Skills trial: remove verification-before-completion under the conflict rule (M4), #381 Amendment 3

[seathatflowsinourveins]

Summary

This PR removes the trial skill verification-before-completion (obra/superpowers@8ca22dba9a94f28898bbce59f2537ff4d87c747d) under the trial's in-force conflict rule, docs/decisions/2026-09-25-skills-trial-and-usage.md:284-286: "gives instructions that conflict with CLAUDE.md/AGENTS.md once actually read in full → remove it immediately".

The conflict:

• The pinned SKILL.md puts "Previous run" under Not Sufficient for "Tests pass" (L42). It also says "If you haven't run the verification command in this message, you cannot claim it passes" (L20) and "Execute the FULL command (fresh, complete)" (L28).
• AGENTS.md:16 says "Reuse passing evidence when its inputs still match and run only checks needed for a concrete gap".

Findings. They are recorded in docs/decisions/2026-09-28-delegated-decisions.md (#462) and evidence/artifacts/delegated-decisions-20260928/:

1. The trial owner's full read.
2. The practice-sweep reviewer.
3. GPT-6 (the pinned request was gpt-6-astra at max, read-only): "VERDICT: conflict", disposition "remove now".

Findings 1 and 2 are not shown to be independent. coordination.md records that one reading was shared before the GPT-6 verdict.

Changes

• Skills manifest: the row moves to excluded, with a dated reason that quotes both sides. The budgets are recomputed: Claude from 7,409 to 7,184, and Codex from 3,054 to 2,829. The template's skillOverrides key is removed.
• isolated-builder (all three copies, byte-identical, sha256 0f8e0834…): the preload and the body line that named the skill are removed.
• adoption/hooks/claude/token-lanes-block.builder.md: gains the default block's evidence sentence, so the builder keeps an evidence rule. The token-lanes addendum records why Token-lanes carrier: evidence sentence replaces the verification-skill line (AN-15, after #447) #456's removal no longer applies. The new block sha256 is c81a91c4ffb2fa906feb403d5a11bd294b6cf2db45795d80f9acd662c5004973.
• Landscape-sweep templates and TEMPLATE_SKILLS: drop the skill. "Evidence before any verdict" stays as plain text, and the prompt hash is updated.
• Workflow suites: they now pin the builder preload set as ['context-mode:context-mode'].
• PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381 token-E2E preregistration, Amendment 3 (pre-execution, authorised by the Gate A owner):
  • a new isolated-builder.md role-body row 0f8e0834… and a new RUNBOOK seal 135340b6…;
  • the d022295a and Amendment 2 rows kept as history;
  • tests/test_token_e2e_preregistration.py pins both.
• Current-state prose: updated to match the removal. Dated records stay unedited.

Host steps after merge

These run on this host only, and a receipt comes in a follow-up:

1. DISABLE_TELEMETRY=1 skills remove verification-before-completion -g -y -a claude-code codex, through the pinned skills 1.7.0 CLI.
   • -a limits the removal to the two agents the installer adds it for (install_skills.py L60-62). Without it, remove targets every known agent.
2. python3 tools/adoption/install_claude_profile.py --only guard --only agents.
3. A python3 scripts/skills_status.py read-back.

Review

• Gate A owner (native-agent-stack-a9) reviewed Amendment 3 at d98ff7d0 and approved it with no blocking findings. It re-derived the role-body and seal hashes and re-ran the preregistration, token-lanes and install-profile tests (109 OK), both node suites and validate.py.

• GPT-6 cross-family review (gpt-6-astra, max, read-only): VERDICT: clean. Usage: 1,207,847 input tokens (1,094,784 of them cached) and 15,607 output.

• Independent stack-verifier: it confirmed the tests, all hashes, the budget recompute (7,184 and 2,829, by both the test predicate and the named method) and the registration (26 of 26). A scratch-HOME install_claude_profile.py --home ... --only guard --only agents showed:

  • the installed builder preloads only context-mode:context-mode;
  • its block holds the evidence sentence;
  • the hook injects that block.

  It found two low defects:

  1. Fixed in 29ed7d98: the post-patch-validation exclusion reason still assumed the removed skill.
  2. Kept, and recorded here: Amendment 3 edits three earlier README spots in place: a sentence at L228, the arms-roles table row at L234, and a note in Sealing. Each edit is dated and listed in Amendment 3's change rows, and none changes a hash or seal. Amendment 2 (PR-H (#381) Amendment 2: adopt the #402 role bodies before execution (Opus verifier and builder, builder base binding) #413) set the same precedent, and the Gate A owner approved it. The Sealing rule's "never a silent rewrite" holds; strict append-only does not.

Checks

These are our integration checks, run after merging origin/main at c61e6657:

• The skills-manifest, skills-status, install-profile, token-lanes, landscape-sweep, docs-consistency, token-E2E-preregistration and grand-dashboard unittests: Ran 350 tests, OK (skipped=7).
• test-envelope.mjs: 254 of 254 pass. test-contract-mutations.mjs: 74 of 74 pass.
• The three registry test methods: OK. python3 scripts/validate.py: passed.
• Failing-first controls on the base: token-lanes failures=2, envelope 253 of 254, skills manifest failures=3, install_claude_profile failures=1, landscape prompt-hash mismatch.

SOTA sources

• obra/superpowers@8ca22dba9a94f28898bbce59f2537ff4d87c747d, skills/verification-before-completion/SKILL.md (sha256 2befe7fc…), read in full on 2026-09-28. Upstream main has no newer commit touching the skill: the last is 3be5aad3, from 2026-07-24.
• The upstream skills CLI 1.7.0: remove <name> -g -y -a <agents>.
• Claude Code sub-agents docs on preloading skills.

🤖 Generated with Claude Code