probe left untouched.
† P11 verified as a rule on a throwaway copy, the real cycle's end date not having arrived.
The two acceptances are one defect, already written up above: the boundary rules are blind to a contract dependency, which is the only dependency this framework allows between modules.
Every criterion of What is measured, against its target. Written before the exit decision, so the decision is taken on the record rather than on an impression.
The criteria, one by one
Criterion
Target
Observed
Tech lead — a compliant repository
≤ 30 min (PDR-0001)
~25 min of effective work, conditions logged (Tasks 1–2)
met
Tech lead — framed, first cycle started
≤ 2 h (PDR-0002)
148 min wall clock, ≈ 53 min of effective work
missed as written
Developer joining — no spoken conversation
0 sync events, each named
0 during the work; 1 between the teams (below)
met, with one named
AI agent — bounded
10 of 10 probes, or a defect
9 of 11 refused, 2 accepted
missed — defect recorded
Second team — moves forward without blocking the first
0 sync events, 0 waits
0 waits.loans merged against the contract alone
met
Evidence (PDR-0003)
verifier ≠ author, evidence on the head; review ≤ 15 min median
3 of 3 deliverables verified by a separate session with machine-produced evidence. Review: one measurement, 10 min
met on evidence; the median rests on n = 1
Cycle (PDR-0002)
every delivery pull request names a deliverable or carries the label; the cycle ends on its end date or earlier, no extension
7 of 7 merged pull requests conform. Cycle closed 7 days early, no extension
met
Questions (PDR-0004)
recorded from Task 4
recorded for every question, with stage, field, recipient
met
The sync events, counted honestly
Zero during the work. Three author sessions and three verifier sessions wrote, delivered and verified without one question they could not answer from the repository. D1: "I asked no question before writing." D2: "I did not need to come back with questions." D3: "Nothing was ambiguous enough to stop on."
One between the teams, and it deserves naming rather than rounding to zero. D1's author left an open question its own session could not settle: contract_double lives inside catalog, so loans cannot reach it without crossing the boundary — does team 2 build its own double, or does a shared one need a home neither module owns? The decider settled it before team 2 started. That is a synchronisation between two teams, mediated by a human, and the framework produced it rather than prevented it. It was cheap here; it is the shape of the expensive one.
PDR-0004's success criterion — met, well before 2026-10-31
It asked for at least three questions the framework should have raised and did not, each attributable to a stage that facts an existing check returns would establish. At least five were observed:
With no deployment and no real user, what counts as a verified deliverable? Nobody raised it; the maintainer did, reading a handover. Stage: a charter after a discovery that dropped its evidence criterion.
This module is user-facing and declares no run verb — while playbooks/verification.md tells the verifier to start the change with it. Both facts are in the manifest.
This module consumes a contract, and the boundary checks cannot see that dependency.consumes non-empty with no import detectable: both computable.
This pull request creates a user-facing module inside an accepted cycle; it will need a deliverable and a test sheet.new-module prints four next steps, mentions none of it, and holds every fact needed to say so.
This charter names tools, while docs/tooling-profile.md claims to be the only place a tool is named.
The defects, ordered by what they cost
1. The boundary rules are blind to a contract dependency — the only dependency this framework allows between modules. Observed three times: B1 accepts an undeclared one (probe P3, every check green), B2 is walked around by a path manipulation (probe P2, fitness exits 0), B4 warns on a real one (D3). docs/os/02-modules.md promises the declared graph is compared with the real one; for the boundary this product exists to enforce, that comparison does not happen. Every NapkinStack project is affected.
2. No check refuses a broken contract version. Renaming a required field of v1 in place, producer edited to match, passes every fitness rule (probe P4). It was caught only by the producer's own e2e, and only because those scenarios read the contract's examples, which the probe left untouched.
3. A human approval survives a force-push. No rule of the checklist requires GitHub's Dismiss stale pull request approvals; only the agent's own --match-head-commit stood between an approval and content its approver never read. Closed on both repositories on 2026-09-18; the checklist gap stands. Candidate G13.
4. T2 cannot tell a verifier from an author. The Verifier: line is free text; its own message says "someone other than the author" and nothing enforces it. Charter C1 clause 2 rests on convention.
5. nstack new-module produces a module that fails its own project, and says nothing about the sheet, the deliverable or the label it will need. check and test exited 1 on the generated module, measured. The field's convention is that a generator's output is green.
6. PDR-0002's criterion measures wall clock, therefore the decider's availability rather than the framework's cost. Hit twice — Tasks 2 and 4.
7. contracts/ and platform/ are modules for one check and invisible to two others.
8. The handbook contradicts itself on criticality; the framer had to arbitrate with no guidance.
9. playbooks/framing.md says nothing about the technical framing — zero mentions of ADR, tooling, stack, architecture. The M9 plan had to require the tooling profile by hand.
10. Technical decisions taken at framing have no artefact to land in, so the stack landed in the charter while tooling-profile.md claims to be the only place a tool is named. A structuring decision can also land with no ADR: new-module recommends one, nothing requires it.
11. provides is never confronted with reality. Nothing checks that its path exists or holds a document.
12. pr-check and pr-scope count modules differently, and print contradictory lines in the same run.
13. Two agent sessions on one checkout corrupt each other's branches. The framework describes parallel teams assuming separate humans on separate machines; nothing warns that "in parallel" may mean two sessions on one working tree. Both deliveries survived because the sessions repaired it themselves.
14. Two teams conflict on .github/CODEOWNERS at their first commit, because the scaffolding appends its owner line at the same place.
15. uv.lock is neither committed nor ignored in either module, so transitive dependencies resolve fresh on every run.
What worked, and it is not nothing
A full cycle ran end to end in one day against a one-week appetite: a discovery challenged by a session that had not written it, a charter and a cycle framed and accepted, two modules with separate owners, a hand-written contract provided by one and consumed by the other, three deliverables each with a test sheet written before the code and run by a session that did not write it, evidence machine-produced and re-runnable, and a closure recorded before the end date.
Nine of eleven barriers refused what they had to refuse, including the two that matter most and that only the maintainer could test: GitHub refused a merge without approval (405, naming the missing rule) and refused the App's self-approval (422).
And twice a barrier produced evidence rather than an obligation: T4 refused a sheet verified on another commit, which is what proves the author never touched the verifier's cells.
Taken by the decider on 2026-09-19. The pilot waits for a NapkinStack version that closes a named set of defects. No GitHub Team seat is bought today; the organisation stays on the Free plan, as decided on 2026-09-17.
The last measurement, which completes PDR-0003's criterion
Median 10 minutes against a target of 15, on three measurements. PDR-0003's criterion is met, and no longer rests on a single observation.
Worth reading with its context rather than alone: each of those reviews stood on a sheet whose scenarios had been written before the code and run by a session that did not write it, and on evidence produced by machine. The ten minutes are what the sheet and the verifier bought. They are not ten minutes of reading 700 lines.
Why fix first rather than go
The cycle ran end to end and every barrier protecting the human held: GitHub refused the merge without approval, refused the App's self-approval, and refused a sheet verified on another commit.
What did not hold are the barriers protecting the architecture. docs/os/02-modules.md promises that the declared graph is compared with the real one. For a contract dependency — the only dependency this framework allows between modules — that comparison does not happen, and three probes showed it: an undeclared dependency passes, an implementation import passes behind a path manipulation, and a real dependency is reported as unused.
Sending a real project onto that would be shipping the differentiator broken. Every competitor surveyed produces documents; this product's claim is that the rules are executable. A rule that cannot refuse is a document with extra steps.
Why not rethink
No decision proved wrong. PDR-0001, PDR-0002, PDR-0003 and ADR-0004 each did what they promised where they were implemented. They proved incompletely applied, which is implementation work, not design work. Reopening them would spend weeks to arrive back where we are.
M10h, step 1 — the update to v0.4.0, paid for real (#31)
NapkinStack v0.4.0 was published on 2026-09-19 to fix what M9 found here (D29 to D33, D35). M10h proves the fixes where they were found; this is its first step, and M9's Task 10: the first update of this project.
What the update asked for
Observed
nstack update, v0.3.1 → v0.4.0
Copier's merge, no conflict, 19 files, nothing edited by hand
What it left red
fitness M7 on contracts/, twice: v0.4.0 no longer ships placeholder verbs, and contracts/ holds catalog-api v1 — the adaptation the release notes announce
The adaptation
check: the document is valid OpenAPI 3.1 (openapi-spec-validator, pinned) and the tests' Python passes the modules' ruff rules. test: every example is what its schema declares — loans builds its double of catalog from them, and the structural check lets a Problem example with status: 200 (minimum 400) through — and a Problem says the status its response answers (RFC 9457 §3.1.3). 5 files, 114 lines, in C2's stack
What pr-check then asked
T1, a test sheet: the adaptation adds tests to contracts, of criticality high — delivery work by the framework's own definition. K1: cycle 1 is closed and no cycle is framed, so the out-of-cycle label, with its justification. pr-scope: over the files budget (22/15), over-budget
The verification
A verifier session (dec9b511…, fresh context, launched from the author's session — said in the pull request) ran the sheet twice. The first run found a real gap: a 404 example whose body says 400 passed both verbs, while the double answers with the response's code and the example's body. It also found referencing unpinned, ruff's target version missing, and two commit hashes swapped in the description. All fixed in 20d484a; the second run, S1 to S5, passed on it, and confirmed that the update's commit is Copier's merge unchanged by hand (same tree from a fresh nstack update)
First run of Module checks here
Green: Changed modules found contracts alone; its job installed the engine from the registry (napkinstack==0.4.0), ran check and test, and compat said No frozen contract version changed
Cost: the update itself, nothing. The adaptation, one contract test and the manifest's verbs — and a sheet, since the module is of criticality high; that sheet is what caught the gap. The plan expected no label, no sheet: true of the update alone (measured before the release), not of the adaptation, which adds code to a high-criticality module.
Found
K1's message is wrong between two cycles. The charter is accepted and cycle 1 closed; K1 says "the project is not framed" and sends the reader to playbooks/framing.md, when only the next cycle is missing. The rule is right; its message sends a team back to framing when it only needs a cycle. For NapkinStack's register.
The step Every touched module passed of module-checks.yml still says "touched" where v0.4.0 says "changed" for what CI lists. Cosmetic. For NapkinStack's register.
Upstream, harmless: Copier prints "Make sure Git >= 2.24 is installed" on Git 2.55 — its first attempt, git diff-tree --inter-hunk-context=-1, is refused by recent Git and it falls back to 0. The merge is unaffected.
Sync events
None.
State
#31: every check green, BLOCKED on the maintainer's approval only. After the merge: Module checks made required on main, and nstack doctor run with the maintainer's own token.
Each a pull request, never merged, closed once logged here. Judged by CI on GitHub, and reproduced beforehand with the project's pinned engine (napkinstack==0.4.0).
✅ Refused.FAIL [B2] modules/loans/src/loans/catalogue.py:75 (the path into catalog's folder) and :76 (the import)
accepted, fitness exit 0
P3
#35 — M9's #23 replayed: consumes emptied while loans reads the contract
Fitness functions — B1
✅ Refused.FAIL [B1] modules/loans/src/loans/contract.py:17 and modules/loans/e2e/conftest.py:4, each with the exact consumes entry to add
accepted, every check green
P12
#37 — a change to loans whose sheet is signed Verifier: session <id> by the session its commit's Agent-Session: trailer names
Test sheet and cycle — T2
✅ Refused.FAIL [T2] Test sheet: the verifier, session d9d08c03-…, wrote commits of this change (Agent-Session trailer). Every other check green: T2 alone refused it
accepted — the Verifier: line was free text
P14
#36 — nstack new-module reminders NapkinStack/maintainers standard --user-facing, alone, untouched
every check green, no label, no sheet
✅ All seven checks green, no label, no sheet: Changed modules listed reminders, whose job said nothing to check yet; pr-checkTest sheet: not required
check and test exited 1
Beside the target, on P2: loans' check refused the probe's unformatted lines (ruff), and Test sheet and cycle asked for a sheet and a cycle (T1, K1) — a real change of this shape would meet three barriers, not one.
P14's next steps, as new-module prints them, now name the sheet, the cycle and M7 — M9's D33 asked for exactly that. One of them repeats K1's wording already reported (step 1): "The project is not framed", when only the next cycle is missing.
V1 runs the merged command in the pull request's tree. A proof that calls a file of the module — a script, or a binary bootstrap installed — can be rewritten by the very change it judges. contracts — a frozen contract version is changed only when oasdiff proves it compatible #32's command therefore carries everything itself. The framework should say so, or run the proof from the base's tree.
V1 says "a breaking change" whenever the proof exits non-zero — also when it could not run at all (no network, a wrong pin). "Could not prove it compatible" would be true in both cases.
A multi-line compat command is printed twice, in the -> line and inside the FAIL [V1] message, burying the comparator's own verdict.
M10h, step 3 — P4, P4c and P4b replayed on v0.4.0, after #32 and #33
Each a pull request, never merged, closed once logged here. Judged by CI on GitHub, and reproduced beforehand with the project's pinned engine (napkinstack==0.4.0). main = 87bb76b: contracts declares compat (#32), the modules' locks are committed (#33).
#
The pull request
Expected on v0.4.0
Observed on v0.4.0
On v0.3.1 (M9)
P4
#38 — M9's #24 replayed as it was: approximateLocation renamed whereabouts in required and in one example, catalog edited to match
Module checks — V1, oasdiff naming the property
✅ Refused — but not by V1 in CI.contracts' test (#31) refused first: the two examples no longer match their schema ("'whereabouts' was unexpected", "'whereabouts' is a required property"), and the job stopped before compat. catalog's test refused too (1 failed). V1, run locally on the same commit: FAIL [V1] catalog-api v1 (consumed by loans), oasdiff response-property-became-optional on approximateLocation, twice
caught by luck, by the producer's own e2e
P4c
#39 — the change M9 said nothing would refuse: the same rename made consistently, in the property, required and every example, and in catalog's code and tests; loans untouched
V1 alone
✅ Refused by V1 alone.contracts check and test (8 passed), catalog check, test (8) and e2e (7 + 2 subtests), Fitness functions: all green. FAIL [V1] catalog-api v1 (consumed by loans), oasdiff response-required-property-removed on approximateLocation, twice. Locally on that commit, loans' test: KeyError: 'approximateLocation', 2 of 15 failed — the consumer broken, and no job of its own ran
not run; M9: "had the probe updated them too, nothing would have refused"
P4b
#40 — instance, RFC 9457's own member, added to Problem in v1, optional
Module checks green: V1 proven compatible
✅ Module checks green: "Contract version catalog-api v1 (consumed by loans): proven compatible."
—
Beside the target.#38 and #39: P1 (two modules), T1 (a sheet: contracts is high criticality, catalog user-facing) and K1. #40: T1 and K1 — a real compatible change to the contract still carries a test sheet, and a cycle or the out-of-cycle label.
Found, for NapkinStack's register
From the probes:
A step failing before compat hides V1's verdict.module-checks.yml runs compat after test, with no condition: on P4, contracts' test failed and compat never ran. Nothing gets through — the pull request is red either way — but the refusal that names the real problem (a frozen version broken) arrives one round trip late, after the author has fixed the examples. Running compat before test, or under if: ${{ !cancelled() }}, would show it at once.
From #33 (the locks, merged as 87bb76b after four verifier rounds on four heads — each round found something real, fixed before the next):
--locked says nothing about uv run --with: resolved in a layer outside the lock. The verbs now use dependency groups. contracts' test still runs uvx --with — the same unlocked layer, with no project of its own to lock.
uv run never removes an extraneous package, and CI's verbs share one environment: a verb could pass on another group's packages. The verbs now run --exact.
setup-uv pins no uv version; with --locked, a uv release that writes locks differently would turn the modules red with no change in the project. The workflow is the framework's.
The runner's Python (3.12) is below the modules' floor, so uv downloaded the newest, 3.14.7, while the charter says 3.13: each module now carries a .python-version.
The verifiers are subagents of the author's session — a fresh context, the brief alone, an identifier of their own. T2 compares names and cannot see that link; disclosed in the pull request's risks, as on NapkinStack v0.3.1 -> v0.4.0, and the contracts module's check and test #31, and approved with it.
P13 — a module test that fails: Module checks red, the pull request BLOCKED, read through the API, no merge attempted. Module checks is a required check of main (read through the rules API).
A pull request, never merged and never approved, closed once logged here. Reproduced beforehand with the project's pinned engine (napkinstack==0.4.0).
#
The pull request
Expected on v0.4.0
Observed on v0.4.0
On v0.3.1 (M9)
P13
#41 — a regression in loans: the guard refusing a return date before the hand-over date removed from Records.record
Module checks red; the pull request BLOCKED, read through the API, no merge attempted
✅ Blocked.loans' test: FAILED tests/test_recording_a_loan.py::test_a_return_date_before_the_hand_over_is_refused_and_nothing_is_recorded — DID NOT RAISE Refused, 1 failed, 14 passed; Module checks: "Changed modules: success. Their checks: failure." Read through GraphQL on head 82244cd: mergeStateStatusBLOCKED, Module checksFAILURE, isRequired: true — a failing required check, which an approval does not lift. No merge attempted
mergeable once approved: Module checks was not a required check (D35)
Beside the target.Test sheet and cycle: T1 (loans is user-facing) and K1. check and Fitness functions green: the regression is well-formed code within its boundaries — only the module's own test knew the rule. Locally, loans' e2e refuses it too (1 of 4), but in CI the job stops at test, so compat and e2e never run — the same ordering as P4's.
Required checks of main, as the API reads them on #41: Fitness functions, PR scope and review budget, Hooks and secrets, Test sheet and cycle, Module checks. Changed modules and each module's own job are not required — Module checks carries them, since a matrix job's name cannot be required in advance.
Every probe replayed
P2, P3, P4, P4b, P4c, P12, P13 and P14 are run on v0.4.0. The closing comment compares each with M9 and sums up the update's cost.
M10h — closed: every probe against M9, the update's cost, the sync events
NapkinStack v0.4.0 was released to fix what this project exposed in M9 (D29 to D35). M10h paid for the update here and replayed the probes. Everything below is linked from the steps above: #31 (step 1), #34–#37 (step 2), #38–#40 (step 3), #41 (step 4).
Eight of eight did what v0.4.0 promises: the six that must be refused were refused, the two that must pass passed. Of M9's eleven probes, the three not refused by a rule — P2, P3 and P4 — now are. The eight it refused rest on rules v0.4.0 did not loosen; P1 (#38, #39) and T1, M9's P5 (#38 to #41), refused again beside the target.
The forge, checked by the maintainer with their own token after the update — nstack doctor: L1 to L7 and G1 to G13 compliant, Module checks among G4's required checks, G13 (stale approvals dismissed) enforced.
the modules' locks committed (M9's D34 #10), every uv run locked and exact, Python 3.13
13 files, +1,842 −13 (1,780 generated)
a sheet: 4 runs on 4 heads, each found something real, fixed before the next
4 h 39
Wall clock includes the wait for the maintainer's approvals. The update itself cost nothing. Its adaptations are what v0.4.0 asks of a project holding a consumed contract: check, test and compat on contracts. The locks were not asked by the update but by M9's own finding. The maintainer's part: three approvals, Module checks added to the ruleset, one nstack doctor.
Deviations from the plan, each recorded where it happened: #31 needed a sheet and two labels (the plan said none — the adaptation adds tests to a high-criticality module, and the sheet caught a gap); compat is self-contained rather than installed by bootstrap (V1 runs it in the pull request's tree); P4c added, because P4 no longer isolates V1; M9's P1 and P5 to P11 not replayed, their rules unchanged.
None during M10h: no question had to be settled in conversation. The maintainer's part — three approvals, the ruleset, doctor — was the plan's, taken on what the pull requests and this log said.
State
main = 87bb76b, no pull request open, every probe closed and its branch deleted. Next, in NapkinStack/engineering-os: the M10 closure — the plan ticked with what was observed, the register above — and the maintainer's exit decision.
Taken by the maintainer, decider, on the closing comment above.
Why. Every probe replayed in M10h did what v0.4.0 promises, and nstack doctor is compliant. But one finding is of the class the v0.4.0 tag was held for (D36 to D39, a barrier that can be walked around): V1 runs the merged proof in the pull request's tree, so a project whose compat calls a file of its module — a script, a binary its bootstrap installed — can have that proof rewritten by the very change it judges. This project is not exposed: its command carries everything itself (#32). A pilot project would be, from the day it freezes its first contract version.
What. A v0.4.1 in which V1 cannot be walked around this way, and its verdict is never hidden by an earlier failing step nor worded as a breaking change when the proof could not run. It is planned in NapkinStack/engineering-os; the M10 closure pull request places every other finding of the closing comment.
Then. The organisation moves to GitHub Team and the private pilot starts on v0.4.1.
M10h's last entry — NapkinStack v0.5.0, and what the update cost this project
NapkinStack v0.5.0 was published on 2026-09-20 to close the defect M10h itself exposed: V1 read the proof as merged but ran it where the change is (D44). This is what the version cost this project, and what it proved here.
The update: nothing was asked
Observed
nstack update, v0.4.0 → v0.5.0
Copier's merge, no conflict, 6 files, 84 lines, nothing edited by hand
What it left red
Nothing. The v0.4.0 update had needed the contracts module's check and test written before it could pass; this one merges and passes as it stands
What pr-check asked
Nothing: compliant with no label and no sheet — an update changes no module beyond its description (D24). The out-of-cycle label was added for K1, cycle 1 being closed
What it brings
commits-on-main.yml, new; the compat step under if: ${{ !cancelled() }} (D46) and its closing step saying "changed" (D47); the handbook on where a contract proof runs (D44) and on what remains when the forge can enforce nothing (PDR-0006); the README's two words, guarded and unguarded
D44 — a proof that judges the change cannot be rewritten by it. Replayed on a clone of this project before the update branch was cleaned: catalog-api v1 broken consistently — the property renamed in the schema, the examples and the producer's code — is refused.
error [response-required-property-removed] ... removed the required property
`tools/items/approximateLocation` from the response with the `200` status
error [response-required-property-removed] ... removed the required property
`approximateLocation` from the response with the `200` status
FAIL [V1] catalog-api v1 (consumed by loans): the compatibility command of module 'contracts'
(contracts/MANIFEST.yaml) did not prove the change compatible (exit 1).
Two differences from v0.4.0, both visible above: the comparator's own output is above the verdict instead of being buried under a command repeated whole (D45), and the verdict says the proof did not prove compatibility rather than asserting a break it may not have observed. This project's compat command is self-contained, which is why the change to where it runs asked nothing of it — a command that had called a script of its own module would have had to become self-contained (D44).
The record — its first real run, on this repository. The merge of #42 pushed to main, and Commits on main ran for the first time:
Judged by NapkinStack 0.5.0, published, project recorded at v0.5.0.
Commit 36f1adf (napkinstack-agent[bot], 2026-09-20): not recorded, older than the record itself.
That is the designed behaviour, seen for real: the commit that brought the record is the one commit it says nothing about (PDR-0006). Its first firing on a commit pushed outside a pull request has not happened yet, and will be a probe, logged here like the others.
Found, in real use
A refusal named an action rather than a command. Running nstack update with an engine older than this project refused correctly — "use nstack 0.4.0 or newer" — and the maintainer then tried uv tool upgrade napkinstack, which does nothing at all on an install pinned to an exact version. Invariant P6 asks for the action; an action nobody can copy is a description. Fixed in NapkinStack (D50): the refusal now carries uv tool install "napkinstack@latest" --with-executables-from pre-commit.
An update commit carries the workstation's git identity. The commit nstack update created on the maintainer's own checkout was authored by the machine's global identity, unrelated to this project; on a public repository that identity would enter the history. The branch was re-authored under the App before being pushed. Nothing to fix in the framework — a note for whoever runs an update on a shared workstation.
Sync events
None.
State
main = 36f1adf, on NapkinStack v0.5.0. No pull request open, no probe open. This project has now been carried across two framework versions, v0.3.1 → v0.4.0 → v0.5.0, the first at the cost of one adaptation and the second at no cost at all.
Activity
napkinstack-agent commentedon Sep 17, 2026
Task 1 — repository, teams, App (2026-09-17): the public repository created empty; teams
catalogandloanscreated, write access given tomaintainers,catalogandloans(the maintainer's statement: the App cannot read team permissions); the repository added to thenapkinstack-agentinstallation — verified: the installation listsengineering-osandtool-library; allowed actions includeastral-sh/setup-uv— verified by the first CI run.Task 2 — timed
nstack init, in progressnstack initfromnapkinstack==0.3.1, commit "Project created, NapkinStack v0.3.1",_commit: v0.3.1.napkinstack-adminonly; gitleaks over the history: no leak.mainpushed by the agent under its App.Governanceon the push:Hooks and secretsandFitness functionsgreen;Modulesgreen.PR scope and review budgetruns on pull requests only.nstack doctor.napkinstack-agent commentedon Sep 17, 2026
Task 2 — first
nstack doctor, run by the maintainer with a short-lived token of their own (Administration: read), before the checklist:Finding F1 (minor, message): with no ruleset on
main, G12 fails with "remove every bypass actor" — there is no actor to remove; the action that applies is G1's, "create the ruleset". Source:_no_bypassindoctor.pyreturns false when no ruleset applies, and the message is G12's only one. P6 asks for the action that fits the case.napkinstack-agent commentedon Sep 17, 2026
Task 2 — the checklist, G11: the three labels
cross-module,over-budgetandout-of-cyclewere created by the agent under its App, at the maintainer's request, with the colours and descriptionsengineering-osuses. Counted as help from the agent: a tech lead alone would have created them by hand in Issues → Labels, three forms.Finding F2 (in NapkinStack's own repository):
engineering-oshascross-moduleandover-budgetbut notout-of-cycle, required by G11 since M8d — its owndoctorwould report the gap. To fix there, not here.napkinstack-agent commentedon Sep 17, 2026
Finding F3 (friction at setup, G4): the checklist asks the tech lead to require four checks in the ruleset, but GitHub only offers a check it has already seen.
Test sheet and cycleandPR scope and review budgetrun on pull requests only, so on a repository that has none they cannot be selected — the maintainer hit exactly that. What unblocked it: opening a first pull request (#2) so the checks run once. The README anddoctor's G4 action say nothing about it.Task 2 — the first pull request (#2):
README: what this repository is for, by the agent under its App, assigned to the maintainer. The four checks ran and passed:Fitness functions,Hooks and secrets,PR scope and review budget,Test sheet and cycle, plus theModulesworkflow (no module touched).Also observed, in a real project this time: a pull request that changes no module beyond its description asks for no test sheet and no cycle — NapkinStack v0.3.1's D24 fix behaving as intended.
napkinstack-agent commentedon Sep 17, 2026
Task 2 — done.
nstack doctor, run by the maintainer with their own short-lived token: 0 gap, 0 not verified. Verified from the agent's side, through the App, which has no Administration permission: the rules applying tomainrequire a pull request, 1 approval, a code owner review, no force-push, no deletion, and the four checksFitness functions,PR scope and review budget,Hooks and secrets,Test sheet and cycle. The bypass list and Advanced Security are readable only with the maintainer's token: theirdoctorrun is the evidence.Duration — PDR-0001's criterion (a compliant repository in under 30 minutes): met.
date, beforeuv tool install)date, after the lastdoctor)The wall clock is not the working time: the session was interrupted several times. The figure that counts is the maintainer's own, 25 minutes.
Conditions, so the figure is not read for more than it is worth:
engineering-os: a first project in a fresh organisation would add that work.The first pull request (#2) is merged, approved by the maintainer, merged by the agent under its App: the barrier held from day one — the pull request stayed
BLOCKEDuntil the approval.napkinstack-agent commentedon Sep 17, 2026
Task 3 — discovery, started.
nstack discoverkept, indocs/project/inputs/, with a line saying it is a translation. The French original stays outside the repository.nstack discoveron branchdiscovery/tool-library: the idea kept,docs/project/discovery.mdcreated, and the command printed what to do next in the agent. It called no model — as PDR-0002's extension requires.napkinstack-agent commentedon Sep 17, 2026
Task 3 — discovery: done, decided
go(pull request #5, merged 2026-09-18 by the agent under its App, after the decider's approval).clarifygo— an acknowledged rehearsal, criterion A only; criterion B (twenty real loans) dropped, not deferred; no test kept as a spikenstack fitnessandnstack planpass; C7 mutation-tested on a throwaway copy: agowith nodecided_onfails, agowith nochallengerfailsWhat the framework earned here. The challenger found in one session what stage 2 had missed and what decides the product:
entrevoisins.euships the same product free in French today;smiile.com, the French reference for free neighbourhood mutual aid, now serves a domain-sale page after 500,000 neighbours; Mutum was liquidated with 113,000 objects listed. It also showed that the success criterion as written could not be passed or failed and had no owner — the ending "framework declared, product never measured" was the likeliest one on the document's own evidence. The decider removed the criterion rather than arrange that ending. The rule that produced this is PDR-0002's extension — a challenger in another session — and it paid for itself on the first use.What it leaves for the framing, and criterion A depends on it: with no deployment and no real user, what counts as a verified deliverable? PDR-0003 asks for test sheets run by a verifier, with evidence; that evidence can no longer come from a neighbourhood. If the charter does not settle it, a cycle can close green while proving only that the paperwork moved — the same structure the challenger found on criterion B, displaced onto criterion A. Eight places in sections 1 to 5 still speak of the dropped criterion B; those sections stay as the round-1 record, so the charter has to settle them.
Sync events so far: 0. The framer and the challenger asked nothing outside this repository; the decider's answers went through the orchestrating session and are posted in #3 and #4.
napkinstack-agent commentedon Sep 18, 2026
Task 4 — framing: the brief is published, the session is not started
2026-09-18. The framer's brief is issue #6, published before the session that will follow it exists — the same isolation rule the discovery used, where it proved its worth.
The timer has not started. It starts when the framer session opens and stops at the merge of the accepted charter and cycle, per the plan. The decider is not available for the interview yet, and starting the session without them would spend the two-hour budget on waiting.
What the brief fixes, so that it cannot drift once the session is running:
Recorded for PDR-0004, accepted the same day: from this task on, every question put to the decider is logged with the stage of the project, the field it comes from, who it was put to, and whether a check could have answered it. The framer is not told about this. Its brief says nothing of it, so that observing the interview does not change it; the orchestrating session records from the outside, since every question passes through it anyway.
napkinstack-agent commentedon Sep 18, 2026
Task 4 — the timer starts
2026-09-18, 10:34 UTC. The framer session is open. It is a fresh session: it carries nothing of the orchestrating session's conversation, and receives only issue #6, published before it existed.
The timer runs from this moment to the merge of the accepted charter and cycle. Budget: two hours (PDR-0002).
What the session was given beyond its brief, recorded so the measurement can be read honestly: the plumbing of this validation, not of the framework — the bot's commit identity, the push form that works under the sandbox, the
gh-agentwrapper, the pinned engine version, and the hosts each command needs. These are artefacts of how this agent is isolated (ADR-0004) and of the workstation's sandbox. A question about any of them would be noise in the sync-event count, not a finding about the framework.What it was not given: anything about the project, the decision, the products the challenger found, or what the decider wants. All of it has to come from the repository or from a question.
Known gap, recorded before it can be claimed as a success. This repository pins
_commit: v0.3.1, so its CI installsnapkinstack==0.3.1. Rules H1 (D27) and the decider check (D28) were merged after that release and are not in it: a charter keeping the template'sdecider: "@<github-handle>"placeholder would pass this repository's CI today. The orchestrating session runs the newer checks against the branch as a second gate. This is what Task 10 will exercise for real.napkinstack-agent commentedon Sep 18, 2026
Task 4 — batch 1: six sync events, and the first PDR-0004 measurement
2026-09-18, 10:41 UTC (7 minutes in). The framer read the brief,
AGENTS.md, both playbooks, all 629 lines of the discovery, the templates, seven chapters of the handbook, the three workflows and the engine's own check code at v0.3.1 before asking anything. It wrote nothing and created no branch until its questions were answered. Questions and answers are on issue #6.How the answers were produced, recorded so this can be read honestly later: the decider asked the orchestrating session, acting as product expert, to draft them; the decider ratified them and amended one. They are the decider's decisions. The six questions are genuine sync events either way — no session could have answered them from the repository — and are counted as such.
Sync events: 6
e2euvas the only prerequisite, phone-first, near-zero budget. The repository states outright that it supplies no stack, so the choice is the decider'sWhat this first sample says. Of six questions, one (Q6) a convention already in the repository could have removed, and one (Q4) the repository answered twice in opposite directions — a category PDR-0004 had not anticipated: not a question the framework should have asked, but a contradiction it should not have left standing. The remaining four are genuine decisions that belong to a human, and every one of them arrives at a stage that predicts it: verification and product at the charter, architecture at the tooling profile, governance before scaffolding.
Defect candidate, for Task 9
The handbook contradicts itself on criticality:
playbooks/framing.mdraises it for a domain handling personal data,docs/os/07-governance.md§6 forbids the ceremony of a critical system on a small module. Both apply here and they point opposite ways. The framer resolved it correctly — the personal-data trigger is about real personal data, and nothing here is deployed — but it resolved it, rather than reading it.Deviation from the M9 plan, recorded
The plan requires "team 2's first deliverable ready only once team 1's first is merged". The framer applied it faithfully and gated
D3(loans) onD2(catalog's user-facing deliverable). The decider amended the constraint:D3gates onD1, the versioned contract, instead.The reason is the plan's own purpose. Gating the consumer on the producer's interface contradicts what cycle 1 exists to demonstrate — that two modules know each other only through their contract — and it concentrates the risk: if
D2slips inside a one-week appetite,D3never starts and the two-team observation, the most valuable thing this validation can produce, is lost entirely. The dependency and the sequencing are still real; they now run through the contract, which is where the framework claims they should.napkinstack-agent commentedon Sep 18, 2026
Task 4 — the clock, recorded honestly while it runs
2026-09-18, 11:46 UTC — 72 minutes since the framer session opened. The pull request (#7) has been waiting for the decider since 11:05. The plan measures the framing as wall clock from the session's start to the merge, so that wait is inside the measurement. It is recorded here rather than explained away at the end, and it follows the precedent of Task 2, where effective work and wall clock were logged separately with their conditions.
What this says about PDR-0002's criterion, and it is not a detail. The criterion is "the framing in one session of two hours". Effective work is well inside it. Wall clock will not be, and that is not a defect of the framing — it is the criterion measuring the decider's availability as if it were the framework's cost. A tech lead evaluating NapkinStack will read "two hours" as two hours of my attention, not two hours of calendar. Candidate finding for Task 9: PDR-0002's criterion should be stated in effective work, with the wall clock recorded beside it. Task 2 already hit this and worked around it by logging both; twice is a pattern, not an accident.
A second observation, for PDR-0004. Part of the wall clock above is a design conversation that had nothing to do with framing this project. The orchestrating session could not have known that in advance, and neither could any timer — which is a reminder that a stage tells you what to ask, never how long a human will take to answer. Nothing in PDR-0004 should ever try to measure the second.
napkinstack-agent commentedon Sep 18, 2026
Task 4 — done. The framing is merged, and PDR-0002's criterion is missed on the measure the plan chose
Merged 2026-09-18 at 13:02 UTC — pull request #7,
73bb72b. The charter and cycle 1 areaccepted. The timer ran from the framer session opening at 10:34 to that merge.The criterion is missed as written, and met on what it was trying to measure. This is the second time: Task 2 hit the same thing and logged both numbers with its conditions. Twice is a pattern. A tech lead reading "the framing in one session of two hours" understands two hours of my attention, not two hours of calendar — and calendar time is mostly the decider's availability, which is not a cost of the framework. Finding for Task 9: state the criterion in effective work, and record wall clock beside it.
What the framing produced
Three deliverables, one week, 2026-09-19 to 2026-09-26.
D1the versioned contract and the double built from its schema,D2andD3one user-facing outcome per module,catalogandloans, bothstandardanduser_facing.D3gates onD1— the decider's amendment to this plan, so that the consumer waits on the contract rather than on the producer's interface.The question the framing existed to settle is answered as charter constraint C1, in four clauses: the sheet is the criteria; the verifier is a separate session and no scenario may be
human onlywhile nothing is deployed; the evidence is machine-produced and re-runnable, never a screenshot pasted by hand; and the criteria are fixed at acceptance, because the sheet is the criteria means nothing if the criteria can move.Six findings, from one framing session
contracts/andplatform/are modules for one check and invisible to two others.MODULE_DIRSincludes them;module-checks.ymlgreps^modules/andpr_scope.shgreps^(modules|services|apps|packages)/. A pull request touchingcontracts/alone requires a test sheet and can produce no.evidence/artefact — it would have no admissible evidence under C1. Found by the framer, verified by the orchestrating session.playbooks/framing.mdraises it for a domain handling personal data;docs/os/07-governance.md§6 forbids the ceremony of a critical system on a small module. Both applied; the framer had to arbitrate with no guidance.playbooks/framing.mdsays nothing about the technical framing. Zero occurrences of ADR, tooling, stack or architecture. Module decomposition is routed todocs/os/02-modules.md§3; nothing else is. The M9 plan had to require the tooling profile by hand, because the playbook does not.docs/tooling-profile.mdstates it is "the only place in the repository where a tool name appears". The framer's bounds forbade touching that file. The charter should point at the profile rather than duplicate it, and nothing says so.nstack new-moduleprints a reminder to write a creation ADR; nothing requires one. Two modules and a versioned contract are exactly whatdocs/os/06-decisions.mdsubmits to the Prior Art Gate.Findings 4, 5 and 6 are the same defect seen three times: technical decisions taken during the framing have no artefact to land in, so they land in the product document.
playbooks/framing.mdis not being corrected now — it is the thing under measurement, and changing it mid-validation would destroy the measurement.napkinstack-agent commentedon Sep 18, 2026
Finding — a human approval survives a force-push, and nothing in the checklist refuses it
Observed 2026-09-18, on pull request #9 of this repository.
The decider approved commit
f40ae95. The branch then had to be rebased ontomain—nstack new-moduleappends a line to.github/CODEOWNERS, so the two module-creation pull requests conflicted on that one file. After the rebase the head became2859e7e, a commit the decider has never seen.GitHub still reports, on the current head:
The merge was refused by the orchestrating session, not by the forge. The only thing standing between an approval and content the approver never read is
--match-head-commit, passed by the agent to its own merge command. That is the agent's discipline, not an enforced rule — and the agent is the party that pushed the new commits.This matters because it is the one barrier ADR-0004 rests on. The decision took care that an agent can never approve; it did not take care that an agent can change what was approved after the fact. The barrier is weaker than the ADR claims, and the gap is closed today only by a convention the ADR does not name.
The setting exists on GitHub — "Dismiss stale pull request approvals when new commits are pushed" — and no rule of the checklist requires it:
doctorruns G1 to G12, covering the pull request requirement, the approval count, code owner review, required checks, secret protection, vulnerability reporting, allowed actions, SHA pinning, workflow approval, workflow token, labels and the empty bypass list. Stale approvals are not among them.Candidate G13, and a candidate defect against ADR-0004. Both for Task 9.
Second finding, from the same event: two teams conflict on their first commit
nstack new-modulewrites its owner line into.github/CODEOWNERS, at the same place for every module. Two teams creating their module in parallel therefore conflict before a single line of code exists — on the file that decides who approves what.The conflict is trivial and resolves in seconds. What it costs is the promise: several teams work in parallel with no synchronisation meeting. The very first action of the second team needs the first team's work to be merged. A file per module, or an ordered insertion, would remove it.
napkinstack-agent commentedon Sep 18, 2026
The stale-approval hole is closed on both repositories — the checklist gap is not
2026-09-18. The maintainer enabled "Dismiss stale pull request approvals when new commits are pushed". Verified from the agent's session, through the API, on both rulesets:
The barrier ADR-0004 rests on is now held by the forge. Replayed against today's event: a rebase after approval would dismiss it, the pull request would return to
REVIEW_REQUIRED, and the agent — which cannot approve — could not merge what it had just pushed.require_last_push_approvalwas leftfalse, deliberately. It was first proposed as complementary; it is redundant once approvals are dismissed on push. It would only cover a case this flow does not have: an approver pushing after their own approval.What is not fixed, and stays a finding for Task 9: nothing in the framework requires this.
nstack doctorruns G1 to G12 and none of them reads this setting, so a project generated tomorrow ships with the hole open anddoctorreports it compliant. Candidate G13. The instance is closed; the defect is not.Minor, same audit
The two repositories disagree on merge methods:
engineering-osallowssquashonly,tool-libraryallowsmerge,squashandrebase. The flow squash-merges everywhere. Not a hole, but a divergence between the reference repository and the project it generates, and the checklist does not cover it either.napkinstack-agent commentedon Sep 18, 2026
Task 5 — D1 delivered, verified twice, merged. Measurements
Merged 2026-09-18, pull request #11 →
08fb6cf. Deliverable D1:contracts/catalog-api/v1/openapi.yaml, hand-written, and a double built from that document alone.What the plan asks to be measured
AGENTS.md/MANIFEST.yamland the engine's own rules were read." Neither verifier was blocked. The definition applied is M9's — a question a session could not answer from the repository and had to put to the maintainer. The budget overrun and the two findings reached the decider through the pull request, which is the review channel, not a synchronisationT2 no verifier namedon the author's commit — by design, the author leaves the sheet empty.T4 scenarios verified on another commit than the headafter the fix — the evidence was produced on the previous head. No spurious failureWhat the barrier caught that a green sheet had not
The sheet was green and complete when the first verifier finished. The review that followed still found two real defects:
No module named contract_double, because the module declares no[build-system]and the scenarios passedPYTHONPATHto their subprocess themselves. It mattered because D3's premise is thatloansbuilds against this double, and the README is where that session will look.The decider's call: fix the first, and for the second withdraw the claims rather than build coercion — a promise known to be false costs more than a limitation that is declared. The author's fix removed the trap instead of documenting around it: the command in the README is now the one the scenarios themselves run, so the two cannot drift apart unnoticed.
Proof of posture, and it is mechanical rather than stated
After the fix, the sheet still carried
729cf67in its three commit cells while the head wasbd1227f, andT4was red for exactly that. Had the author updated those cells,T4would have gone green and the verification would have been a lie. Its red is the evidence that the author stayed inside their posture — no human had to check.Framework findings from this task
Verifier:line is free text; the check only refuses a placeholder. Its own message says "someone other than the author of the change" and nothing enforces it. A sheet filled by the authoring session would pass identically. Charter C1 clause 2 rests on convention alone. Same shape as the stale-approval hole: the barrier is a habit, not a constraint.providesis never confronted with reality.boundaries.pynever readsprovides; nothing checks thatprovides[].pathexists or holds a document. The rule that kept it empty until D1 was held by hand alone.nstack pr-checkandnstack pr-scopecount modules differently and print contradictory lines in the same CI run on the same commit.nstack e2ehas no way to say a scenario is about the module's own envelope. Noted for what it is: not a defect, an absence observed while writing the first real sheet.napkinstack-agent commentedon Sep 18, 2026
Task 7 — the drift probes: 6 of 8 refused, 2 accepted what should have been refused
P1 to P8 run on 2026-09-19, one pull request each, none merged. P9 and P10 are the maintainer's, P11 belongs to Task 8.
Deviation from the plan, recorded: the plan asks for an author session per probe. They were written by the orchestrating session instead. Eight sessions for eight deliberate one-line violations is disproportionate, and what is measured — does the check refuse? — does not depend on who wrote the line. Every probe was also reproduced locally on the project's own pinned engine, so each result below is a rule identifier, not a red cross.
cross-modulelabelPR scope and review budgetrefusesloansimportscatalog's implementationnstack fitnessexits 0. Accepted.loansuses the contract withconsumesemptiedv1broken in place, producer edited to matche2eFAIL [T1] Test sheet missingpassedwith no evidenceFAIL [T3] passed without evidence and the commit verifiedFAIL [K3] no deliverable namedout-of-cyclewith no justificationFAIL [K4] label out-of-cycle without its justificationThe two failures are one defect, and it is the most serious this validation has produced
The boundary rules are blind to a contract-based dependency — which is the only dependency this framework allows between modules.
boundaries.pybuilds the real graph from import lines. A module that consumes an HTTP contract imports nothing of its producer, so the graph stays empty. Three consequences, all now observed rather than argued:consumeswhile the module reads the contract is accepted by every check. The declared graph can lie freely.from catalog.listing import Toolpreceded bysys.path.insert(0, "../catalog/src")is accepted: the marker B2 looks for (/src/) is on a line that carries no import keyword, so the detector never reads it.docs/os/02-modules.mdand the fitness functions promise that the declared graph is compared with the real one. For the boundary this product exists to enforce, that comparison does not happen. Every NapkinStack project is affected, not this one.P4 — what catches a broken contract, and what does not
Renaming a required field of
v1in place, with the producer edited to match, is caught by no fitness rule.nstack fitnessexits 0. What caught it isnstack e2e catalog— D1's own scenarios, withKeyError: 'approximateLocation'— and only because those scenarios read the contract's own examples, which the probe left untouched. Had the examples been updated in the same pull request, nothing in CI would have refused it, and every consumer ofv1would have broken silently.This confirms what the contract rules already mark "to automate": compatibility of a contract version is checked by nobody. The producer's own tests are not a substitute — they pass by construction once the producer follows its own break.
Probes kept open
P1, P4, P5, P6, P7 and P8 are closed: the refusal was observed and is logged above. P2 (#22) and P3 (#23) stay open as live evidence of the defect, until the decider says what to do with them.
The plan requires a decision here: "a probe that passes when a refusal was expected: a defect in the register, and the maintainer decides whether to stop the validation."
napkinstack-agent commentedon Sep 18, 2026
Task 8 — the cycle closed,
completed, seven days earlyPull request #29 awaits the decider. The cycle's state once it merges:
closed,outcome: completed,ended_on: 2026-09-19.The playbook's first closure path applies — every deliverable accepted — so the circuit breaker was never reached.
P11, verified without being faked
The breaker needs a delivery pull request after the end date, and 2026-09-26 has not arrived. Moving the real cycle's dates to manufacture it would have corrupted the record of a validation whose whole value is that its records are true. The rule was reproduced on a throwaway copy of the tree instead, with the window moved into the past:
Deviation recorded: P11 is verified as a rule, not as a pull request in this project.
And the honest limitation: the circuit breaker was never exercised on the real cycle, because the work finished early. The most interesting case of Task 8 did not occur, and no amount of reasoning replaces having seen it fire on a real cycle. It stays unobserved until a cycle actually overruns.
What the closure records, and what it refuses to hide
Every clause of the charter's only success criterion holds, and none of its three misses occurred. Criterion A is met as it is written.
The closure also states, in the cycle file itself, what the green cannot say: Task 7 showed that two of the barriers the criterion counts on cannot refuse anything for a contract dependency —
B1for an undeclared one,B2for an implementation import reached through a path manipulation. Neither was bypassed here. They were never able to refuse.Both statements are true, they are recorded together, and the exit decision has to weigh them together.
napkinstack-agent commentedon Sep 18, 2026
Task 7 — P9 and P10, run by the maintainer as the App. Both refused by GitHub itself
Target: pull request #30, head
408be1b, CI green, deliberately unapproved, so that nothing but the rule under test could explain a refusal. Run from the session's shell prompt, because the agent's own attempt is stopped by its harness before the request leaves the machine (M8f) — and a refusal by the harness would have proved nothing about the forge.P9 — merge with no approval, as the App:
P10 — the App approves its own pull request:
Both refusals come from GitHub, and P9's names the missing rule rather than failing vaguely. ADR-0004's barrier holds on the validation project, under the ruleset a generated project receives — not only on the repository that develops the framework, where it was first observed in M8f.
Pull request #30 closed, never merged, branch deleted.
Task 7 complete — 11 probes
* P4 was a "to observe": no fitness rule refuses a broken contract version. It was caught by the producer module's own
e2e, and only because those scenarios read the contract's examples, which the probe left untouched.† P11 verified as a rule on a throwaway copy, the real cycle's end date not having arrived.
The two acceptances are one defect, already written up above: the boundary rules are blind to a contract dependency, which is the only dependency this framework allows between modules.
napkinstack-agent commentedon Sep 18, 2026
Task 9 — Findings report
Every criterion of What is measured, against its target. Written before the exit decision, so the decision is taken on the record rather than on an impression.
The criteria, one by one
loansmerged against the contract aloneThe sync events, counted honestly
Zero during the work. Three author sessions and three verifier sessions wrote, delivered and verified without one question they could not answer from the repository. D1: "I asked no question before writing." D2: "I did not need to come back with questions." D3: "Nothing was ambiguous enough to stop on."
One between the teams, and it deserves naming rather than rounding to zero. D1's author left an open question its own session could not settle:
contract_doublelives insidecatalog, soloanscannot reach it without crossing the boundary — does team 2 build its own double, or does a shared one need a home neither module owns? The decider settled it before team 2 started. That is a synchronisation between two teams, mediated by a human, and the framework produced it rather than prevented it. It was cheap here; it is the shape of the expensive one.PDR-0004's success criterion — met, well before 2026-10-31
It asked for at least three questions the framework should have raised and did not, each attributable to a stage that facts an existing check returns would establish. At least five were observed:
runverb — whileplaybooks/verification.mdtells the verifier to start the change with it. Both facts are in the manifest.consumesnon-empty with no import detectable: both computable.new-moduleprints four next steps, mentions none of it, and holds every fact needed to say so.docs/tooling-profile.mdclaims to be the only place a tool is named.The defects, ordered by what they cost
1. The boundary rules are blind to a contract dependency — the only dependency this framework allows between modules. Observed three times:
B1accepts an undeclared one (probe P3, every check green),B2is walked around by a path manipulation (probe P2,fitnessexits 0),B4warns on a real one (D3).docs/os/02-modules.mdpromises the declared graph is compared with the real one; for the boundary this product exists to enforce, that comparison does not happen. Every NapkinStack project is affected.2. No check refuses a broken contract version. Renaming a required field of
v1in place, producer edited to match, passes every fitness rule (probe P4). It was caught only by the producer's owne2e, and only because those scenarios read the contract's examples, which the probe left untouched.3. A human approval survives a force-push. No rule of the checklist requires GitHub's Dismiss stale pull request approvals; only the agent's own
--match-head-commitstood between an approval and content its approver never read. Closed on both repositories on 2026-09-18; the checklist gap stands. Candidate G13.4.
T2cannot tell a verifier from an author. TheVerifier:line is free text; its own message says "someone other than the author" and nothing enforces it. Charter C1 clause 2 rests on convention.5.
nstack new-moduleproduces a module that fails its own project, and says nothing about the sheet, the deliverable or the label it will need.checkandtestexited 1 on the generated module, measured. The field's convention is that a generator's output is green.6. PDR-0002's criterion measures wall clock, therefore the decider's availability rather than the framework's cost. Hit twice — Tasks 2 and 4.
7.
contracts/andplatform/are modules for one check and invisible to two others.8. The handbook contradicts itself on criticality; the framer had to arbitrate with no guidance.
9.
playbooks/framing.mdsays nothing about the technical framing — zero mentions of ADR, tooling, stack, architecture. The M9 plan had to require the tooling profile by hand.10. Technical decisions taken at framing have no artefact to land in, so the stack landed in the charter while
tooling-profile.mdclaims to be the only place a tool is named. A structuring decision can also land with no ADR:new-modulerecommends one, nothing requires it.11.
providesis never confronted with reality. Nothing checks that its path exists or holds a document.12.
pr-checkandpr-scopecount modules differently, and print contradictory lines in the same run.13. Two agent sessions on one checkout corrupt each other's branches. The framework describes parallel teams assuming separate humans on separate machines; nothing warns that "in parallel" may mean two sessions on one working tree. Both deliveries survived because the sessions repaired it themselves.
14. Two teams conflict on
.github/CODEOWNERSat their first commit, because the scaffolding appends its owner line at the same place.15.
uv.lockis neither committed nor ignored in either module, so transitive dependencies resolve fresh on every run.What worked, and it is not nothing
A full cycle ran end to end in one day against a one-week appetite: a discovery challenged by a session that had not written it, a charter and a cycle framed and accepted, two modules with separate owners, a hand-written contract provided by one and consumed by the other, three deliverables each with a test sheet written before the code and run by a session that did not write it, evidence machine-produced and re-runnable, and a closure recorded before the end date.
Nine of eleven barriers refused what they had to refuse, including the two that matter most and that only the maintainer could test: GitHub refused a merge without approval (405, naming the missing rule) and refused the App's self-approval (422).
And twice a barrier produced evidence rather than an obligation:
T4refused a sheet verified on another commit, which is what proves the author never touched the verifier's cells.napkinstack-agent commentedon Sep 18, 2026
Task 9 — the exit decision: fix first
Taken by the decider on 2026-09-19. The pilot waits for a NapkinStack version that closes a named set of defects. No GitHub Team seat is bought today; the organisation stays on the Free plan, as decided on 2026-09-17.
The last measurement, which completes PDR-0003's criterion
Median 10 minutes against a target of 15, on three measurements. PDR-0003's criterion is met, and no longer rests on a single observation.
Worth reading with its context rather than alone: each of those reviews stood on a sheet whose scenarios had been written before the code and run by a session that did not write it, and on evidence produced by machine. The ten minutes are what the sheet and the verifier bought. They are not ten minutes of reading 700 lines.
Why
fix firstrather thangoThe cycle ran end to end and every barrier protecting the human held: GitHub refused the merge without approval, refused the App's self-approval, and refused a sheet verified on another commit.
What did not hold are the barriers protecting the architecture.
docs/os/02-modules.mdpromises that the declared graph is compared with the real one. For a contract dependency — the only dependency this framework allows between modules — that comparison does not happen, and three probes showed it: an undeclared dependency passes, an implementation import passes behind a path manipulation, and a real dependency is reported as unused.Sending a real project onto that would be shipping the differentiator broken. Every competitor surveyed produces documents; this product's claim is that the rules are executable. A rule that cannot refuse is a document with extra steps.
Why not
rethinkNo decision proved wrong. PDR-0001, PDR-0002, PDR-0003 and ADR-0004 each did what they promised where they were implemented. They proved incompletely applied, which is implementation work, not design work. Reopening them would spend weeks to arrive back where we are.
napkinstack-agent commentedon Sep 19, 2026
M10h, step 1 — the update to v0.4.0, paid for real (#31)
NapkinStack v0.4.0 was published on 2026-09-19 to fix what M9 found here (D29 to D33, D35). M10h proves the fixes where they were found; this is its first step, and M9's Task 10: the first update of this project.
What the update asked for
nstack update, v0.3.1 → v0.4.0fitnessM7 oncontracts/, twice: v0.4.0 no longer ships placeholder verbs, andcontracts/holdscatalog-apiv1 — the adaptation the release notes announcecheck: the document is valid OpenAPI 3.1 (openapi-spec-validator, pinned) and the tests' Python passes the modules' ruff rules.test: every example is what its schema declares —loansbuilds its double ofcatalogfrom them, and the structural check lets aProblemexample withstatus: 200(minimum 400) through — and aProblemsays the status its response answers (RFC 9457 §3.1.3). 5 files, 114 lines, in C2's stackpr-checkthen askedcontracts, of criticality high — delivery work by the framework's own definition. K1: cycle 1 is closed and no cycle is framed, so theout-of-cyclelabel, with its justification.pr-scope: over the files budget (22/15),over-budgetdec9b511…, fresh context, launched from the author's session — said in the pull request) ran the sheet twice. The first run found a real gap: a 404 example whose body says 400 passed both verbs, while the double answers with the response's code and the example's body. It also foundreferencingunpinned, ruff's target version missing, and two commit hashes swapped in the description. All fixed in20d484a; the second run, S1 to S5, passed on it, and confirmed that the update's commit is Copier's merge unchanged by hand (same tree from a freshnstack update)Module checkshereChanged modulesfoundcontractsalone; its job installed the engine from the registry (napkinstack==0.4.0), rancheckandtest, andcompatsaid No frozen contract version changedCost: the update itself, nothing. The adaptation, one contract test and the manifest's verbs — and a sheet, since the module is of criticality high; that sheet is what caught the gap. The plan expected no label, no sheet: true of the update alone (measured before the release), not of the adaptation, which adds code to a high-criticality module.
Found
playbooks/framing.md, when only the next cycle is missing. The rule is right; its message sends a team back to framing when it only needs a cycle. For NapkinStack's register.Every touched module passedofmodule-checks.ymlstill says "touched" where v0.4.0 says "changed" for what CI lists. Cosmetic. For NapkinStack's register.git diff-tree --inter-hunk-context=-1, is refused by recent Git and it falls back to0. The merge is unaffected.Sync events
None.
State
#31: every check green,
BLOCKEDon the maintainer's approval only. After the merge:Module checksmade required onmain, andnstack doctorrun with the maintainer's own token.napkinstack-agent commentedon Sep 19, 2026
M10h, step 2 — four probes replayed on v0.4.0
Each a pull request, never merged, closed once logged here. Judged by CI on GitHub, and reproduced beforehand with the project's pinned engine (
napkinstack==0.4.0).loansimportscatalog's code behindsys.path.insert(0, "../catalog/src")Fitness functions— B2, twiceFAIL [B2] modules/loans/src/loans/catalogue.py:75(the path intocatalog's folder) and:76(the import)fitnessexit 0consumesemptied whileloansreads the contractFitness functions— B1FAIL [B1] modules/loans/src/loans/contract.py:17andmodules/loans/e2e/conftest.py:4, each with the exactconsumesentry to addloanswhose sheet is signedVerifier: session <id>by the session its commit'sAgent-Session:trailer namesTest sheet and cycle— T2FAIL [T2] Test sheet: the verifier, session d9d08c03-…, wrote commits of this change (Agent-Session trailer).Every other check green: T2 alone refused itVerifier:line was free textnstack new-module reminders NapkinStack/maintainers standard --user-facing, alone, untouchedChanged moduleslistedreminders, whose job said nothing to check yet;pr-checkTest sheet: not requiredcheckandtestexited 1Beside the target, on P2:
loans'checkrefused the probe's unformatted lines (ruff), andTest sheet and cycleasked for a sheet and a cycle (T1, K1) — a real change of this shape would meet three barriers, not one.P14's next steps, as
new-moduleprints them, now name the sheet, the cycle and M7 — M9's D33 asked for exactly that. One of them repeats K1's wording already reported (step 1): "The project is not framed", when only the next cycle is missing.Still to run
v1broken in place, producer edited to match) and P4b (an optional property added tov1) — once contracts — a frozen contract version is changed only when oasdiff proves it compatible #32,contracts'compat, is merged: V1 reads the proof as merged.Module checksred, the pull requestBLOCKED) — once the maintainer has madeModule checksrequired onmain.Found, for NapkinStack's register
From #32 (
compat):bootstrapinstalled — can be rewritten by the very change it judges. contracts — a frozen contract version is changed only when oasdiff proves it compatible #32's command therefore carries everything itself. The framework should say so, or run the proof from the base's tree.compatcommand is printed twice, in the->line and inside theFAIL [V1]message, burying the comparator's own verdict.Sync events
None.
napkinstack-agent commentedon Sep 19, 2026
M10h, step 3 — P4, P4c and P4b replayed on v0.4.0, after #32 and #33
Each a pull request, never merged, closed once logged here. Judged by CI on GitHub, and reproduced beforehand with the project's pinned engine (
napkinstack==0.4.0).main= 87bb76b:contractsdeclarescompat(#32), the modules' locks are committed (#33).approximateLocationrenamedwhereaboutsinrequiredand in one example,catalogedited to matchModule checks— V1, oasdiff naming the propertycontracts'test(#31) refused first: the two examples no longer match their schema ("'whereabouts' was unexpected", "'whereabouts' is a required property"), and the job stopped beforecompat.catalog'stestrefused too (1 failed). V1, run locally on the same commit:FAIL [V1] catalog-api v1 (consumed by loans), oasdiffresponse-property-became-optionalonapproximateLocation, twicee2erequiredand every example, and incatalog's code and tests;loansuntouchedcontractscheck and test (8 passed),catalogcheck, test (8) and e2e (7 + 2 subtests),Fitness functions: all green.FAIL [V1] catalog-api v1 (consumed by loans), oasdiffresponse-required-property-removedonapproximateLocation, twice. Locally on that commit,loans'test:KeyError: 'approximateLocation', 2 of 15 failed — the consumer broken, and no job of its own raninstance, RFC 9457's own member, added toProblemin v1, optionalModule checksgreen: V1 proven compatibleModule checksgreen: "Contract version catalog-api v1 (consumed by loans): proven compatible."Beside the target. #38 and #39:
P1(two modules),T1(a sheet:contractsis high criticality,cataloguser-facing) andK1. #40:T1andK1— a real compatible change to the contract still carries a test sheet, and a cycle or theout-of-cyclelabel.Found, for NapkinStack's register
From the probes:
compathides V1's verdict.module-checks.ymlrunscompataftertest, with no condition: on P4,contracts'testfailed andcompatnever ran. Nothing gets through — the pull request is red either way — but the refusal that names the real problem (a frozen version broken) arrives one round trip late, after the author has fixed the examples. Runningcompatbeforetest, or underif: ${{ !cancelled() }}, would show it at once.contracts'test(NapkinStack v0.3.1 -> v0.4.0, and the contracts module's check and test #31) catches examples that contradict their schema. P4c is the probe that does, and V1 alone refused it.From #33 (the locks, merged as 87bb76b after four verifier rounds on four heads — each round found something real, fixed before the next):
--lockedsays nothing aboutuv run --with: resolved in a layer outside the lock. The verbs now use dependency groups.contracts'teststill runsuvx --with— the same unlocked layer, with no project of its own to lock.uv runnever removes an extraneous package, and CI's verbs share one environment: a verb could pass on another group's packages. The verbs now run--exact.setup-uvpins nouvversion; with--locked, auvrelease that writes locks differently would turn the modules red with no change in the project. The workflow is the framework's.uvdownloaded the newest, 3.14.7, while the charter says 3.13: each module now carries a.python-version.catalog's README says its start commands are checked on every run; the scenarios start the same invocation from thee2einterpreter, not throughuv run, and CI never runs therunverb (predates catalog, loans — the lock committed, and everyuv runrefuses a stale one #33).Still to run
Module checksred, the pull requestBLOCKED, read through the API, no merge attempted.Module checksis a required check ofmain(read through the rules API).Sync events
None.
napkinstack-agent commentedon Sep 19, 2026
M10h, step 4 — P13 replayed on v0.4.0
A pull request, never merged and never approved, closed once logged here. Reproduced beforehand with the project's pinned engine (
napkinstack==0.4.0).loans: the guard refusing a return date before the hand-over date removed fromRecords.recordModule checksred; the pull requestBLOCKED, read through the API, no merge attemptedloans'test:FAILED tests/test_recording_a_loan.py::test_a_return_date_before_the_hand_over_is_refused_and_nothing_is_recorded— DID NOT RAISERefused, 1 failed, 14 passed;Module checks: "Changed modules: success. Their checks: failure." Read through GraphQL on head82244cd:mergeStateStatusBLOCKED,Module checksFAILURE,isRequired: true— a failing required check, which an approval does not lift. No merge attemptedModule checkswas not a required check (D35)Beside the target.
Test sheet and cycle:T1(loansis user-facing) andK1.checkandFitness functionsgreen: the regression is well-formed code within its boundaries — only the module's own test knew the rule. Locally,loans'e2erefuses it too (1 of 4), but in CI the job stops attest, socompatande2enever run — the same ordering as P4's.Required checks of
main, as the API reads them on #41:Fitness functions,PR scope and review budget,Hooks and secrets,Test sheet and cycle,Module checks.Changed modulesand each module's own job are not required —Module checkscarries them, since a matrix job's name cannot be required in advance.Every probe replayed
P2, P3, P4, P4b, P4c, P12, P13 and P14 are run on v0.4.0. The closing comment compares each with M9 and sums up the update's cost.
Sync events
None.
napkinstack-agent commentedon Sep 19, 2026
M10h — closed: every probe against M9, the update's cost, the sync events
NapkinStack v0.4.0 was released to fix what this project exposed in M9 (D29 to D35). M10h paid for the update here and replayed the probes. Everything below is linked from the steps above: #31 (step 1), #34–#37 (step 2), #38–#40 (step 3), #41 (step 4).
The probes, v0.3.1 → v0.4.0
loansimportscatalog's code behindsys.path.insertfitnessexit 0Fitness functions, B2, twiceloansreads the contract withconsumesemptiedFitness functions, B1v1broken in place, producer edited to match, examples left behinde2econtracts'test(the examples) andcatalog'stest, before V1 could run; V1 refuses it locallyapproximateLocation;loansbroken, no job of its own ranv1Verifier:was free text (D32)Test sheet and cycle, T2Module checks, a failing required check (isRequired: true,BLOCKED)nstack new-module --user-facing, alonecheckandtestexit 1 (D33)Eight of eight did what v0.4.0 promises: the six that must be refused were refused, the two that must pass passed. Of M9's eleven probes, the three not refused by a rule — P2, P3 and P4 — now are. The eight it refused rest on rules v0.4.0 did not loosen; P1 (#38, #39) and T1, M9's P5 (#38 to #41), refused again beside the target.
The forge, checked by the maintainer with their own token after the update —
nstack doctor: L1 to L7 and G1 to G13 compliant,Module checksamong G4's required checks, G13 (stale approvals dismissed) enforced.The update's cost
nstack updatev0.3.1 → v0.4.0 — Copier's merge, no conflict, nothing by hand — and the adaptation it asked for:contracts'checkandtestcontractsdeclarescompat: oasdiff 1.32.1, pinned by checksum, self-containeduv runlocked and exact, Python 3.13Wall clock includes the wait for the maintainer's approvals. The update itself cost nothing. Its adaptations are what v0.4.0 asks of a project holding a consumed contract:
check,testandcompatoncontracts. The locks were not asked by the update but by M9's own finding. The maintainer's part: three approvals,Module checksadded to the ruleset, onenstack doctor.Deviations from the plan, each recorded where it happened: #31 needed a sheet and two labels (the plan said none — the adaptation adds tests to a high-criticality module, and the sheet caught a gap);
compatis self-contained rather than installed bybootstrap(V1 runs it in the pull request's tree); P4c added, because P4 no longer isolates V1; M9's P1 and P5 to P11 not replayed, their rules unchanged.For NapkinStack's register
compatcommand in the pull request's tree: a proof calling a file of the module can be rewritten by the change it judges (contracts — a frozen contract version is changed only when oasdiff proves it compatible #32).compatcommand is printed twice, burying the comparator's verdict (contracts — a frozen contract version is changed only when oasdiff proves it compatible #32, Probe P4c — v1 of the contract broken consistently, examples and producer's tests included #39).compathides V1's verdict:module-checks.ymlruns it aftertest, with no condition (P4, P13).new-module's next steps repeat it (NapkinStack v0.3.1 -> v0.4.0, and the contracts module's check and test #31, P14).setup-uvpins nouvversion: with--locked, auvrelease can turn a project red with no change in it (catalog, loans — the lock committed, and everyuv runrefuses a stale one #33).uv runrefuses a stale one #33 (catalog, loans — the lock committed, and everyuv runrefuses a stale one #33).contracts'testrunsuvx --with: what the pins pull in is unlocked (catalog, loans — the lock committed, and everyuv runrefuses a stale one #33) — the project's, reported here for the pattern.Every touched module passedstill says "touched" (NapkinStack v0.3.1 -> v0.4.0, and the contracts module's check and test #31). Upstream and harmless: Copier's "Git >= 2.24" warning on Git 2.55 (NapkinStack v0.3.1 -> v0.4.0, and the contracts module's check and test #31).Sync events
None during M10h: no question had to be settled in conversation. The maintainer's part — three approvals, the ruleset,
doctor— was the plan's, taken on what the pull requests and this log said.State
main= 87bb76b, no pull request open, every probe closed and its branch deleted. Next, inNapkinStack/engineering-os: the M10 closure — the plan ticked with what was observed, the register above — and the maintainer's exit decision.napkinstack-agent commentedon Sep 19, 2026
M10's exit decision: fix first (2026-09-19)
Taken by the maintainer, decider, on the closing comment above.
Why. Every probe replayed in M10h did what v0.4.0 promises, and
nstack doctoris compliant. But one finding is of the class the v0.4.0 tag was held for (D36 to D39, a barrier that can be walked around): V1 runs the merged proof in the pull request's tree, so a project whosecompatcalls a file of its module — a script, a binary itsbootstrapinstalled — can have that proof rewritten by the very change it judges. This project is not exposed: its command carries everything itself (#32). A pilot project would be, from the day it freezes its first contract version.What. A v0.4.1 in which V1 cannot be walked around this way, and its verdict is never hidden by an earlier failing step nor worded as a breaking change when the proof could not run. It is planned in
NapkinStack/engineering-os; the M10 closure pull request places every other finding of the closing comment.Then. The organisation moves to GitHub Team and the private pilot starts on v0.4.1.
Sync events
None.
napkinstack-agent commentedon Sep 20, 2026
M10h's last entry — NapkinStack v0.5.0, and what the update cost this project
NapkinStack v0.5.0 was published on 2026-09-20 to close the defect M10h itself exposed: V1 read the proof as merged but ran it where the change is (D44). This is what the version cost this project, and what it proved here.
The update: nothing was asked
nstack update, v0.4.0 → v0.5.0checkandtestwritten before it could pass; this one merges and passes as it standspr-checkaskedout-of-cyclelabel was added for K1, cycle 1 being closedcommits-on-main.yml, new; thecompatstep underif: ${{ !cancelled() }}(D46) and its closing step saying "changed" (D47); the handbook on where a contract proof runs (D44) and on what remains when the forge can enforce nothing (PDR-0006); the README's two words, guarded and unguardedWhat v0.5.0 was for, replayed here
D44 — a proof that judges the change cannot be rewritten by it. Replayed on a clone of this project before the update branch was cleaned:
catalog-apiv1 broken consistently — the property renamed in the schema, the examples and the producer's code — is refused.Two differences from v0.4.0, both visible above: the comparator's own output is above the verdict instead of being buried under a command repeated whole (D45), and the verdict says the proof did not prove compatibility rather than asserting a break it may not have observed. This project's
compatcommand is self-contained, which is why the change to where it runs asked nothing of it — a command that had called a script of its own module would have had to become self-contained (D44).The record — its first real run, on this repository. The merge of #42 pushed to
main, andCommits on mainran for the first time:That is the designed behaviour, seen for real: the commit that brought the record is the one commit it says nothing about (PDR-0006). Its first firing on a commit pushed outside a pull request has not happened yet, and will be a probe, logged here like the others.
Found, in real use
nstack updatewith an engine older than this project refused correctly — "use nstack 0.4.0 or newer" — and the maintainer then trieduv tool upgrade napkinstack, which does nothing at all on an install pinned to an exact version. Invariant P6 asks for the action; an action nobody can copy is a description. Fixed in NapkinStack (D50): the refusal now carriesuv tool install "napkinstack@latest" --with-executables-from pre-commit.nstack updatecreated on the maintainer's own checkout was authored by the machine's global identity, unrelated to this project; on a public repository that identity would enter the history. The branch was re-authored under the App before being pushed. Nothing to fix in the framework — a note for whoever runs an update on a shared workstation.Sync events
None.
State
main= 36f1adf, on NapkinStack v0.5.0. No pull request open, no probe open. This project has now been carried across two framework versions, v0.3.1 → v0.4.0 → v0.5.0, the first at the cost of one adaptation and the second at no cost at all.