Promote 2026-09-29 inbox notes: exec.sh refusal rule, UI fix rounds, transcript usage, autocompact advisory

[MendixMau]

Triage of the 2026-09-29 field-comparison inbox notes into on-demand skills. Baseline is unchanged at 79,997 words (render-routing.sh --check, before and after). Nothing new went into a baseline file. ui-loop.md and learned-css-that-never-applied.md are both baseline, so the #170 text went to the on-demand learned-page-patterns.md.

#69: "always exec.sh" (Refs #69)

Problem: agents hitting an exec.sh guard had no stated move, and the predictable one is FORCE_EXEC=1. The MCP path's recording status was undecided.
Fix: a new section in skills/iterative-build-loop.md (on-demand):

• Every model write goes through exec.sh.
• A refusal table: commit (--commit-first), wait (exec lock), or stop and report (Studio Pro holds the model; gate cannot run → doctor; ALLOW_UNVERIFIED only on the person's say).
• Never FORCE_EXEC on the agent's own initiative, with the failure example (silent loss of MCP work on auto-restore).
• MCP writes stay allowed and are accepted as unrecorded. Their record is an MCP: commit.
• Phase-close denominator: model commits = exec rows + MCP: commits.

The mdl stub gets this only through routing: the trigger is now "…, or an exec.sh guard just refused a write". Stubs were re-rendered.
Verified against project-bin/exec.sh: the uncommitted and --commit-first refusals log a 🚫 row; the Studio Pro and lock refusals only print. The text says this.
Left open (why Refs): decision 4, which projects get --upgrade-bin exec.sh. That is a sync/rollout call, not skill text. Decision 3 is already answered by #304's mxbuild auto-download plus refuse.

#170: UI fix rounds, N1–N4 (Refs #170)

Fix: skills/learned-page-patterns.md → "UI fix rounds" covers:

• the round order: SCSS first; apply directly via exec.sh with no scratch pre-apply; re-look only changed pages, N of N;
• a pre-apply table for N1 (page-column on the layout root), N2 (.d-none losing to a display class), N3 (DG2 autoFit at 32 px) and N4 (pointer to the existing focus-ring section);
• an M-of-M denominator.

The "halves calls" figure is labelled a hypothesis.
Left open: N1–N3 as a mechanical check in project-bin/check-page-shell.sh. That needs captured golden page MDL/SCSS, and none was available.

#171: batched gate and usage measurement (Refs #171)

Fix:

• The build loop states what the gate already saves: the stamp skips the baseline mxbuild, and the lock refuses a second apply (it does not queue); re-run after the first finishes.
• It also says there is no batched-gate mode, so do not imitate one with SKIP_*, FORCE_EXEC or raw exec.
• New skills/measured-claims.md §7: read each message's last usage record. Re-measured on real Claude Code 2.1.296 transcripts:
  • the main loop had 941 messages with 0 disagreeing, so it is exact;
  • sub-agent first-record output was up to 7x low (1,151 vs 8,241).

Proposed, not done (owned elsewhere):

• a batched-gate mode in exec.sh;
• bin/token-burn.sh keeps the FIRST record per message.id (it skips ids already seen), so it undercounts sub-agent output. Fix: keep the last record's usage per id.

#173: save-then-check and MCP asks (Closes #173)

Fix:

• A save-then-check subsection in the build loop (failure example: 229 errors hidden by a stale disk copy).
• skills/upstream-feedback.md gets a third destination, the Studio Pro MCP server's owners: draft locally, never file in either repo. It carries the asks with their evidence rows and a "done when" bound.
• The per-page visual check already existed (build loop step 8 / ui-loop.md).

#166: context diet, measured (Refs #166)

Measured with chars/4 on this branch, lead role, baseline skills only. CLAUDE.md and CLAUDE.local.md are not included.

Stage	Full runbook read	Runbook as §1b + own stage section
0	~43k	~22k
1	~52k	~30k
2	~55k	~32k
3	~55k	~32k
4	~55k	~32k

The runbook alone is 15,216 words (~24.7k tokens). The ≤25k target is not met for stages 1–4 even with the slice. No change in this PR; a spine split is still needed.

#167: autocompact advisory (Refs #167)

Fix: bin/init-project.sh prints:

• a WARNING when CLAUDE_CODE_AUTO_COMPACT_WINDOW, or an autoCompactWindow number in the project or user settings.json, is above 250k;
• an "ok" line when it is at or below 250k;
• otherwise a /autocompact 250k recommendation.

It writes nothing and never fails the scaffold. An unparseable value is ignored. The setting names were checked against the installed Claude Code 2.1.296 binary, and the per-model settings shape matches a real ~/.claude/settings.json.
Left open: lean per-role payloads, a bundle cap, and per-role context in the time log.

Inbox

• 2026-09-29-mcp-build-observations.md: deleted, fully promoted.
• ui-loop-efficiency, parallel-build-hygiene, subagent-context-cost: trimmed to what stays open.
• context-diet-early-stages: untouched, since nothing from it was promoted.

Checks run

• New fixture tests/wave2/test-init-autocompact.sh bin/init-project.sh: 15/15 pass; 4 fail on master's init-project.sh.
• Scoped fixtures for changed files:
  • test-bug09-13.sh (iterative-build-loop): 22/0
  • test-docs-numbering.sh: 16/0
  • test-routing-tier.sh: 12/0
  • test-baseline-pack.sh: 16/0
  • test-render-routing.sh: 5/0. The first run reported 1 FAIL that did not reproduce on rerun, on this branch or on master.
• check-scripts, check-portability, check-no-private-citations, render-routing --check, the leak guard (no denylist) and check-pr-discipline origin/master: all clean.
• Banned-word grep of added lines: 0 hits.

Field run

The init advisory was run against four scratch scaffolds (no setting, env 1M, settings 250000, settings 1000000), and each printed the expected line. A field run on a real project's scaffold is still owed. The skill text has no instrument to field-run.

Review

Independent review: MERGE-AFTER-SMALL-FIX; fix round landed; confirmation round: MERGE-AFTER-SMALL-FIX (kept-write wording, PARTIAL coverage, print-only list, this body) — fixes applied verbatim as specified.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VJgWP5vEoAsNsJYqCDMGNw

---

Generated by Claude Code

---

Generated by Claude Code

[MendixMau]

Landed on master as part of the integration squash #329 (commit 31e90ae), after independent review — its Review: line is in that commit message. Closing as superseded.

---

Generated by Claude Code