Repository navigation
FIX-2087: closure QA for one task primitive: the walk, the sweep in CI, and 13 findings - #3692
Conversation
goals/workforce-coordination/runs-on-one-task/ runs the closure's plan on one main commit: the concept's walk (leg a, with journeys j1 and j2), the server killed after an ending (b) and a task failing out (c), on a real model, on a Lab of WORKER.md files served by Shift Manager's command; the task-surface sweep; every child's check with its controls; and the 31 in-scope gaps rows' map. Each assertion is tagged structural or judgment, and a structural miss on any attempt is never retried past (D1). The sweep is one rule engine (rules.mjs, engine.mjs) the goal and CI share (D2): the tools rule reads every tool the task-tools capability declares, the removed rule reads the epic's migration table as it stands, and the endings rule reads callers of a task collection's complete and fail, and claim-passing cancels, by type through the TypeScript program. --control plants a sixth tool, a removed name, a write through a renamed variable and a claim-passing cancel. It joins CI's Typecheck job, which installs the TypeScript it reads with. FIX-2078's and FIX-2080's controls move to a controls.mts beside their goals, so the closure reuses them rather than copying them. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
|
The latest updates on your projects. Learn more about Vercel for GitHub. 3 Skipped Deployments
|
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: 609f5384-5981-4adb-8824-dcf71d9cade7) |
jhoffner
left a comment
There was a problem hiding this comment.
Stand-in review (Bugbot is out of usage). Lenses: coherence, restraint, correctness, completeness against specs/issues/FIX-2087/. I did not count the sweep's intended red (FIX-2119) as a finding.
Verdict: one blocking issue. The CI gate's tools rule can be dodged by renaming the capability's local defineTool alias, and its control plant doesn't go through the real path. Everything else is a fix or a nit.
What I ran (scratch worktree on 63da04f, no real-model legs):
sweep.mjs --rows: FAIL, 9 live doc lines (FIX-2119). It reads 7 tools, 14 ending writes (all by recorders) and 48 migration rows, and every row is covered. This matches the report.sweep.mjs --control: CONTROL OK, exactly 4 plants.- 13 extra in-memory plants run through
sweepTaskSurface. Results are in the inline comments. The sweep caught: a direct pass-through call, a claim cancel via spread, and a removed name re-exported under an alias. It missed: the alias rename, ahandlertool in a sibling file, a renamedhandlerimport, a renamed or re-exported pass-through, destructured / element-access /.callendings, a spoofed recorder name, and a tool name built from a template string. - Moved controls:
controls.mtsin both goals is byte-identical to the inline blocks they replace, and the imports are still used.modulePatchEnvappliesKILL_AFTER_ENDINGandno-cascade.no-sweeper-paymentthrows: its anchor has been gone since 0d24ee6 (FIX-2082). The problem predates this PR, and the report already files it as FIX-2118, so the move changed no behaviour and the report matches.
Findings
- 🔴
engine.mjs:142: tools are matched by callee spelling. Renaming the alias reads 2 of 7 tools, and a sixth tool passes with no CAN'T TELL. The plant imports adefineToolthat core doesn't export. - 🟡
engine.mjs:242and:231: endings miss renamed or re-exported pass-through helpers, destructuredcomplete, element access, and.call. - 🟡
ci.yml:101: the--controlstep is skipped whenever the sweep is red. It was skipped on this PR's run, so CI never checks the plants while the sweep waits on FIX-2119. - 🟡
run.mts:368/:217: "no routed task" is tagged judgment in c and j1 but structural in a, so a routing miss can be retried past, against D1. - 🟣
engine.mjs:219: a recorder is any function with that name in the package. - 🟣
rules.mjs:75: thegoalsbranch ofSWEPT_TEXTis dead because of thegoals/exclude.
Report vs evidence: the sweep numbers, the control results and the FIX-2118 control break all reproduce. Two claims overstate the evidence:
- D2's "a tool built apart can't escape it" and the body's "a renamed variable can't hide one" hold only for receiver aliasing (🔴 and 🟡 above).
- "its
--controlstep PASS" is from the local run. In CI that step was skipped.
Generated by Claude Code
|
CI on 63da04f: the only red step is One task surface, in the Typecheck job. The "Typecheck & Test" summary fails only because it mirrors that step. Every other step is green. That step is the closure's new sweep. It is red on purpose: it reports the 9 lines of published docs that still teach removed task surfaces (FIX-2119), which is the finding this PR exists to surface. The fix belongs in FIX-2119's own PR, which is queued. It is not fixed here because this PR changes no product code or docs. I'll bring Generated by Claude Code |
… run in CI Review 5626363033 on #3692. - tools: a tool is a call of core's handler, read by symbol (a renamed import, a namespace member, a local alias, a .withDefaults result), in the capability's file and the files it imports from its directory. Every allowed name must be read, or the sweep can't tell (exit 3). - endings: pass-through helpers resolve by aliased symbol (renamed imports, re-exports, namespace calls). A collection's complete or fail read by element access, destructuring, .call or .bind is an ending. Each recorder, pass-through and collection wrapper key must name exactly one function. The three collection wrappers the wider read found (workforce's guarded, byRow, notOnAStep) are FORWARDERS: their callers are read in their place. - --control plants each of these through the real path (tools in a copy of the capability and in sibling files it imports), and a floor plant must read CAN'T TELL. CI runs it even when the sweep is red. - run.mts: no routed task is structural in b, c and j1 (b:routed, c:routed, j1:routed); b:files-research is the model's choice, not b:killed. - rules.mjs: the dead goals branch of SWEPT_TEXT goes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: 1fb54dde-72bd-4915-a9cc-521a26940c4a) |
jhoffner
left a comment
There was a problem hiding this comment.
Delta review of 63da04f00..f6d06f135 (stand-in for Bugbot). Nothing here is blocking. Two 🟡 are worth fixing before merge. Each is a small change.
What I ran, in a scratch worktree at f6d06f135, through sweepTaskSurface with in-memory plants:
sweep.mjs --rowsgives exactly the 9 FIX-2119 doc lines (8 forTaskCollection.block, 1 forparkOnWait'stimeoutMs), withtools 0 · endings 0and no CAN'T TELL. ✔sweep.mjs --controlgivesCONTROL OK: 13 findings from 12 plants, no strays, and the floor reports CAN'T TELL. ✔- The six earlier threads, re-probed. All closed:
- The capability's local alias renamed to
makeToolplus a sixth tool, a renamedhandlerimport, an imported sibling-file tool and a template name are all reported. - A renamed pass-through import and a renamed re-export are caught.
const { complete } = rows, a destructured parameter,rows["fail"],rows[k],.call,.bindandrows?.failare all caught.- A duplicate
releaseRowand a duplicateguardedeach report CAN'T TELL. - The CI
--controlstep now hasif: ${{ !cancelled() }}. - The
run.mtstags are split the way that thread asked:c:routed,j1:routedandb:routedare structural,b:files-researchis judgment. Untaggedsweep:cant-telldefaults to structural, which is correct. goalsis gone fromSWEPT_TEXT.
- The capability's local alias renamed to
Your questions:
- Is the FORWARDERS list justified and narrow? The list is: three functions, four forwarding sites, and each returns a
TaskCollectionRefwhose callers are read. The exemption isn't: it covers the whole function body. A directref.fail()planted inguardedwas caught at63da04f00and is silent atf6d06f135(🟡, inline). - Can the floor pass vacuously? Not when the rule stops reading: the floor control proves it fires. But it only checks that the seven names were read. A sixth tool built as a sequencer or generator (both valid
GeneratorTools) gets past the rule and the floor alike (🟡, inline). - Is a forwarder's caller read soundly? Yes. The returned wrappers are typed as collections, and
guarded'sownClaimonly strips a foreign claim. It never adds one, so the cancel/claim read at the call site stays conservative. - Does
--controlreport every plant and nothing else? Yes. Its exact-count and stray checks hold.
🟣 Two shapes of pass-through are still missed: destructured from a namespace, and passed as a value. Neither is likely (inline).
Generated by Claude Code
…builder makes a tool Delta review 5626603412 on #3692. - endings: in a collection wrapper on FORWARDERS, only the shapes that hand a write on are exempt (a .bind/.call/.apply receiver, an element access by a non-literal key, a property access not called directly). A direct complete()/fail() call inside one is a finding again. - tools: any of core's five block builders (handler, generator, sequencer, router, evaluator) builds a tool, since a generator's tool is any block. - pass-throughs: a destructured binding is followed, and a helper handed on as a value is an ending. - --control plants a direct write inside guarded, a sequencer- and a generator-built tool, a destructured and a passed-on pass-through. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: 5421fb3a-2292-4691-87b4-0a26b4bcd728) |
…ard-once FIX-2112 (#3697) stops the second notice turn when a task ends before its filer's getTask wait. The ending still starts one onTaskSettled request, which finds the notice no longer owed and runs no turn (output { owed: false }). j1:read-once now counts only the requests that admitted the notice (owed: true), as goals/task-board/waits-on-a-task-it-filed does. A new structural j1:heard-once holds that to its claim: the ending is heard exactly once, by editor's getTask read in one request, and by no notice turn. The coordinators store reader carries each request's result output. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: f853ba04-ca34-4af2-affe-614c50545db3) |
Brings in FIX-2114, 2116, 2118, 2119, 2120, 2121 and 2124. The every-ending-reaches-its-filer conflict keeps the branch's controls.mts import and ports FIX-2118's re-anchored no-sweeper-payment patch into it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
|
The full real-model closure run waits until FIX-2097, 2113, 2115, 2117, 2122, 2126 and 2127 land. Generated by Claude Code |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: 0a4745f9-1f1b-4919-ae93-ebdce47d47b2) |
The closure's findings (FIX-2097, FIX-2110 to FIX-2122) are children of FIX-2056 blocking FIX-2087, so row 3 reads them from Linear and failed each as "has no check here". Each now runs its regression tests, and names the run elsewhere in the closure that grades it end to end (`by`: a goal some entry runs, a row-1 assertion, or the sweep). A `by` that names no such run is a failure, so a stale reference can't pass. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: 09c9f2ea-ae43-4ca7-bff4-9d07eb02b485) |
FIX-2123 made a notice's admission a claim: a notice delivered twice starts
a second onTaskSettled request, which finds it taken and runs no turn
(`{ owed: false }`). a:one-settled counted every request and failed on one
such duplicate. a, b and j2 now count the requests that admitted a notice,
as j1 does since FIX-2112, and still want exactly one per ending.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
There was a problem hiding this comment.
Simplify review (closure QA + task-surface sweep)
What this PR is solving: FIX-2056 needs more than “re-run every child’s goal on one commit” — none of those walks a real routed hand-off on a live model or proves the gaps between children. This adds the closure runner (runs-on-one-task) plus a shared static-analysis engine so CI and the goal cannot disagree on the task surface (D2).
Did the approach go too far? Partially, but mostly in carrying cost, not in wrong direction.
| Alternative | Verdict |
|---|---|
| ESLint / knip / semgrep only | Would not reliably catch the plants this PR documents (renamed handler imports, named() tools, claim-passing cancel, pass-through recorders). You would still end up with bespoke rules, without a single program context. |
Drop the TS program; regex-only like one-word-per-thing |
Cheaper CI, but would not satisfy the spec’s endings/tools rules or D2’s “read by type” bar. |
One mega run.mts without engine.mjs / rules.mjs split |
Simpler file count, worse maintainability — the vocabulary goal already separates data (rules.mjs) from driver (sweep.mjs); this PR extends that pattern correctly. |
| What I’d actually simplify next (optional follow-ups, not blockers) | (1) Reuse one Program/TypeChecker across plain + planted + floor passes in sweep.mjs — today --control builds the program three times, and CI runs plain + control as two Node processes (~4 full programs per Typecheck job; ~17–80s sweep-only depending on runner). (2) Factor shared sweep plumbing with goals/vocabulary/one-word-per-thing (ROOTS/REFUSAL_MARKER are already shared; EXCLUDE + asset-extension regex are duplicated and will drift). (3) Consider ts-morph only if engine.mjs keeps growing — it shrinks boilerplate, not rule count. |
What already looks like the right kind of simple: extracting controls.mts from the task-board goals so the closure imports patches instead of copying regex anchors; placing the goal under workforce-coordination/ (epic-wide, not a single board behaviour); keeping the sweep CLI as the CI entrypoint while engine.mjs holds the logic.
Approval: I don’t see a smaller design that still meets the spec’s D2 surface gate and the closure plan. The remaining simplify work is performance and shared goals/lib sweep helpers — worth tracking, not worth blocking merge.
Verified locally: node goals/workforce-coordination/runs-on-one-task/sweep.mjs → PASS (~17s on this runner).
Sent by Cursor Automation: Review For Best Approach
|
@codex review |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
…ow 3 (FIX-2087) FIX-2147 runs the privacy goal's model-free check-talk.mts and is graded end to end by that goal, which row 3 already runs. FIX-2149 and FIX-2150 are park-window races no real-model leg lands in, so their regression tests are their whole check (no by). Note in FIX-2078's and FIX-2080's controls.mts that the closure imports their anchors. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
There was a problem hiding this comment.
Code Snob. A shorter shape of the same contract is on draft #3790 (poc/shorter-one-task-notes @ 083cef8a6).
Fold 083cef8a6 onto cde0a95ae. Do not reset onto the POC. Close #3790 unmerged after this PR lands.
Net −60 (36/96). Comments that restated goal.md or the next assertion. run.mts's header drops the parts, D1, and controls paragraphs (goal.md already says them; TAGS and attempts still state D1). The FIX-2112 note stays, in four lines: an ending before the getTask wait still starts a request, and { owed: false } ran no turn. j1 counts that through admittedRequests instead of a second copy of the owed === true filter.
TOOLS_ALLOWED is a set of the same seven names, same order. The reason strings were never read; they sit on those lines as comments. The rules.mjs header no longer repeats the export notes. The kind list stays on REMOVED. A row with neither a matcher nor a goal still fails, and each recorder key still names one function. The tools note in engine.mjs still says five builders, read by symbol, same directory only, and an unread allowed name is CAN'T TELL. CI still says why this job needs TypeScript, and that the plant step runs when the sweep is red. It no longer lists every plant.
sweep.mjs --control on 083cef8a6: CONTROL OK. Plain run 0 live lines, 18 planted lines from 15 plants, floor CAN'T TELL. The report matches cde0a95ae aside from the commit line. Red: deleting getTask from the set reported one tools finding and nothing else. eslint exit 0. The walk was not re-run; admittedRequests is the filter j1 already used.
Left as recorded: the plants, the symbol reads, forwarders, recorders, removed matchers, the walk, children, gaps, the verdict log, goal.md, D1, D2. FIX-2101 stays known-open.
Tried and left: one AST walk for declarations and endings (the duplicate-name check still has to finish first). A shared unwrap (same length). Turning RECORDERS into a set (one distinct reason per key).
Sent by Cursor Automation: Code Snob
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cde0a95ae7
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
… (FIX-2087) Cherry-picked from poc/shorter-one-task-notes: comments that restated goal.md go, j1 counts admittedRequests, TOOLS_ALLOWED is a set of the same 7 names, and the rules.mjs, engine.mjs and ci.yml headers are shorter. Two trimmed comments are kept, since they state facts written nowhere else: buildsTool's (it reads local variables and conditionals) and the addTask matcher's (the collection's addTask method stays). Co-Authored-By: Jake Hoffner <jhoffner@users.noreply.github.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
…ow-up as judgment (FIX-2087) From Codex's review of #3692: - row 3 fails (children:cant-tell) when the live Linear set can't be read: no key, an HTTP error, GraphQL errors or no relations in the answer; - a closure control that leaves its tag green fails the run (control:broken); - the sweep reads a claim-passing cancel through .call, .apply, element access and destructuring, and counts one handed on by .bind; --control plants each shape, and the same shapes with no claim, which must not be reported (17 plants, 23 lines); - row 4 passes a gaps row only when every check it names ran in the same closure run and passed (a goal, its named controls, a test file, a row-1 assertion whose leg passed, the sweep); a stale or unrun label fails. From round 4's walk: writer's follow-up turn wrote the original draft word for word, so the two result lines read alike and a:lands-once, a:once and a:follow-up failed on a correct run. a:follow-up now wants a turn of writer's after the follow-up was filed; the new version is judgment (a:follow-up-differs); lines that read alike are told apart by when each landed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cde0a95ae7
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
… every worker flow (FIX-2087) From Codex's second review of #3692: - row 3 runs hands-each-post-to-its-delegates' six controls and split-to-roster's hold-session; - a control passes only when every tag its goal says it must fail is red (a bullet of its FAIL list starting with the tag; a leg tag, any bullet in that leg), not one of them; the lists now match each goal.md; - leg c fails c:writer-ends (structural) when writer's task never ends, then wants its one line; - j1:heard-once counts every completed getTask read, two in one request included; - the door rule reads every core defineFlow (by symbol, under a cast) that declares respond or whose configSchema is typed by the worker contract, computed keys resolved; it must read the agent's and the coordinator's flows or it can't tell. --control brings back a removed door in each (19 plants, 25 lines); the plain sweep reads 6 worker flows, 0 live lines. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
|
Architect: not approving yet. Several Codex P1s on the closure harness look real, chiefly that the live child lookup passes when Linear fails ( — FSD Architect |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: cde0a95ae7
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…three controls up to date (FIX-2087) From Codex's third review of #3692, and round 4's row 3: - a control is red only when its goal exits 1, every required tag has a FAIL bullet of its own (the whole tag, bare or bracketed), and no bullet is the goal saying a control of its own didn't go red. A diagnostic like 'control no-plan: a1:x did NOT fail' names the tag but is no red state. - in a forwarder, ref.fail.call(...) and ref.complete.apply(...) are writes made there, not forwarding; --control plants both inside guarded (19 plants, 27 lines; the plain sweep is still 0 live lines). A plant's findings are matched by path, not substring. - j1:tools fails on any tool offered beyond the allow list and the roster's two own tools, not only one named like a task tool. Controls out of date (the goals, not the product): - files-tasks-down-the-owners-chain no-follow-up: Workforce's conversation board maps the notice decision to a line itself for a turn-filed and an unrouted failure (FIX-2115, FIX-2122), so leg e's failure was still said. The control now patches both maps; e:one-notice red 2 of 2. - a-worker-plans-its-own-steps no-plan: a1:no-child-session needs the planless model to file for itself, which it no longer does (4 of 4); only a1:steps-in-task-session is required. - hands-each-post-to-its-delegates no-self-choice: best fit places the project ask on the EM now, not the hire (4 of 4); any of leg h's own-job asks counts. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
… j1:tools reads (FIX-2087) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A


The problem
Every child of FIX-2056 passed its own check on its own commit. Nothing had used the assembled set the way a person would: hand work to a Workforce app on a real model, wait on it, steer it, and hear how it ended. The closure does that on one
maincommit. Round 1 (1477834) filed 13 findings. Round 2 (e07f360) found the late cascade (FIX-2136). Round 3 (af27e48) found the privacy goal misreading a routing coordinator (FIX-2147). This is round 4, onmainat ce2282b, with FIX-2147, FIX-2149 and FIX-2150 in. The whole plan ran on b6516a8; the head adds only the verdict log.On the final commit, everything passes except one leg. Two things are open:
in_progressafter the restart. The worker's notice is then held for a request that will never end, and it doesn't hear its task's ending in time. Leg b failed on this once (structural, not retried). It's described below.j:names-ticket, 1 of 1). It waits on a product decision.What this does
It adds the closure's goal and a permanent CI guard. No product code changes.
goals/workforce-coordination/runs-on-one-task/runs the spec's plan. It starts a Lab: the concept's cast asWORKER.mdfiles on the built-in flows, served by Shift Manager's command. On that Lab it runs leg a (the walk) with journeys j1 (a worker waits on what it filed) and j2 (Alice steers one routed task and cancels another), leg b (the server killed after an ending), and leg c (a task fails out). It also runs the sweep, every child's check with its controls, and the 31-row gaps map. Each assertion is tagged structural or judgment. A structural miss is never retried past (D1).rules.mjsplusengine.mjs, one engine for the goal and for CI, D2) joins CI's Typecheck job. It fails a later PR that offers a sixth task tool, brings back a surface the migration table removed, or records an ending outside the claim gate or the drain. What it reads, exactly:.withDefaultsresult, in the capability's file and the files it imports from its own directory. It does not read a tool imported from another directory. If any allowed name goes unread, it reports CAN'T TELL (exit 3) rather than passing.completeorfailcounts in any shape, and so does acancelunder the caller's own claim (called,.call,.apply, element access, destructured, or handed on by.bind). In a collection wrapper only a reference handed on forwards; a call,.callor.applythere is a write. Each recorder key must name exactly one function. Not covered: a write through a value typedany.defineFlow(by symbol, under a cast) that declaresrespondor whoseconfigSchemais typed with the worker contract's keys is a worker flow, with computed action keys resolved. It reads 6 today and must read the agent's and the coordinator's, or it can't tell.--controlplants each of these shapes through the real path and must report exactly those; a claim-free cancel in the same shapes must not be reported; a floor plant must produce CAN'T TELL.controls.mtsbeside their goals, so the closure reuses them. Each header now says the closure imports its anchors.by, the run here that grades it end to end. A stalebyfails. If Linear's live set can't be read, row 3 fails. A control passes only when its goal exits 1, every tag its goal.md says it must fail has aFAILbullet of its own, and the goal prints no diagnostic that one of its controls stayed green.What's asked of you
One finding to file (below, under The report). FIX-2101 stays known-open.
FIX-2123, FIX-2135, FIX-2136, FIX-2138, FIX-2147, FIX-2149 and FIX-2150 are children of FIX-2056 but not marked as blocking FIX-2087 in Linear. Row 3 runs them anyway. Add the blocking relation if you want Linear to show them gating the closure.
Parts worth reviewing closely
children.mts:controlVerdict(how a control's red is read), the control tag lists against each goal.md,CHECKSfor FIX-2147, FIX-2149 and FIX-2150, and the fail-closed Linear read.coordinators/files-tasks-down-the-owners-chain(no-follow-upnow patches Workforce's own decision maps too),task-board/a-worker-plans-its-own-steps(no-planrequires onlya1:steps-in-task-session), andcoordinators/hands-each-post-to-its-delegates(no-self-choicetakes any of leg h's own-job asks).gaps.mts,unresolved: how a label resolves to a run.engine.mjs: the cancel shapes, the forwarder rule, the door rule's worker flows.run.mts:a:follow-upanda:follow-up-differs,c:writer-ends,j1:heard-onceandj1:tools.FIX-2087 · epic FIX-2056 · spec
specs/issues/FIX-2087/· draft, does not merge until the open finding is settled and the whole plan passes on a freshmaincommitHow to review this: closure QA, a report plus durable checks
This is the closure PR of an epic. Its body is the closure report, the record of one run of the plan on one
maincommit. Review the checks (do they read what they claim to read, and can they fail?) and the findings (is each one real, and filed against the right owner?). Do not review the product here: every product gap is a child of FIX-2056 that blocks this issue.The report: round 4, every check, its verdict and its evidence
Round 4 merged
maince2282b (merge only), with everyFSDEV_*override unset andvercel:openai/gpt-5.4-miniunless a goal names its own model. It folds in Codex's three reviews and #3790's shorter notes (083cef8, cherry-picked; two trimmed comments kept because they state facts written nowhere else). The plan ran three times. Each run's misses were fixed in the goal before the next, and the last ran whole:Row 1 and 2: the walk, the crash, the failure, the journeys, the sweep (b6516a8)
a:draft-content, thenj2:steer-showsb:hears-in-due-timeandb:once(structural, not retried) → the finding below. 5 diagnostic reruns PASS, 857 to 996 ms past due--controlGOAL_CONTROL=no-sweeper-paymentb:hears-in-due-timeGOAL_CONTROL=no-cascadec:dependent-cancelleda:lands-once, goal out of date (4ec08e6). writer's follow-up turn wrote the original draft word for word. Each line landed once, from its own request, but the two read alike, so the goal counted two lines for each.a:follow-upnow wants a turn of writer's after the follow-up was filed. The new version is judgment (a:follow-up-differs), and lines that read alike are told apart by when each landed.Row 3: every child's check with its controls (b6516a8)
There are 27 children blocking FIX-2087 in Linear, read live, plus 7 with a check here that aren't blocking in Linear.
hand-offs/survives-a-restart-and-runs-once(J4 below).hands-each-post-to-its-delegates's six andsplit-to-roster'shold-session, which row 3 didn't run before.check-talk.mts, the privacy goal's model-free check, PASS.delivery-after-ending,notice-claimandtask-notice, plus workforce'sworker-plans-its-steps, PASS.wait-resumeand orchestration'swait-on-task, PASS.byresolved. FIX-2147 →workforce-privacy/two-users-share-a-project-and-nothing-else. FIX-2149 and FIX-2150 have noby: each is a race inside the park window, milliseconds wide, that no real-model leg lands in. Their regression tests inject the timing, so they are the whole check.general-lead,chief-of-stafftwice andresearchertwice, andeng.triageat 683c151.hand-offs/survives-a-restart-and-runs-once: P1.a to P1.d PASS, with every control red exactly its expected set. J4 failedj:names-ticket→ KNOWN-OPEN FIX-2101, not re-filed.Five controls fixed at 683c151 (all goal-side). Requiring every documented tag, as Codex asked, read five controls as broken:
every-coordinator-is-a-worker-filemailbox-left: its failures are bracketed ([a:no-mailbox-file]). The closure's reader now takes that form.files-tasks-down-the-owners-chainno-follow-up: stale since FIX-2115 and FIX-2122. Workforce's conversation board maps the notice decision to a line itself for a turn-filed or an unrouted failure, so leg e's failure was still said. The control now patches both maps;e:one-noticered 2 of 2.a-worker-plans-its-own-stepsno-plan:a1:no-child-sessionneeds the planless model to file for itself, which it no longer does (4 of 4). Onlya1:steps-in-task-sessionis required now.hands-each-post-to-its-delegatesno-self-choice: best fit now places the project ask on the EM, not the hire (4 of 4). Any of leg h's own-job asks counts.ask-survives-a-restartno-waker: ungraded 4 of 4 standalone runs at 683c151. The keeper ended before the asker'sgetTaskparked, so the wait read an ended row and nothing needed resuming. It went red insidesurvives-a-restart's P1.a, and in row 3 at b6516a8. It is timing, not stale. It could use a held keeper; that's for the hand-offs goal's owner.Row 4: the gap sweep
31 of 31 rows mapped. 30 of 31 have every check run and passing on b6516a8. Every debt a row owes carries a due time fails, because it names
closure b:hears-in-due-time. This is honest under the new grading: a label is green only if its run passed.Row 4 now checks that each row's checks ran and passed. It doesn't check that they reach both sides of every seam, follow the published task-board page, or check the epic's
Not done ifstates. Those stay out of scope here.Finding, for you to file
After a crash, a worker can miss its task's ending: the notice waits on a request the crash left looking alive. Product bug. It hit 1 of 6 leg-b runs on b6516a8.
req_1792800872068_…still readsin_progress, and itsactive_requestsrow is still there;{ owed: false };noticeDeferred:1:completed {"whileRunning":"req_1792800872068_…"};heldWhileRunning, read through the engine'swaits.running), or the engine's recovery of requests orphaned by a crash. The store is kept at/tmp/fsd-goal-one-task-1792800645733_16612/b-1-1792800645733_16612.sqliteon this container.FIX-2101 cause 2: KNOWN-OPEN (Backlog), J4
j:names-ticket.Pre-push checks
On eb33735 (b6516a8 plus the verdict log):
pnpm typecheckPASS ·pnpm lintPASS ·pnpm fallow:ciPASS · sweep PASS and--controlOK ·node scripts/check-retired-terms.mjsPASS · FIX-1792's census--afterPASS.Round 3, in brief
mainaf27e48: everything the closure walks passed except FIX-2005's privacy check on 1 of 3 runs. A message to a worker came back "completed" with no answer, and the answer landed later. That became FIX-2147 (#3764). It was not a product bug: the goal had forkedeng.triage, a coordinator withrouting: everyone, whose delegate answers in a request of its own after the post's. The goal now grades a routed post on its delegate's answer (people.mtstalk), and m2 allows the coordinator's own passes. It's graded here by the privacy goal andcheck-talk.mts.Round 2, in brief
maine07f360: every round-1 finding green. One new finding, a failure's notice missing a dependent the cascade cancelled late, so the filer never finished. That became FIX-2136, merged and graded above. FIX-2097 to FIX-2122 are as recorded ingoal.md's verdict log.Engineering calls: decided along the way, no product sign-off needed
owed: true), not a request started. This follows FIX-2123's claim. If wrong: a duplicate request that did run a turn would go unseen. It can't: the entry returns{ owed: false }before any turn.by), not run twice. A stale reference fails.GUARDin FIX-1792's census, following FIX-2011.GOAL_TOOLS_LOG) and never changes them.https://claude.ai/code/session_01VaN4uTiYqiabCbsFLWM52A
Generated by Claude Code