This is a point-in-time record of the week of 2026-08-31 to 2026-09-06. It was written from the repository's history for that week and is not revised afterwards, so later pages may supersede what it says.

Week of 2026-08-31 to 2026-09-06

Window: 2026-08-31 00:00:00 to 2026-09-06 23:59:59, local author time Recapped: 2026-09-07 Commits in window: 392 (290 non-merge, 102 merges)

The largest week in this repository's history by some distance, and worth saying so before anything else: the prescribed read for a recap was measured at ~19k tokens on a normal week and came to ~190k here. Six plans archived, two of which were also written in the same week; twelve new plan documents were opened. Nearly half of everything — 129 of 290 non-merge commits — belongs to a single plan.

What shipped

Plan 162 — the testing census and CI restructure (129 commits)

Plan 162 exists to run a census of the whole automated test suite for coverage, dead assertions and drift between what CI checks and what the code does, then to close the gaps it finds and restructure CI around what the census showed actually mattered. It inherited its inbox from Plan 161, which had deliberately fixed nothing it discovered.

The week ran the plan from its census to its sixteenth closed stage. The census itself was scoped and numbered (dfa55ae, a0a49c8), then the gaps came off in order: the orphaned test suites wired into CI (5862935), coverage unblinded so its number means something (8c10d95), the two health-sensor censuses given one declared source (17d4fab, a5fda6c, 92ef62b), and the 267-second dbt build + test job decomposed and split into four (5097569, 5dd6bb7, 70d2411, 9f21f87).

Three themes are worth separating out.

Mechanising rules that were previously habits. A sweep moved 34 files to mocker and 16 layer numbers to the contract (5de9a3c); an encoding guard was built at runtime, then rebuilt statically once the runtime version proved to be the wrong shape (a3eea2b, f9a702b, c7d1d33, 8325046); every skip in CI must now be declared or the run fails, demonstrated by adding an undeclared skip to watch the gate fail and then removing it (03553fb, fcec5ac, 7b8ea40).

SQL stopped being paraphrased. G5 — no production module holds SQL at a call site — and G14 — every production .sql file is executed — were closed (9b2f685, 54b516d, dad8f33, 3bcac3a), then Stage X went further: a test may not author SQL either (b45702d), a test may not invent the shape of a relation production defines (815fcf2), and a recorder now captures what text ran against which engine (ae97f83). The resulting execution-coverage gate reports 161 of 161 production .sql files executing, from 9,775 executions across 12 per-job records (2fc6184). The commit is candid that every reading below 161 turned out to be the instrument rather than the repository — CI discarding most of the record, two loaders returning a plain str, fourteen selectors credited only to their wrapper — each found by the gate and fixed on the same branch. The waiver ledger it replaced was then deleted (d7981b3), and 14 SQL waivers were retired, 37 down to 23 (e54c98d).

dbt now builds against production-shaped data (35ed492, da1f3f0), with the gate proved by injecting a deliberate data-test violation to watch it fail and then reverting it (37406c1, e6e3c13).

Two stages were added by production rather than by the suite. Stage K (95cc107, af8f28d) fixed the defect where dashboard and pgadmin — the two services designed to pause no surface — were the two that could never be deployed alone, because V043's CHECK required a non-empty scope on every active record. The second half of that commit is the more useful finding: _set_intent caught bare Exception and rendered every failure as 503 "Database unavailable", so the production diagnosis pointed at the one component that was healthy throughout.

Mid-week the plan was restructured onto Plan 172's new plan-document contract and its stages renumbered from digits to letters (b76904a, 1364808, f0edeca), with a follow-up commit for the bare back-references the sweep missed (9075bbf). Sixteen of twenty-two stages are done; the slice pointer ends the week on Stage S (6d3e3d4).

Plan 138 — public surface refresh (67 commits, archived 2026-09-04)

Plan 138 existed because the live page and README claimed a system the repository no longer was: 36 migrations against 49, 971 tests against 3,661, eleven containers against more than two dozen, and three different make/model counts on one page.

Both surfaces were rewritten from a reconciled contract (1c47893, 515bb82), the landing page moved to / with /info a permanent redirect (c80f494), the page taken same-origin under a strict CSP after a 41.7 MB eager hero video was removed (6695b60), and the hand-maintained roadmap and recap copy replaced with build-time projections CI validates (d411445, 15d028d, e211ecc). Stage 7 made published content servable from the checkout rather than the image (5f00c60), which is why a merged recap now goes live on a git pull with no image build.

Stage 2 shipped, was deployed, broke /dashboard, and was reverted the same night (2b2808b, then 433cc4d to reapply). The post-mortem (a1f1e4b) is the part worth keeping: the plan's analysis of the Streamlit coupling was simply wrong. It had predicted one failure mode and written an assertion that catches it — verified by mutating the Caddyfile three ways — and that assertion was blind to what actually happened. Streamlit serves one SPA shell for every path and routes client-side, so /dashboard was never a route it recognised; it worked because the fallback landed back on Streamlit through the catch-all, and taking / away removed the fallback. The fix confined Streamlit to a base path (22032bc), the mechanism was measured rather than left as a guess (37f4ba3), and the assertion was rewritten against the base path (f3cc48b).

The plan added three mechanisms to keep the surfaces honest — a commit gate that holds a staged surface until it has been read (3d8ea58), a closeout question that fires when the tree moves under a surface nobody edited (a4166ee), and docs/PUBLIC_SURFACE.md as the one contract its three forms cite (c66ea98). Both gates earned their keep late: Stage 4's found the freshness label wrong in two directions on the same field (db02910), and Stage 6's route matrix, run in full for the first time, found HEAD answering 405 on every public route (b5b4eb2). Archived 13ce139, with presentation work carved out to Plan 174 (9f2984f).

Plan 172 — the plan document contract (15 commits, opened and archived 2026-09-03)

Plan 172 was raised out of Plan 138 Stage 1d and archived two days later. Its case: 98 plan documents had no shape rule — 4% of work-unit sections declared an exit, evidence appeared at four heading depths, and 43 of 44 live plans carried no public section.

It produced docs/PLAN_DOCUMENT.md (8e9bdc0) and five skills at three grains — plan-draft (27c783e), plan-start (5de445e), and note-evidence plus the stage-close/close-out split (0d26683) — along with the plans skill's open-plan operation (380bcf8) and a contract test (5a40ff6). Stage G's planned re-measure was retired in favour of a ratchet asserting the waiver list only shrinks (00d4b74), after three of its four before-numbers proved unreplayable; the ratchet was paid down to 38 on archival (46782cd).

Plan 161 — the testing contract (12 commits, archived 2026-08-31)

Plan 161 replaced a Testing Strategy paragraph that had been accurate in April and quietly false by August with a contract in three forms — docs/TESTING.md, a testing-contract skill, and 17 assertions in CI (cad95c8, 06f90e3, f2f5ab3, 2836a53). Mechanising three measurements the prose half had made by eye found all three low, and the asserting test then failed its own first CI run and proved its own thesis: FastAPI changed include_router between 0.128 and 0.141, so a shallow app.routes walk returned 4 routes for ops instead of 54 without erroring, and only the stale-waiver assertion caught it (2324b0a, e46ce27, 313c914). The unpinned pre-1.0 dependency underneath was recorded against Plan 121 (8c854e0).

Plan 164 — the cycle-close ritual (10 commits, moved to closeout)

Plan 164 gives a home to three steps of the cycle-close routine that had no owner and are only ever correct at the close (266b4d5). It built the ref-hygiene step (df8ea36), made it prevention-first and cross-machine (96b3f8e), and enforced the settings and the recap habit rather than the sweep (0643cbd, c6bcadc). The one-time backfill took 69 remote refs down to 4, deliberately leaving one unmerged branch and two that active worktrees were sitting on; every deleted tip was a strict ancestor of origin/master, so no commit went anywhere, only names did. The merged-branch budget dropped 67 → 5 (7e49b71).

Plan 175 — the commit gate's bundling hole (7 commits, opened and archived 2026-09-04)

Plan 175 was found on 2026-09-04 while reading the gate Plan 138 had just built: the PreToolUse hook ran before the command, so it read an empty index whenever a commit staged its own changes — a bundled git add … && git commit fired the gate, found nothing, and returned 0, indistinguishable from a surface that had been read (1cd7755). Enforcement moved inside git commit (9874473).

The stage reintroduced its own defect twice, which is the finding worth keeping. chmod +x on Windows never reaches the index, so the hook committed as 100644 and git silently declines to run a hook without the execute bit — it would have checked out on macOS installed, configured and never firing (5e3f09d). Then hooks_installed() compared core.hooksPath against the literal string .githooks, and within hours the value in this clone had become an absolute path to the same directory, so git was running the hook while the gate called it uninstalled and refused every agent commit (d9eaf8d).

Plan 146 — the planning system (7 commits)

Last week's recap landed (54fee43), and writing it exposed that the state parser could not see the closeout table (bf989fc) — a defect recorded against the plan alongside a correction to the recap's rename sha (83b9245). CI's Documentation tests job was given PYTHONPATH and workflow_dispatch (404f521, 78e5aad, CAR-42). Handoffs and reports were filed out of docs/plans/ (538c9fa) with every broken reference repointed (4bd29dc).

Plan 142 — scoped operational coordination (6 commits, moved to closeout)

The first real maintenance window produced findings, and this week fixed them: four defects that made the window unrunnable (42a81d6), an observability_fresh check that could never pass on any host (9594d19), and a mounts_expected gate that had to read the findmnt tree (1815b39), with the abort criterion coupled to the boundary table (f5e577a) and the first window's findings closed (0b6191a). Deployed and verified in production (2ff4918).

Plan 139 — test suite maintenance (5 commits, archived 2026-08-31)

Stage F gave tests/integration/sql the real Airflow schema so ops queries crossing into the airflow schema execute at all (f3b4406, 9ea2e27, 9e56012). Stage H was folded into Plan 162 and the plan archived (1107433), with the references its archival left stale cleared (34bc12a).

Plan 149 — Linear as the weekly execution layer (6 commits)

Cycle 1's close was recorded (7fb6a92), and four defects in the seeding skills were fixed: seed to a points budget rather than an issue count (bafcebf), fill-cycle creates the projects it seeds into (a30cdb0), a plan that gets an issue gets a project in both skills (7c2bada), and a Backlog-status issue cannot hold a cycleId (7412a25). ticket-now now always passes assignee explicitly (03abeb1).

Plan 170 and the twelve new plan documents (3 commits, plus singletons)

Plan 170 opened with a container image reclaim policy (a831eb2), resolved its Stage 0 trend question and pulled build cache into scope (aa1c82e), then corrected its own measurement claim — containerd is already on a panel (c4cbcc6).

Eleven other plans were opened this week and filed: 165 subdomain routing, scoped out of Plan 138 (40dda68); 166 container PID 1 does not reap its adopted orphans (640877b); 167 the solver truth in .env.example (67c982f); 168 the generated knowledge substrate at build-order 8 (d8b9913); 169 the public analytics agent (624df20); 171 the coordination failure vocabulary (6d563e9), filed directly out of Stage K's two commits; 173 machine credential lifecycle (f5adf55); and 174 the public site shape (9f2984f). Plans 172 and 175 are covered above; 164 opened and reached closeout in the same week.

Also this week: Plan 129's closeout documentation (11aa6d9), Plan 145's post-mortem closed against the finished plan (972dfa5) with its design attribution corrected against git (1fbf526), Plan 125's four new obligations from the testing contract recorded (5c1944c) and Gate C's evidence archived out of the gitignored cache (e0b33d6), Plan 154's expected-running service count corrected twice (81a5c3d, c60ff6d), and Plan 121 given a record of what local bring-up actually is after scripts/setup.ps1 was deleted (f55b1f5).

Every commit, by plan

The narrative above cites the commits that carry the argument. This is the full roll, so the denominator reconciles. Headings are plan numbers; the last two groups are commits that declare no plan and commits that name none.

162 (129)

• dfa55ae docs(plan-162): scope the plan now that the census is done (CAR-40)
• a0a49c8 docs(plan-162): number the census as Stage 0 and record its evidence (CAR-40)
• 5862935 test(plan-162): run the orphaned suites and wire the survivors into CI (CAR-45)
• 296016c docs(plan-162): re-own G13 and fold it into Stage 5 (CAR-45)
• 8c10d95 test(plan-162): unblind coverage and make its number mean something (CAR-46)
• 92cb718 docs(plan-162): scope Stage 5b, separating production scripts from spent ones (CAR-55)
• 2b12294 docs(plan-162): reconcile Stage 2 with Stage 1's gap-list convention (CAR-46)
• c711dcc docs(plan-162): record Stage 2's evidence and unstick the index row (CAR-46)
• 4b6a9c8 docs(plan-162): record what CI verified about the coverage gate (CAR-46)
• 17d4fab test(plan-162): give the two health-sensor censuses one declared source
• a5fda6c test(plan-162): read the census by path, not by importing it
• 92ef62b test(plan-162): make DAG_SPECS name every DAG the DagBag builds
• 06547b3 docs(plan-162): record Stage 3's evidence and point the index at Stage 4
• d9597ef docs(plan-162): mark Stage 2 complete in the stage table
• 5dd6bb7 ci(plan-162): split the dbt build + test job into four jobs (CAR-48)
• 8c54915 ci(plan-162): stop two comments impersonating the Airflow pin (CAR-48)
• 5097569 measure(plan-162): decompose the per-invocation dbt cost (CAR-48)
• 70d2411 ci(plan-162): install per job only what its own suites import (CAR-48)
• e3b4c82 measure(plan-162): probe DuckDB coexistence with dbt in-process (CAR-48)
• 9f21f87 perf(plan-162): run the real-build dbt tests in-process (CAR-48)
• 848b22e docs(plan-162): record Stage 4's evidence, measured (CAR-48)
• 4647e6d test(plan-162): put the read-only guard back, on statements (CAR-48)
• 802adac docs(plan-162): advance Plan 162's build-order slice to Stage 5 (CAR-48)
• 783ab79 ci(plan-162): delete the Stage C measurement rig (CAR-48)
• 4c5c7dd docs(plan-162): bring the coverage section up to date with Stages 2 and 4 (CAR-48)
• 103d987 docs(plan-162): record the third reading, and widen every range to match (CAR-48)
• 5de9a3c test(plan-162): sweep 34 files to mocker and 16 layer numbers to the contract (CAR-49)
• eaec028 test(plan-162): refile test_scrape_listings as the unit test it is (CAR-49)
• e954306 refactor(plan-162): split scripts/ into production and spent (CAR-55)
• 10140ef docs(plan-162): Stage 5b evidence, and scope the incremental-CI question to Stage 10 (CAR-55)
• 0b9d0c4 docs(plan-162): add Stage 6b, the guard G13's exception said would be needed
• 1172a7a feat(plan-162): empty ROUTE_WAIVERS and give container_health a test home (CAR-50)
• d8d607d fix(plan-162): install container_health's deps in the contract job (CAR-50)
• 9741370 test(plan-162): unit-test the Docker contract verifier to 99% (CAR-50)
• 184ca11 test(plan-162): cover the Promtail checker's replay setup too (CAR-50)
• c5c7650 docs(plan-162): record Stage 6's CI evidence (CAR-50)
• 9dd99f5 docs(plan-162): advance the slice pointer to Stage 6b (CAR-50)
• 97f4af8 docs(plan-162): keep the status lines general so they stop going stale
• 8325046 fix(plan-162): name the encoding on every production text read and write
• a3eea2b test(plan-162): fail when a text read or write lets the locale decide
• f9a702b test(plan-162): catch the encoding shapes no static rule can name
• 27288e6 fix(plan-162): scope the encoding guard's ignores to third-party modules
• c7d1d33 refactor(plan-162): check the encoding shapes statically instead of at runtime
• cc264db docs(plan-162): record what CI said about the encoding guard
• b15c506 docs(plan-162): close out Stage 6b and advance the slice pointer to Stage 7
• 0d7b436 fix(plan-162): skip gitignored files in the encoding guard's file walk
• 9b2f685 feat(plan-162): mechanise G5, the inline-SQL gap nothing has ever checked
• 0e4be9a test(plan-162): execute the six Data Health statements against real DuckDB
• cad86d5 refactor(plan-162): extract the seven single-site statements, and execute them
• 54b516d feat(plan-162): close G5 -- no production module holds SQL at a call site
• 566032c test(plan-162): close the two holes that closing G5 exposed
• 2af4d5f docs(plan-162): record Stage 11 and G16, the question Plan 161 did not ask
• 404ef08 fix(plan-162): scan production Python, not service packages, for SQL
• 4b05d9d refactor(plan-162): sort the scripts' SQL into what is production and what is not
• 13d8de6 fix(plan-162): let the Layer 2 census read where the selector tests already are
• dad8f33 feat(plan-162): drain G14 to one entry, and stop a statement being filed twice
• 23e2913 docs(plan-162): add Stage 12, the fixture duplication this plan created
• 3bcac3a feat(plan-162): close G14 -- every production .sql file is executed
• da72354 test(plan-162): execute ops' statements instead of retyping them
• 53f7bf0 refactor(plan-162): file admin.py's statements, and unblock the last paraphrases
• 2a98d4b docs(plan-162): measure the assertion duplication Stage 12 inherits
• 35ed492 fix(plan-162): a placeholder in a .sql comment is part of the statement
• ffa8225 docs(plan-162): record Stage 7's evidence, and close it
• 88b185d docs(plan-162): advance the slice pointer to Stage 8 (CAR-51)
• dfa0fac docs(plan-162): add Stage 6c, the contract-versus-constraint gap a deploy found
• fe0ace8 docs(plan-162): point the slice cell at Stage 6c, not Stage 8
• 086724f docs(plan-162): record Stage 7's production deploy
• 95cc107 fix(coordination): let a service that pauses no surface be deployed alone
• af8f28d fix(coordination): report why a deploy release failed
• 6b66333 docs(plan-162): narrow Stage 8, hand the dashboard Python to Plan 150
• f6a7077 docs(plan-162): close out Stage 6c with its production evidence
• ceb4f07 test(plan-162): assert Layer 2's second clause, and run the scraper's write path unmocked
• 5ffd3fb test(plan-162): close G8 by running the SRP fetch path unmocked
• 5b8b144 docs(plan-162): document SCRAPER_RESULTS_BASE_URL and why it stays unset
• cdd2906 docs(plan-162): record Stage 8's evidence from CI
• 575ae72 docs(plan-162): record Stage 8's deploy, and the half it does not confirm
• b86fd51 docs(plan-162): record Stage 8's cost as 2 estimated, 1 actual
• 9044b8e feat(plan-162): give airflow/dags a reachable .sql convention, and replace two SQL rules with one
• 7011aa1 docs(plan-162): close out Stage 8 with its production confirmation
• 3398299 docs(plan-162): point Plan 162's slice cell at Stage 9
• afb3100 docs(plan-162): close out Stage 9, and point the slice cell at Stage 10
• 99f7f73 docs(plan-162): scope Stage 10 and split it into 10, 10b and 10c
• 35532b4 feat(plan-162): build dbt against a production-shaped snapshot (CAR-54)
• e54c98d refactor(plan-162): retire 14 SQL waivers, 37 -> 23 (CAR-54)
• 2e346cd ci(plan-162): hold the snapshot-dbt gate for its own PR (CAR-54)
• af8b2ec docs(plan-162): record Stage 10's split, the DAG defect, and the credential trail
• da1f3f0 feat(plan-162): gate dbt on a production snapshot, and fix the DAG that exports it (CAR-54)
• 37406c1 test(plan-162): TEMPORARY — inject a data-test violation to prove the gate fails
• e6e3c13 docs(plan-162): record Stage 10's deliberate violation, and revert it (CAR-54)
• a70fb12 docs(plan-162): close Stage 10, and open Stage 13 for the skips it exposed (CAR-54)
• fa2f4c7 docs(plan-162): give Stages 11 and 12 the exits they never had
• 71f6b1d docs(plan-162): open Stage 15 for the defect class the DAG fix only patched
• d6a8cb3 docs(plan-162): open Stage 16 for SQL a test authors itself
• 44946a3 docs(plan-162): record how Stage 16 was arrived at
• ef16b1c docs(plan-162): size Stage 16 and repoint the slice cell to 10b
• 7ccb735 docs(plan-162): rescope Stage 10c and record Stage 10b's decisions
• b76904a docs(plan-162): restructure onto the plan-document contract
• 1364808 docs(plan-162): repoint every external stage reference at the new letters
• 9075bbf docs(plan-162): convert the bare stage back-references the sweep missed
• f0edeca docs(plan-162): reorder the remaining eight stages, and repoint the index
• a0c47a1 docs(plan-162): estimate Stages R and V at 1 point each
• 30e2b94 docs(plan-162): rewrite Stage S around branch coverage, move the recorder to X
• 70127be docs(plan-162): point Stage T's dependency at X, where the recorder now lives
• 03553fb feat(plan-162): declare every skip in CI, or fail the run
• fcec5ac test(plan-162): an undeclared skip, to watch the gate fail
• 7b8ea40 test(plan-162): remove the demonstration skip, green again
• f111d02 docs(plan-162): close Stage U
• 9253ae7 refactor(plan-162): move every executed test statement into tests/sql/
• b45702d feat(plan-162): a test may not author SQL either
• 815fcf2 feat(plan-162): a test may not invent the shape of a relation production defines
• ae97f83 feat(plan-162): record what text ran against which engine
• ec9ad67 fix(plan-162): keep the execution record CI was throwing away
• e044708 feat(plan-162): the SQL execution coverage gate, reporting once before it bites
• 52ee24e fix(plan-162): load the last two statements that were read rather than loaded
• fdba0de fix(plan-162): SqlText carries the set of files a statement came from
• 2fc6184 feat(plan-162): the SQL execution coverage gate bites -- 161 of 161
• 9cd23ba test(plan-162): the execution gate's own CI wiring is asserted, not assumed
• d7981b3 refactor(plan-162): delete the execution-coverage waiver ledger
• 1cc1c67 refactor(plan-162): the recorder keys on the argument, not the method name
• a616d0c feat(plan-162): declare the routes SQL travels, and assert them both ways
• d2d0561 test(plan-162): guard the denominator every coverage number is a fraction of
• 79d606b docs(plan-162): record the contract-drift findings this stage did not fix
• a3d2b17 test(plan-162): plan every templated test statement instead of waiving it
• 301beb1 feat(plan-162): Plan 129's sampling statements move to files and gain a Layer 2 test
• 6c6dcc0 docs(plan-162): a skill for adding SQL, because the reader starts cold
• ae74cf3 docs(plan-162): the prior-art comparison conceded more than it measured
• 2906edf fix(plan-162): ask for bindings only where there is a placeholder to bind
• 6d3e3d4 docs(plan-162): close Stage X, and point the slice at Stage S
• de188c8 docs(plan-162): the depends-on cell states a dependency, not progress

138 (67)

• e0b2054 docs(plan-138): publish the recaps, and name the coupling the root move breaks
• 385c2be docs(plan-138): fix the generator contract Stage 1d and 1e could not be built from
• b869275 docs(plan-138): reconcile the overview's solver naming with the live path (CAR-37)
• 3cb759b docs(plan-138): record the Stage 0 baseline and dispose of every drift row (CAR-44)
• f0a5f81 docs(plan-138): correct the baseline's stated measurement basis (CAR-44)
• fa9ef9b docs(plan-138): close Gate 0 and open Stage 1a (CAR-38)
• 1c47893 docs(plan-138): rebuild the README against the public truth contract (CAR-38)
• 0f3f622 docs(plan-138): retitle the cover badge its content no longer matches (CAR-38)
• 3a48feb docs(plan-138): mark both overviews by register, not by disclosure
• 3bc23f1 docs(plan-138): bring the storage overview up to date with Plan 145
• b33c4f9 docs(plan-138): drop the README's pointer to a setup script that cannot run (CAR-38)
• 515bb82 feat(plan-138): rebuild the landing page against the public truth contract
• adbc8d0 docs(plan-138): record Stage 1b, open Stage 1f, and bring Stages 3 and 5 current
• 7a47f18 docs(plan-138): point the build-order row at Stage 1c (CAR-56)
• e9302d9 docs(plan-138): bring the Stage 1 slice table current for 1b and 1c (CAR-56)
• 3d8ea58 feat(plan-138): gate public-surface commits on a review skill (CAR-56)
• 436842c docs(plan-138): record the Stage 1c evidence and point the row at 1d (CAR-56)
• d411445 feat(plan-138): generate the public roadmap projection (CAR-57)
• b0cec16 docs(plan-138): record Stage 1d and promote 1e to the next slice (CAR-57)
• 4a9de43 fix(plan-138): regenerate the stale public roadmap projection
• 469d2ed fix(plan-138): write test fixtures as UTF-8, not the Windows locale codepage
• c204d22 fix(plan-138): make the public-surface hook run on Windows
• 15d028d feat(plan-138): project the weekly recaps onto the public site (CAR-58)
• 64ac21c docs(plan-138): record Stage 1e's publication policy and what it corrected (CAR-58)
• 9d46e6b fix(plan-138): make plan-week regenerate the recap projection it invalidates (CAR-58)
• ee80754 docs(plan-138): record Stage 1e and promote 1f to the next slice (CAR-58)
• e5452ca fix(plan-138): regenerate the roadmap projection the slice-cell edit staled (CAR-58)
• 85f29d8 docs(plan-138): teach the plans skill that its two files are published (CAR-58)
• f632613 docs(plan-138): draw the corpus framings onto both surfaces (CAR-59)
• 5367ab0 docs(plan-138): advance the slice pointer to Stage 1h (CAR-59)
• 4960b52 docs(plan-138): replace the PR sequence with a build order (CAR-61)
• 59b5b15 docs(plan-138): draft article D on the April cutover
• e6b6a48 docs(plan-138): rewrite article D's companion post as a teaser
• a4166ee docs(plan-138): add the closeout public-surface question (CAR-61)
• 331b980 docs(plan-138): record Stage 1h evidence and advance the slice pointer (CAR-61)
• 3907c2f docs(plan-138): close the seam that let a slice-pointer edit publish unnoticed (CAR-61)
• 81ddeff docs(plan-138): point the slice cell at Stage 2 now the ops deploy has landed (CAR-61)
• a20b94f docs(plan-138): add Stage 7 -- publish content without an image rebuild
• 5f00c60 feat(plan-138): serve generated public content from the checkout, not the image
• 3675e47 docs(plan-138): point the remaining stage specs at the moved content path
• eb45064 docs(plan-138): correct the deploy state and the unreachability claim
• 3c53a35 docs(plan-138): close out Stage 7, and correct a claim it was publishing
• c80f494 feat(plan-138): make / the public landing page and give the recaps a route
• afac29c docs(plan-138): add the Stage 2 deploy run sheet
• 2b2808b Revert "Merge pull request #338 from whitewalls86/car-62-plan-138-stage-2"
• a1f1e4b docs(plan-138): record why Stage 2 was deployed and reverted
• 22032bc fix(plan-138): confine Streamlit to /dashboard/ with a base path
• 37f4ba3 docs(plan-138): replace "mechanism unknown" with the measured mechanism
• 433cc4d Reapply "Merge pull request #338 from whitewalls86/car-62-plan-138-stage-2"
• f3cc48b test(plan-138): assert Streamlit under its base path, not under the catch-all
• 6060495 docs(plan-138): close out Stage 2 with Gate 2 met, and advance the slice pointer
• 74d17aa docs(plan-138): add Stage 8, extract the public surface contract
• c66ea98 docs(plan-138): extract docs/PUBLIC_SURFACE.md, the public surface contract
• 1cd9afa docs(plan-138): re-measure the status block against the tree
• 4842209 docs(plan-138): point the slice cell at Stage 8, not Stage 4
• 847d76e docs(plan-138): record Stage 3's three decisions, which were made and unwritten
• 6695b60 feat(plan-138): take the landing page same-origin and drop the hero video
• 50ebf55 docs(plan-138): record Stage 3b and 3c evidence, and correct a stale routing comment
• 38eaa66 docs(plan-138): bring the Caddy route test's header up to date
• 8761f7c docs(plan-138): redesign 1g as its own page, and sequence Stage 9 first
• e211ecc feat(plan-138): publish what a plan is for, not which stage is next
• 0ec6d5a feat(plan-138): close Stage 4 and add Stage 10, covering the surfaces Stage 9 created
• db02910 fix(plan-138): show the analytics bucket's end, not its label
• b8dcc79 test(plan-138): close Stage 5 with a README contract and a fallback assertion
• b5b4eb2 fix(plan-138): answer HEAD on every public route, and revalidate the generated artifacts
• 780d393 docs(plan-138): record Gate 1h as met, which it has been since Stage 2
• 13ce139 docs(plan-138): archive the public surface refresh

172 (15)

• 3f8ca3b docs(plan-172): open the plan-authoring plan, and Plan 138's stage that consumes it
• 5d45022 docs(plan-172): rescope to the plan document, and give it four questions
• 9054f0c docs(plan-172): bring the index row up to the rescoped plan
• 8e9bdc0 docs(plan-172): write the plan-document contract
• 380bcf8 feat(plan-172): add the plans skill's open-plan operation
• 27c783e feat(plan-172): add the plan-draft skill
• 9bd25d4 docs(plan-172): ask how stages chunk into issues, not one-per-stage
• 5de445e feat(plan-172): add the plan-start skill
• 5a40ff6 test(plan-172): mechanise the plan-document contract's public sections
• db0142e fix(plan-172): address PR #352 review findings
• 38da869 docs(plan-172): point Stage G at the measurement it actually re-runs
• 0d26683 feat(plan-172): split stage-close out of close-out, add note-evidence
• 00d4b74 test(plan-172): assert the waiver list only shrinks, retiring Stage G's re-measure
• 219d9ec docs(plan-172): repoint the five references the close-out split orphaned
• 46782cd docs(plan-172): archive the plan, and pay the waiver ratchet down to 38

161 (12)

• cad95c8 docs(plan-161): the testing contract, and what it is measured against
• 06f90e3 docs(plan-161): one mocking convention, layers ordered by cost, ast as a symptom
• 9a56635 docs(plan-161): the restructuring has a plan number, and 139's split is decided
• ab6c5b3 docs(plan-161): CAR-33 closeout evidence
• db89cbb docs(plan-161): the harness must not decide the outcome
• f2f5ab3 test(plan-161): the reviewer skill and the test that asserts the contract (CAR-34)
• 313c914 test(plan-161): fail on an empty routing table, and report what was enumerated (CAR-34)
• e46ce27 test(plan-161): enumerate routes from the OpenAPI schema, not a routes walk (CAR-34)
• 2324b0a docs(plan-161): record the FastAPI enumeration failure as CAR-34 evidence
• 5147200 test(plan-161): assert gap-list owners against the archive, and reconcile Plan 139
• 92d2f99 docs(plans): archive Plan 161 and move Plan 142 to closeout
• 2836a53 test(plan-161): assert the contract's claims about its own enforcement (CAR-43)

164 (10)

• 266b4d5 docs(plan-164): a cycle-close ritual, for the three steps nothing owns
• 8bc1c35 docs(plan-164): git branch -d checks the upstream, not HEAD
• df8ea36 feat(plan-164): write the cycle-close order and build the ref-hygiene step
• 96b3f8e feat(plan-164): make ref hygiene prevention-first and cross-machine
• 0643cbd test(plan-164): enforce the settings and the recap habit, not the sweep
• c6bcadc test(plan-164): tighten the recap deadline to Wednesday, and pin the boundary
• 7e49b71 feat(plan-164): drop the ref ratchet to 5 after the backfill swept 65 branches
• 95ed7f9 feat(plan-164): build Stages 1 and 2, and correct the premise that blocked them
• 181b8e1 docs(plan-164): close Stages 0, 1, 2 and 3 with their record entries
• fee0089 docs(plan-164): move to closeout, and backfill Plan 154's public section

146 (7)

• 54fee43 docs(recaps): recap the week of 2026-08-24 to 2026-08-30
• bf989fc fix(plan-146): the state parser could not see the closeout table
• 83b9245 docs(plan-146): correct the recap's rename sha, and record the parser defect
• 404f521 fix(plan-146): give the Documentation tests job PYTHONPATH (CAR-42)
• 78e5aad fix(plan-146): add workflow_dispatch so CI can be re-run manually (CAR-42)
• 538c9fa docs(plan-146): file handoffs and reports out of docs/plans/
• 4bd29dc docs(plan-146): repoint every reference the move broke

175 (7)

• 1cd7755 docs(plan-175): open the commit gate bundling hole plan
• 9874473 feat(plan-175): move the public-surface gate inside git commit
• 5e3f09d fix(plan-175): mark the pre-commit hook executable in the index
• d9eaf8d fix(plan-175): compare the hooks path as a path, not as a string
• b211212 docs(plan-175): close Stage A with its record entry
• 2f4cb7f docs(plan-175): archive the plan with its public summary
• 098eefe fix(plan-175): consume the attribution stamp on every accepted message

142 (6)

• 42a81d6 fix(plan-142): the four defects that made a maintenance window unrunnable
• 9594d19 fix(plan-142): observability_fresh could never pass, on any host
• 2ff4918 docs(plan-142): Stage 4's defects are deployed and verified in production
• 1815b39 fix(plan-142): read the findmnt tree so mounts_expected can pass (CAR-17)
• 0b6191a fix(plan-142): close the findings the first real window produced (CAR-17)
• f5e577a test(plan-142): couple the abort criterion to the boundary table (CAR-17)

149 (6)

• bafcebf chore(plan-149): seed cycles to a points budget, not an issue count
• 7fb6a92 docs(plan-149): record Cycle 1's close, and make category 3 test the slice
• a30cdb0 chore(plan-149): fill-cycle creates the projects it seeds into
• 7c2bada chore(plan-149): a plan that gets an issue gets a project, in both skills
• 7412a25 chore(plan-149): a Backlog-status issue cannot hold a cycleId
• 03abeb1 fix(ticket-now): always pass assignee explicitly on issue creation

139 (5)

• f3b4406 test(plan-139): give tests/integration/sql the real Airflow schema
• 9ea2e27 docs(plan-139): Stage F's CI evidence, measured rather than estimated
• 9e56012 docs(plan-139): mark Stage F complete
• 1107433 docs(plan-139): fold Stage H into Plan 162 and archive Plan 139
• 34bc12a docs(plan-139): clear the references Plan 139's archival left stale

170 (3)

• a831eb2 docs(plan-170): add container image reclaim policy
• aa1c82e docs(plan-170): resolve Stage 0's trend question, pull build cache into scope
• c4cbcc6 docs(plan-170): correct the measurement claim -- containerd is already on a panel

121 (2)

• 8c854e0 docs(plan-121): record the unpinned-FastAPI incident against the constraints item
• f55b1f5 docs(plan-121): record what local bring-up actually is, after deleting setup.ps1

125 (2)

• 5c1944c docs(plan-125): record the four obligations the testing contract now places on its gates
• e0b33d6 docs(plan-125): archive Gate C's evidence out of the gitignored cache

145 (2)

• 972dfa5 docs(plan-145): close the post-mortem against the finished plan
• 1fbf526 docs(plan-145): correct the design attribution against git

154 (2)

• 81a5c3d docs(plan-154): correct the expected-running service count in the build order
• c60ff6d docs(plan-154): correct the expected-running set and every count derived from it

129 (1)

• 11aa6d9 Plan 129 Closout documentation.

165 (1)

• 40dda68 docs(plan-165): subdomain routing, scoped out of Plan 138

166 (1)

• 640877b docs(plan-166): container PID 1 does not reap its adopted orphans

167 (1)

• 67c982f docs(plan-167): open Plan 167 and state the solver truth in .env.example

168 (1)

• d8b9913 docs(plan-168): schedule the generated knowledge substrate at build-order 8

169 (1)

• 624df20 docs(plan-169): file the public analytics agent in the backlog at 46

171 (1)

• 6d563e9 docs(plan-171): file the coordination failure vocabulary

173 (1)

• f5adf55 docs(plan-173): open machine credential lifecycle in the build order (CAR-77)

174 (1)

• 9f2984f docs(plan-174): open the public site shape plan, carved out of Plan 138

none-declared (3)

• aaeb4b6 chore: delete scripts/setup.ps1, superseded and non-functional since April
• d2cc8ba feat: enforce the commit-attribution skill instead of remembering it
• b0e3d6a docs: drop the CI lake snapshot from the not-production-serving lists

unattributed (3)

• 6b9d4f2 feat(ops): name and scope the snapshot download tokens
• aebcacf refactor(ops): put snapshot credential lookup behind a storage seam
• 8acf088 fix(ops): deliver SNAPSHOT_DOWNLOAD_TOKENS to the container, and open Stage 14

What moved between states

Read from docs/PLANS.md's own history across the window — 57 index revisions, comparing the last revision before 2026-08-31 with the last inside 2026-09-06.

Two archivals are invisible to that diff and must be named separately. Plans 172 and 175 were both opened and archived inside the window, so they read as absent at each endpoint and cancel out. Both appear in completed_plans.md, dated 2026-09-03 and 2026-09-04. This is a limit of an endpoint comparison, not a finding about the week: any plan whose whole life fits inside one window will do the same.

Six plans archived in seven days — 129, 138, 139, 161, 172, 175 — which is the largest number this repository has recorded in one week.

Twelve plan documents were created in the window (164 through 175), and every one of them appears in a table. There is no unindexed plan document this week, which is the defect Plan 146 was written to prevent.

What is still owed

• Plan 162 has six stages left — S, T, W, Q, V and R — of twenty-two. The slice pointer ends the week on Stage S, branch coverage for the dbt models (6d3e3d4). Sixteen closed in one week; the remaining six were reordered on 2026-09-04 (f0edeca, a0c47a1).

• Three .env.example keys have no established answer. 8acf088 measured 37 keys, four of which no docker-compose*.yml references. One was the defect it fixed; FASTAPI_ADMIN_KEY, MLFLOW_TRACKING_URI and PROVENANCE_ENV are pre-existing and each needs wiring, deletion, or a declaration that it is script-only. Stage V is deliberately unsized until that investigation runs.

• Stage X recorded contract-drift findings it did not fix (79d606b).

• Plan 138 archived with two deliverables never produced — the mobile Lighthouse report and the before/after measurements. Four of its completion criteria moved to Plan 174 along with the presentation work. The archive row says so; it is recorded here because an archived plan stops being read.

• Four closeout gates are open and dated. Plan 164 lands on Cycle 2's close (2026-09-09), Plan 142 on a second maintenance window run through the checked-in procedure (2026-09-30), Plan 146 on docs/recaps/ holding a file for every complete week from 2026-08-17 to 2026-09-13, and Plan 149 on 2026-09-15. Plans 136 and 160 also sit in closeout with dates of 2026-09-17 and 2026-09-13.

  This recap is itself one of the four files Plan 146's gate counts.

• Plan 138 Stage 2's revert is a resolved question, not an open one. The mechanism was unknown at the time of a1f1e4b and measured in 37f4ba3; it is listed here only so a reader of the revert does not go looking for a loose end.

Unattributed commits

Three commits name no plan in either their subject or their body. All three are the snapshot-download credential work on the ops service:

• 6b9d4f2 feat(ops): name and scope the snapshot download tokens
• aebcacf refactor(ops): put snapshot credential lookup behind a storage seam
• 8acf088 fix(ops): deliver SNAPSHOT_DOWNLOAD_TOKENS to the container, and open Stage 14

Their bodies mention Plans 108 and 112 as context for why the token format needed scopes, which is not the same as claiming either plan. The enclosing merge commits say where to look — 05de301 (PR #358, car-54-plan-162-stage-10-gate) for the first two and 381d66b (PR #359, car-plan-162-compose-env-wiring) for the third — and 8acf088 discusses Plan 162's Stages 13 and 14 by number. That is a strong hint and not an attribution, so it is recorded as a hint. The work itself is recapped above under Plan 162's Stage P and in What is still owed.

Three further commits state explicitly that they belong to no plan, which is an answer rather than a gap:

• aaeb4b6 chore: delete scripts/setup.ps1, superseded and non-functional since April
• d2cc8ba feat: enforce the commit-attribution skill instead of remembering it
• b0e3d6a docs: drop the CI lake snapshot from the not-production-serving lists

b0e3d6a is worth quoting on the point: it says attributing the fix to either Plan 138 or Plan 174 "would be inventing a mapping". d2cc8ba built the mechanism that makes this section shrink — a PreToolUse hook plus a commit-msg hook enforcing that every commit is routed through the attribution skill, written after five commits went through one invocation of it in a single session.

Merges

102 merges: 78 GitHub pull-request merges (#298 through #375) and 24 local integration merges, none of which is recapped — the work they carry is the non-merge commits above, and recapping both would double the week.

Pull-request merges (78)

Local integration merges (24)

These carry no PR number. Most are origin/master merged down into a working branch mid-stage; three at the start of the window are branch merges from the docs/plans/ relocation.

Two notes on this table. The three merges dated 2026-08-31 that name a branch rather than origin/master — c47f6b4, 15bb60e and 5a7b74a — are the tail of the relocation of plan documents into docs/plans/. The other 21 are origin/master merged down into a working branch mid-stage; they carry work already recapped, from this window or an earlier one, and introduce nothing new.

2b2808b and 433cc4d are absent from both merge tables on purpose. They revert and reapply PR #338's merge, and reverting a merge produces an ordinary single-parent commit, so git counts them among the 290 non-merge commits. They are recapped under Plan 138.

Deferred to the next recap

Zero commits. The window closed at 2026-09-06 23:59:59 and this recap was written on 2026-09-07, so nothing has landed since:

git rev-list --count --all --since='2026-09-06 23:59:59'
# 0

This is the first recap in some time with an empty tail — the measured figure on 2026-08-21 was 117 deferred against 55 in the window. It reflects when the recap was written rather than anything about the week.

docs/recaps/ now holds a file for every complete week from 2026-04-19 to 2026-09-06 with no gap. No week is owed.