[lukemaj]

What: The anti-pattern catalog gains nine routing and harness failure modes measured in the AgentsMD benchmarks, and the mechanism page gains two AgentsMD cases for measuring and rewording before building a hook.
Why: The catalog covered only karpathy-wiki failures, so authors of routing tables and trigger harnesses had no checklist entry for the failures that cost AgentsMD invalid benchmarks and missed procedures.
So what: Reviewed and approved on the exact head; the user merges.

Closes #13
Refs #7

Changes

docs/10-anti-patterns.md: nine entries, each with Definition, Evidence and a Counter linking /docs/05-authoring/triggers, /docs/06-testing/trigger-benchmarks or the case study. The unsourced "OpenCode ... less affected" flag (:53) is replaced by the claim boundary: the reported case is on Claude; no other host has been measured.

docs/07-mechanism-vs-decoration.md: new section "Measure, reword, mechanism last: two AgentsMD cases"; the paths: trade-off now names Claude Code and Grok Build with primary sources.

public/llms.txt: regenerated.

Numbers and sources

Entry	Number	Source
Trigger that does not name the action	verification before gh pr create 0/5, then 5/5	toolboxmd/agentsmd#181, #184
Bare conjunct	SKILL-MECHANICS.md 15/15, prose.md 4/15 attempts; both 5/5 with own before-clause	174-routing-injection/records-prefix.jsonl (first batch, arms A/B/C, recounted), #179
Abstract task-start rule	0/40 naive; 20/20 with pointer; fix1 verification 5/5 to 3/5, research on question 2/5	#164 results, #174 results, #184; fix1 3/5 regenerated with --summary on 182-routing-sweep/records.jsonl
Read in the same call	diagnosis 8/10 on fix2, 5/5 on fix3	#182 results (regenerated with --summary), #184
Always-loaded text tuned on one host	OpenCode research on question 5/5 on fix2 (0/5 clean), 5/5 clean after exclusion	case study, #184
Refused read scored as a read	17/20 with every read refused; 10 of 17 never edited	records-prefix.jsonl, #179
No-action run scored as pass or miss	10 of 17 never edited; 12 verdicts changed on #182 rescoring, 10 no-action	#179, case study, #184
Green harness tests	349 tests, self-checks on four hosts, APPROVE	#168, case study, #179
Unconfined harness shell	pilot wrote to the canonical checkout; Bash open; write-only sandbox	#164 results, #181, #183
Mechanism page	9 of 11, then 5 of 13 Issue-creating sessions skipped the Elon method; hook added	toolboxmd/agentsmd#138, #161
Mechanism page	verification hook proposed, conditional, not built after 5/5	#181, #182, #184
paths: hosts	Claude Code, Grok Build	code.claude.com/docs/en/skills; xai-org/grok-build conditional.rs at 97f190f

Deviation from the ticket: entry 2 uses 4/15 against 15/15 (as published in triggers.md, recounted from the records) instead of the ticket's 7/22 against 22/22. #161 extended the existing PreToolUse hook rather than registering a new one; the page says so. No Haiku numbers.

Elon record

• Requirements and who asked: user, 2026-09-29 (Clean up the repository and apply the routing and benchmark learnings #7, Add the routing and harness failure modes to the anti-pattern catalog and the mechanism page #13): catalog entries for every measured routing and harness failure, and the reword-before-mechanism evidence.
• Deleted: restating the case study (entries link it and the method pages); a separate harness page; the ticket's 7/22 figure, which the records do not reproduce for one batch.
• Bottleneck: the self-audit catalog had no routing or harness entry.
• Checked myself: regenerated 182-routing-sweep with trigger_test.py --summary at 06b7af6 (diagnosis 8/10 fix2, verification 3/5 fix1); recounted the writing-for-agents attempts in records-prefix.jsonl; read #138, #161, #181, #182, #184 for the mechanism sequence.

Proof

• npm ci && npm test under Node 22.23.1: all checks pass (links, llms freshness and coverage, absolute paths, bare-path gate).
• npm run build:llms regenerated public/llms.txt.
• Independent review: Luna (gpt-6-luna) max, REQUEST_CHANGES on da85d36 (bare-conjunct figures blurred attempts and reads; fixed), then APPROVE on 79d0591.

🤖 Generated with Claude Code