Skip to content

fast lane: cut pre-build ceremony (#819, #821) - #828

Draft
MendixMau wants to merge 8 commits into
masterfrom
triage/fast-lane
Draft

MendixMau wants to merge 8 commits into
masterfrom
triage/fast-lane

Conversation

@MendixMau

@MendixMau MendixMau commented Nov 20, 2026 •

Copy link
Copy Markdown
Owner

Direct lane

Fixes #819 (options 1 and 3). Addresses #821 points 1, 2 and 3. Point 4 (Size tier: small changes what is owed) is not in this PR.

What changed and why (one paragraph):
Two simulated builds on 2026-11-20 spent 52% (POC) and 75% (requirements-driven) of their cost before the first model write. Neither run had an exec.sh refusal. The cost came from gate ceremony: agents learned each parsed format one FAIL at a time by reading gate source, filled agent-stub slots for subagents they never dispatched, a POC fabricated an open-questions file to clear a build-ready it should not owe, and kickoff was assembled by hand with Python heredocs. This prototype branch has four commits. Each commit can be cherry-picked on its own: the code applies cleanly alone, and only CHANGELOG.md and two doc lines conflict when the commits are taken out of order.

Commit Effect Fixture: branch / master
fda21137 fix(gate-check, poc) — #819 Under Build type: poc, gate-check.sh <p> build-ready still prints its checklist. It ends BUILD-READY: N/A (Build type: poc) with exit 0, raises no open-questions escalation and writes no .build-ready-pass stamp. The exec.sh plan rule already exempted poc and is unchanged. Stage 5's look check reports MANUAL and still names the unseen pages. poc is now an Entry mode: token that implies Build type poc and removes the "No entry mode declared" banner. Artifact and obligation rows stay owed, exactly as for Build type: poc. Kick-off brief line 3: build-ready applies to full builds only. test-fast-lane-poc.sh 15/15 / 6/15 (25/25 after the review fixes)
82625470 fix(gate-check) — #821 pt 2 Unfilled {{…}} slots in agent stubs become a ⚠ WARN that still names every slot. They FAIL again for mdl-agent once its dispatch log (.claude/loop/dispatch/<Module>/scripts.tsv) exists. Every stub still refuses to run while a slot remains (line 11). This supersedes the 2026-11-13 maintainer decision that any unfilled slot FAILs: the maintainer needs to accept that. test-fast-lane-agent-slots.sh 8/8 / 3/8
1f50ec8a new(gate-check) — #821 pt 1 New bin/lib/format-hints.sh holds one minimal passing example per parsed format: intake markers, build plan, coverage ledger, LOOK report, prototype LOOK, improvement-register row. After the stage list, gate-check prints the shapes that this run's FAILs owe. Obligation FAULTs (look, final-look, prototype-look) print theirs after the obligation block. Each shape prints at most once per run. Informational only. test-fast-lane-format-hints.sh 13/13 / 7/13 (the 6 master failures are the missing shapes). Every example is run through its real parser. Two hand-drafted examples failed that check before it was in place: a coverage wildcard hit SECTION-WIDE, and a prototype denominator was too small
b2b9e923 new(kickoff) — #821 pt 3 bin/kickoff.sh <p> --mode <poc|requirements-driven|migration|change> [--build-type] [--interview] [--answers file] [--by person] writes the register lines, one Stage-P Decisions row each (CONFIRMED only with --by), and one answer marker per intake section, then prints the Stage P verdict. Under steering/assist it lists the unasked questions and never invents answers. It is idempotent. Unknown values exit 2 with the accepted list. Changing an already declared mode exits 3. test-kickoff.sh 26/26 / n/a (no kickoff.sh on master; 31/31 after the review fixes)

Review fixes (independent review of #828: A and D MERGE-AFTER-FIX, B MERGE, C HOLD). These were added as new commits; no history was rewritten.

  • 9a144ba7 (A):
    • When Entry mode: poc and a Build type: line naming full disagree, full wins. gate-check and exec.sh print the conflict, and under auto exec.sh holds the project to the requirements-driven plan rule.
    • Build type: poc → full (CR-01) now reads as full, both in _position.sh and in gate-check.
    • The POC Stage 5 MANUAL line keeps the look: N of M built page(s) never looked at … text that bin/ready.sh parses. ready.sh now names a POC's unseen pages instead of saying "nothing measured".
    • test-fast-lane-poc.sh P5/P6: 21/21.
  • e18f18bc (D):
    • Rows and intake markers are CONFIRMED only with --by <person>. Without it they are ASSUMED, "no person named", and that includes Interview mode: auto. A later --by run upgrades kickoff's own rows, then is idempotent.
    • Under steering/assist, an unstated --build-type / --interview leaves Q16 / Q9 unanswered and listed first, with no register line and no row.
    • The refusal wording no longer calls an ASSUMED default "decided".
    • Kick-off brief line 2 now says: ask first, then record with kickoff.sh … --by <name>.
    • test-kickoff.sh K7/K8: 30/30.
  • Re-review fixes (at e18f18b).
    • A, "fullscreen" no longer flips a POC. A substring test read poc — CONFIRMED (user, fullscreen kiosk) as a full build, silently. One shared test, mxtk_bt_names_full (project-bin/_position.sh), now matches full only as a whole word. Whenever a poc-prefixed Build type line, or Entry mode: poc, is read as full, gate-check and exec.sh both print a ⚠. test-fast-lane-poc.sh P5/P7: 25/25. On the reviewer's probe (check5.sh), X1/X2 now print the ⚠, X1 is REFUSED under auto, and X4 ("fullscreen") reaches check with no ⚠.
    • D, a POC needs no --interview. kickoff.sh --mode poc --by X without --interview now gives Q9 poc-build.md's POC marker, so Stage P passes. test-kickoff.sh K3 variant: 31/31.
  • D ships with A. bin/kickoff.sh --mode poc writes Entry mode: poc, a token only A teaches the tokeniser. D's doc lines also sit next to A's, so cherry-picking D without A conflicts. Merge or drop them together.
  • C is held for a maintainer ruling; its code is unchanged. The comment at bin/gate-check.sh:2498-2503 ("review-agent … a project could pass build-ready with the reviewer an inert stub … The count is six") is now stale under C. Build-ready no longer FAILs on an unfilled review-agent stub, so that gap is reopened and left to the look obligation. If C is accepted, the comment must be rewritten and the ruling recorded in process/process-learnings.md. Also: the mdl-agent re-FAIL fires only after mdl-agent has run (agents/README.md step 8a writes the log), and it cannot revoke a sticky .build-ready-pass.
  • Found by the reviewer, not fixed here (pre-existing): the stray .leakguard-deny comes from test-mode-contract.sh:165,173 and test-ci-gates.sh:163. Both call init-project.sh without MXTK_LEAKGUARD_DENYFILE.

Existing fixtures run against the branch (subject = the changed files: gate-check.sh, obligation-check.sh, _position.sh, status.sh, ready.sh, ci-gates.sh): every fixture listed below, all green.

  • Gate fixtures: auto-mode-hand-gates, bug02/03/04/05/06, change-log, closeout, decisions-blank-line, fitgap-review, gate-cache, gate-false-greens, gate-s4cov-cache, interaction-model-gate, marketplace-install.
  • Obligation fixtures: due-and-noise, fidelity, final-look, journeys-ran, prototype-look, quality-floor, valid-at.
  • Source, stage and verdict fixtures: source-ledger, source-sufficiency-gate, stage-header, stage-p, stage6-report-read, wireframe-count, wrong-verdicts, surface-column, open-questions.
  • Mode and readiness fixtures: mode-contract, entry-mode-plan-rule, greenfield-retired, ci-gates, ready, progress-report-polish, guide-reopen, runbook-slice.
  • Status fixtures: status-bare-var, -existing-app, -extension-modules, -lint-unversioned, -teamserver.

One exception: test-progress-report-field-fixes.sh gave 97/98, then 96/98 with a different failure, then 98/98 on a rerun of the same tree. Master gave 98/98. Its card … | grep -q assertions look like the pipefail/SIGPIPE flake. That fixture does not touch any changed file. The full suites were not run.

Impact

Row Answer
Process Removed for POC: build-ready as a gate, and Stage 5 FAIL on look (now MANUAL, pages still named). Removed for every build: the build-ready FAIL on unused agent-stub slots (now WARN; FAIL again once mdl-agent is dispatched). Added: format examples printed under FAILs, and an optional bin/kickoff.sh. Nothing new blocks. One verdict flip for existing projects after sync: a full-build project whose only open build-ready item was stub slots now PASSes build-ready. Under Interview mode: auto that pass writes the stamp that clears exec.sh's plan rule. Review-agent's stub is no longer forced complete before Stage 5, but the look obligation still owes the LOOK.
Dependencies Shared files touched: bin/gate-check.sh, project-bin/_position.sh (installed beside exec.sh; pure bash and fork-free, so it is still within the exec process budget), bin/lib/artifact-check.sh (one case arm), bin/lib/obligation-check.sh. New: bin/lib/format-hints.sh (sourced by gate-check), bin/kickoff.sh (reads PROJECT.md and intake.md, produced by bin/init-project.sh, the mandatory step 1 of the kick-off brief). New tokeniser output poc: every reader was checked. exec.sh / status.sh only test for migration, requirements-driven and greenfield-retired; artifact-check maps poc to "every row owed". No new tools. Order vs other open PRs: not checked.
Context cost none: no always-loaded file touched. bin/render-routing.sh --check on master: Baseline 65568 words (budget 80000). On the branch: 65568. skills/poc-build.md is on-demand (+~45 words). CONVERSION-RUNBOOK.md is not routed (+~25 words).
Behaviour NOT RUN: no A/B of the agent-read text changes (poc-build.md, kick-off brief) — maintainer to accept or request A/B. Another experiment will run simulated builds against this branch; its before/after is the intended measurement.
Runtime Full gate-check.sh on a fresh scaffold: 1.81–2.40 s on master vs 2.01–2.70 s on the branch (two rounds of 3–5 runs, this cloud container; noisy, about +0.1–0.3 s from sourcing one more lib and a per-file slot scan). Output grows by the printed shapes only when something FAILs (+10 lines on a fresh scaffold: the intake shape). kickoff.sh costs about 1.5 s, most of it the gate-check P it runs. Time cost: per gate run · +0.1–0.3 s on a Linux cloud container · can force rework: n · offset: removes gate reruns and gate-source reading (sim: 10 and 19 REVERSE calls, 8 and 28 GATE calls before the first exec).

Field evidence: two simulated claude -p builds on fresh init-project.sh scaffolds, with Mendix tools stubbed: a POC from a brief, and a requirements-driven build from a 232-word spec, 2026-11-20 (SIM-REPORT F1–F4). The fixtures replay the specific failures:

  • the POC build-ready FAIL and the fabricated sme-questions.md;
  • the slot FAILs with 0 dispatches;
  • the five-run prototype LOOK;
  • R03 row ids;
  • hand-written intake markers.

Not yet field-run on a real project. The follow-up simulated-build experiment against this branch is the first measurement.

  • No client data anywhere in the diff. Leak guard (no denylist on this machine): clean. Its warn-only capitalised words are "Rooms" and "Staff", fixture example names.
  • Size cap: 8 non-fixture files, +505/−33 lines, so it exceeds 400. This is a prototype branch meant to be split per commit (the four commits are independent).
  • Test tier: scoped fixtures, run against real init-project.sh scaffolds and the real gate. Not T3.
  • Golden input: the fixtures scaffold with the real init-project.sh (MXTK_LEAKGUARD_DENYFILE set to a temp file) and judge with the real gate-check / coverage-check. No tool output is parsed by the new code. Both layouts: the changes are register and intake readers plus architecture/, and the two-tree app/ layout was not separately exercised. Platforms: bash 3.2-safe constructs, check-portability clean, check-scripts.sh 147/147. Not run on macOS or Git Bash.
  • CHANGELOG.md line per commit, credited.
  • No new bug entries.
  • check-pr-discipline: clean.

Review: MERGE-AFTER-FIX (re-review of #828 at e18f18b: A/D must-fixes applied in the two commits after it, pending re-review; B MERGE; C HOLD for a maintainer ruling)

https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC


Generated by Claude Code

…mode: poc (#819)

Build type: poc now reads build-ready as N/A (exit 0, checklist still printed,
no open-questions escalation, no build-ready stamp), Stage 5's look check
reports MANUAL with the unseen pages named, and `poc` is an Entry mode token
that implies Build type: poc. The kick-off brief applies build-ready to full
builds only. Fixture tests/wave2/test-fast-lane-poc.sh: 15/15 on the branch,
6/15 on master.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
… is dispatched (#821)

Unfilled {{...}} slots in the six agent stubs are a named WARN, not a FAIL;
they FAIL again for mdl-agent once its dispatch log exists
(.claude/loop/dispatch/<Module>/scripts.tsv). Each stub still refuses to
run with a slot left. Fixture tests/wave2/test-fast-lane-agent-slots.sh:
8/8 on the branch, 3/8 on master; test-bug03-gates.sh T18 still names the
multi-line slot (99 ok).

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
…at FAIL (#821)

bin/lib/format-hints.sh holds a minimal passing example for intake markers,
the build plan, the coverage ledger, the LOOK report, the prototype LOOK and
the improvement-register row. gate-check.sh prints the shapes its FAILs owe
after the stage list; obligation FAULTs print theirs after the obligation
block; each shape at most once per run. No verdict or exit code changes.
Fixture tests/wave2/test-fast-lane-format-hints.sh runs every example through
its real parser: 13/13 on the branch, 7/13 on master.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
… lines (#821)

One call writes Entry mode / Build type / Interview mode lines, one Stage-P
Decisions row each, and an answer marker per intake section in the form
check_stage_P parses (flags, --answers verbatim, poc-build.md markers, auto
defaults ASSUMED; steering lists the unasked questions), then prints the
Stage P verdict. Idempotent; unknown values exit 2 with the accepted list;
changing a declared mode exits 3. Fixture tests/wave2/test-kickoff.sh: 26/26
(master has no kickoff.sh).

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
@MendixMau

Copy link
Copy Markdown
Owner Author

Review: MERGE-AFTER-FIX (per commit: A MERGE-AFTER-FIX · C HOLD · B MERGE · D MERGE-AFTER-FIX). Top finding: under A, Entry mode: poc overrides an explicit Build type: full (bin/gate-check.sh:697, project-bin/_position.sh:161). A POC→full switch that edits only the Build type line stays poc: build-ready N/A, Stages 0-4/6/7 waived, plan rule exempt. I reproduced this.
Value: A fixes #819 options 1 and 3. B and D fix #821 points 1 and 3. C fixes point 2 by overturning a ruling (see C). Nothing is overbuilt.
Regression risk:

  • A. ready.sh:178-217 greps look: in Stage 5. The new POC MANUAL line (gate-check.sh:2293) has no look: . ready.sh on a POC with 2 unseen pages now says "MANUAL with nothing measured — no page recorded as built" and "NOT MEASURED". Before, it named the 2 pages. I reproduced this. Existing limitation, now covering build-ready too: poc[!a-z]* (:695) reads Build type: poc → full (CR-01) as poc. sync-project.sh, status.sh and exec.sh only test other tokens.
  • C. This overturns the 2026-11-13 maintainer ruling (CHANGELOG.md:84, test-bug03-gates.sh:366). It also overturns the older rule that any single-line slot FAILs. After sync, a full build whose only open item was the stub slots flips to PASS. Under auto, once open questions evaluate, that PASS writes the sticky .build-ready-pass and clears exec's plan rule (exec.sh:815). The mdl-agent re-FAIL only fires after the agent has already run, since step 8a writes the log (agents/README.md:152). It cannot revoke the sticky stamp. The other five roles have no dispatch record. That reopens the "reviewer an inert stub" gap that gate-check.sh:2498-2503 documents, and that comment is now stale. Brief, BRD and open-question conditions remain, so the change is safe only if the maintainer accepts it.
  • B/D. Output only. Both layouts were probed: kickoff passes Stage P on an app/App.mpr scaffold.
    Correctness:
  • B. I fed each printed example to its gate. All pass, including report-disposition, which the builder left untested: report-disposition-check.sh went rc 1 → 0 with the example row. LOOK passes without the fixture's <br>.
  • D. Real init-project.sh scaffold, all 4 modes × steering/auto. Idempotent. Exit codes 2 and 3 are correct. Stage P PASSes under poc and auto. Under steering it FAILs, listing the unasked questions. But:
    • The default --by writes CONFIRMED … consentBy: the kick-off brief, which is not a person. That includes Interview mode: auto.
    • Under steering, an unstated build type is written Assumed: Full build with an ASSUMED row consentBy: auto-mode. Q16 is the question asked first (poc-build.md:8).
    • Brief line 2 tells agents to run kickoff before asking anyone.
    • D needs A: it cherry-picks with conflicts, and Entry mode: poc is unknown without A.
  • Fixtures (detached worktree): fast-lane-poc 15/15, agent-slots 8/8, format-hints 13/13, kickoff 26/26, bug03 99/0, auto-mode-hand-gates 28/0, stage-p 16/0, bug02 8/0, entry-mode-plan-rule 47/0, mode-contract 31/0, obligation-prototype-look ok, obligation-final-look ok, ready 29/0.
  • Toolkit bug (pre-existing, not this PR): the stray .leakguard-deny comes from test-mode-contract.sh:165,173 (project hand) and test-ci-gates.sh:163 (c6). Both call init-project.sh without MXTK_LEAKGUARD_DENYFILE, so the denylist path defaults to the clone's file (init-project.sh:631).
    Context cost: Baseline 65,568 → 65,568, measured.
    Time cost: Full gate-check on a fresh scaffold: master 2.38-2.70 s, branch 1.99-2.90 s. That is noise, roughly +0.3 s at most. No build-row cost.
    Exact fixes required:
  1. A: when Entry mode: poc and Build type: explicitly says full, resolve to full and print the conflict. Add a fixture case.
  2. A: keep look: N of M built page(s) never looked at … in the POC MANUAL line, or teach ready.sh the new line.
  3. C: maintainer ruling, recorded in process-learnings, and update gate-check.sh:2498-2503.
  4. D: CONFIRMED rows only with an explicit --by <person>; otherwise write open/unasked. Under steering/assist, leave an unstated Q16 unanswered and list it, with no ASSUMED row.
  5. D: brief line 2 should read "after the user answers Q1/Q9/Q16, record them with bin/kickoff.sh … --by ". Ship D only together with A.

Generated by Claude Code

…c; ready names POC pages (review of #828)

- Entry mode: poc with a Build type line naming full reads as a full build;
  gate-check and exec.sh print the conflict, and under auto exec.sh holds it
  to the requirements-driven plan rule.
- "Build type: poc → full (CR-01)" reads as full (gate-check and _position.sh).
- The POC Stage 5 MANUAL line keeps the "look: N of M ... never looked at"
  text bin/ready.sh parses; ready.sh reports that MANUAL as a named FAIL row.
Fixture test-fast-lane-poc.sh P5/P6: 21/21.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
… (review of #828)

- Rows and markers are CONFIRMED only with --by <person>; without it they are
  ASSUMED "no person named" (Interview mode: auto included); a later --by run
  upgrades kickoff's own rows and markers, then is idempotent.
- Under steering/assist an unstated --build-type / --interview leaves Q16 / Q9
  unanswered and listed first (no register line, no row); auto still assumes full.
- Refusal no longer calls an ASSUMED default row "decided".
- Kick-off brief line 2: ask first, then record with kickoff.sh --by <name>.
Fixture test-kickoff.sh K7/K8: 30/30.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
@MendixMau

Copy link
Copy Markdown
Owner Author

Review: MERGE-AFTER-FIX. Re-review of e18f18b. Per commit: A + 9a144ba MERGE-AFTER-FIX · C HOLD (unchanged, maintainer ruling) · B MERGE · D + e18f18b MERGE-AFTER-FIX. Top finding: the fix reads full whenever full appears anywhere in the Build type line (gate-check.sh:698, _position.sh:156). An existing POC whose line says e.g. poc — CONFIRMED (user, fullscreen kiosk) silently stops being a POC.
Value:

  • All five fixes from the first review landed and I re-probed each:
    • Entry mode: poc + Build type: full is now full, with a ⚠ from both gate-check and exec.
    • poc → full (CR-01) is now full.
    • The POC Stage 5 line keeps look: N of M … never looked at, and ready.sh names the unseen pages again.
    • kickoff writes CONFIRMED only with --by. Without it, rows are ASSUMED "no person named". A later run with --by upgrades only kickoff's own rows and leaves a person's row alone.
    • Under steering, Q16 stays unanswered and is listed first.
    • Brief line 2 now says to ask first and then record with --by.
  • I checked the unrequested exec.sh:802-809 addition (auto + conflict → requirements-driven plan rule) on master, the old head and the new head, with a stubbed exec. It is not a regression. On master, Entry mode: poc was an unknown token, and under auto that already refused every write. The old head let it through. The new head refuses again, now with a correct reason. Attended runs only get the ⚠. Entry mode: poc could not have been honoured before this PR, so no existing project's verdict tightens.
    Regression risk: There is one silent flip. Any existing register where Build type: poc … contains the substring full changes:
  • gate-check: Stages 0-4 go from WAIVED to PENDING.
  • exec: under auto, writes go from passing to REFUSED (my probe X4).
  • Neither prints a ⚠, because the conflict note fires only for Entry mode: poc.

It fails in the louder direction, but nothing says why.
Correctness:

  • New in e18f18b: kickoff.sh --mode poc --by X with no --interview leaves Q9 unanswered, so Stage P FAILs. The old head assumed steering. poc-build.md says Q1–Q15 take the POC marker and "kickoff … writes all of the above". Test K3 passes only because it uses --interview auto.
  • Everything else holds:
    • kickoff is idempotent in 7 flag combinations.
    • Exit codes are right: 2 for usage errors and a --by containing (/), 3 for a mode change.
    • The two-tree layout passes Stage P.
    • Every format-hint example still passes its parser.
  • Fixtures (detached worktree), all green:
Fixture Result
fast-lane-poc 21/21
agent-slots 8/8
format-hints 13/13
kickoff 30/30
bug03 99/0
auto-mode-hand-gates 28/0
stage-p 16/0
bug02 8/0
entry-mode-plan-rule 47/0
mode-contract 31/0
obligation prototype-look ok
obligation final-look ok
ready 29/0
exec-process-budget 4/0
greenfield-retired 13/0

The host disk filled up mid-run, so the affected fixtures were rerun once space was back. The stray .leakguard-deny reappeared after mode-contract. That is the pre-existing fixture bug (test-mode-contract.sh:165,173, test-ci-gates.sh:163), not this PR.
Context cost: Baseline 65,568 → 65,568, measured.
Time cost: exec-process-budget is green (132 process starts). The gate-check delta is within noise.
Exact fixes required:

  1. Read full only for an explicit switch: full as a whole word, or better, the (→|->|to) full form. Print the ⚠ whenever a Build type: line that starts with poc is read as full, in gate-check.sh and exec.sh both. Add a fixture case for a non-switch word containing "full".
  2. kickoff: when the build type is poc and --interview is unstated, give Q9 poc-build.md's POC marker. Add a K3 variant without --interview that asserts Stage P PASS.
  3. C: maintainer ruling (unchanged from the first review).

Generated by Claude Code

… warns (re-review of #828)

A substring test read `Build type: poc — CONFIRMED (user, fullscreen kiosk)` as a
full build, silently: Stages 0-4 owed and exec refused under auto. One shared
test, mxtk_bt_names_full (project-bin/_position.sh), now matches full as a
word ("full", "poc → full (CR-01)"), never inside one. A poc-prefixed Build
type line read as full sets the conflict, and gate-check and exec.sh both
print the warning. exec.sh only remaps the plan rule when the Entry mode is
poc. Fixture test-fast-lane-poc.sh P5/P7: 25/25.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
…eview of #828)

`kickoff.sh --mode poc --by X` left Q9 unanswered, so Stage P FAILed against
poc-build.md ("the other intake questions are not asked"). Q9 now takes the
POC marker when the build is poc and --interview is not given.
Fixture test-kickoff.sh K3 variant: 31/31.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01HXpQSjXefSNicGGqP4aiSC
@MendixMau

Copy link
Copy Markdown
Owner Author

Review: MERGE-AFTER-FIX. Third review, of head 616d67d. Per commit: A + 9a144ba + 51d72e7 MERGE · C HOLD (unchanged, waits on the maintainer's ruling) · B MERGE · D + e18f18b + 616d67d MERGE. The one remaining fix is C's ruling. Top finding: both must-fixes from my second review landed and are verified.
Value:

  • Fix 1, whole-word "full" (51d72e7).
    • One test, mxtk_bt_names_full (_position.sh:82), is shared by gate-check and exec.
    • poc → full (CR-01) now reads as a full build and prints a ⚠ in both tools.
    • poc (fullscreen) and poc — CONFIRMED (user; fullscreen kiosk demo) stay poc.
    • Entry mode: poc with no Build type line stays poc.
  • Fix 2, Q9 without --interview (616d67d). kickoff.sh --mode poc, with or without --by, now gives Q9 the POC marker, and Stage P PASSes.
    Regression risk:
  • The builder's exec.sh change (exec.sh:806-812, the auto remap now applies only when Entry mode itself is poc): not a regression.
    • With a real full-build Entry mode (requirements-driven or migration), the normal plan rule already applies once POS_BUILD reads full. Probe X6 is REFUSED with ⚠.
    • With an existing-app or a-la-carte entry mode, the project behaves like any full build of that mode.
    • With no Entry mode line, the stricter "undeclared" hold still applies under auto.
    • Remapping those cases would have changed their mode contract. Limiting it to Entry mode: poc is correct.
  • One intended flip for synced projects. A register whose Build type: poc … line names the word full changes from POC to full, now loudly (⚠ in gate-check and exec).
    Correctness:
  • Stubbed-exec probes:
Probe Register Result
X1 Entry mode poc, Build type full, auto REFUSED + ⚠
X2 same, attended ⚠ only
X3, X5, X7, X8 poc readings pass, no ⚠
X4 "fullscreen" in the Build type line stays poc (fixed)
X6 poc → full REFUSED + ⚠
  • kickoff probes (check4): all 7 flag combinations stay idempotent.
  • Fixtures in a detached worktree: test-fast-lane-poc 25/25, test-kickoff 31/31, test-entry-mode-plan-rule 47/0. Nothing stray was left in the tree.
    Context cost: Baseline unchanged at 65,568 words. The two commits touch no routed file.
    Time cost: One extra bash regex per position read, no fork. exec-process-budget was green at e18f18b; I did not rerun it here, because this change adds no process.
    Exact fixes required:
  1. C: maintainer ruling on stub slots becoming a WARN, which overturns the 2026-11-13 decision. Record it in process-learnings and update the stale comment at gate-check.sh:2498-2503. Otherwise drop C from the PR: A, B and D merge without it. B and C cherry-pick cleanly apart from CHANGELOG.md.

Generated by Claude Code

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

2 participants