Repository navigation
Weekly context audit 2026-10-26 #575
Description
First weekly audit. Read-only against origin/master @ 6615869. No earlier audit issue exists, so the "ref" column uses the 2026-10-23 measurement on 888b4b2 (process/task-routed-rules-proposal-2026-10-23.md on branch experiment/task-routed-rules).
Numbers (bin/render-routing.sh --check)
| Now | Ref 10-23 | Δ | |
|---|---|---|---|
| Baseline total / budget | 63,877 / 80,000 | ~79,900 | ≈ −16,000 |
| Stage 4: mdl / gate / test / review | 11,029 / 7,894 / 7,894 / 6,842 | n/a | n/a |
| Stage 5: mdl / gate / test / review | 14,600 / 22,425 / 21,259 / 30,856 | 30,443 / n/a | mdl −15,843 |
| Stage 6: mdl / gate / test / review | 9,476 / 21,437 / 21,046 / 25,394 | n/a | n/a |
The lazy-load work (#570–#572) cut the mdl role's preload by half. review at Stage 5 is now the heaviest role, at 12 files.
Largest Stage 5 files (words, roles that read them):
testing-shape.md: 7,224 (gate, test, review)module-review.md: 5,917 (test, review)bin/mdl-rules.sh: 3,692 (mdl)module-brief.md: 3,245 (mdl)learned-detection-gaps.md: 3,240 (gate, review)ui-loop.md: 3,148 (review)
Largest sections:
testing-shape.md: §4 false-green register 2,141 · §5 stack precondition 1,523 · §3 1,036module-review.md: "Stage 4 — intelligence check" 2,438 · "Stage 5 — report" 1,014
Top 5 move candidates (words saved per load)
| # | Where | Words | Why movable | Destination |
|---|---|---|---|---|
| 1 | bin/mdl-rules.sh:31–end |
~3,250 | The routing row counts a script, so the mdl role preloads its source code. Only the header (443 words, the usage contract) is needed; the helper runs it. | Count/read the header only. This is the "script headers" piece on the experiment branch. |
| 2 | testing-shape.md:350 §5 |
~1,100 | Docker/stack mechanics with a narrow trigger: it only matters when e2e is about to run. Keep the rule ("you bring the stack up; never infer up from no error"). | Move the command table and mechanics to e2e-harness-base.md and leave a pointer. |
| 3 | module-review.md:149 §4d + :232–260 4e/4f |
~900 | Page/design look checks, read by test too, which never does a LOOK. | Make the row on demand for test (test?). The review role keeps it. |
| 4 | ui-loop.md:153 "What the loop costs — measured" |
~370 | Past-numbers table (movable per CLAUDE.md budget rule 2). | process/ doc, with a one-line pointer. |
| 5 | testing-shape.md:207 §4 register: incident dates and narrative clauses |
~600 (est.) | Guards and the failure column are protected; the dated narratives inside the cells are not. | Move the narratives to process/false-green-incidents.md, keeping one line each. |
authoring-audit.py dupes: all top pairs are skills/generated/rules-*.md against their sources. That is expected (they are rendered), so there are no new duplicates in baseline files.
Re-probe list (newest stamp in repo: mxcli 0.27; ≤ 0.24 is 3+ minor versions behind)
mxcli is not available here, so none of these are claimed obsolete. Re-probe per skills/retesting-learned-rules.md.
learned-detection-gaps.md:45:ifwithoutelseinsideon error {}(v0.24.0)learned-detection-gaps.md:46: reserved-word entity/attribute names (v0.24)learned-detection-gaps.md:52:get workflows for $Xwith no context workflow (v0.24.0)learned-mdl-preflight.md:13: "on v0.24.0 it writes everything MDL can express" (the default-path claim)
Nothing found in the 2026-09-30 research doc §6 marked fixed-in-mxcli or caught-by-mxcli-check under those labels.
Task-routing proposal (A/B/C)
B overlaps, A partly overlaps. Lazy-load PRs 1–3 delivered per-statement-kind rules via mdl-rules.sh. Still unlanded from the branch: task packs in context-pack.sh, script-header counting (candidate 1), the #565 trigger rule, and the A/B tooling. C (reference projects) has not started.
Decision needed: which moves to make
- 1. Script-header counting for
mdl-rules.sh(−3,250 mdl, Stage 5) - 2. Move the
testing-shape.md§5 mechanics toe2e-harness-base.md(−1,100 gate/test/review) - 3.
module-review.mdon demand for test (−900+ test) - 4. Move the
ui-loop.mdcost table toprocess/(−370 review) - 5. Move the false-green incident narratives (−~600 gate/test/review)
- Re-probe items 1–4 on the current mxcli
Activity
MendixMau commentedon Oct 26, 2026
Audit read and logged on the roadmap: #579. Its five text moves (−3,250 / −1,100 / −900 / −370 / −600 words) are each a trim PR for the maintainer to pick. The 4 rules written against v0.24 need re-probing on v0.28.
Generated by Claude Code