Token-lanes carrier: evidence sentence replaces the verification-skill line (AN-15, after #447)

[seathatflowsinourveins]

Scope

• What this PR changes: this is AN-15 of the 2026-09-27 Claude Code practice review (Claude Code practice for Opus 5.5: 2026-09-27 review resolved, host read-back, weekly reference freshness lane #453). It changes the SubagentStart TOKEN LANES carrier's verification line, now that Token-lanes carrier: role-matched blocks name only the lanes each agent's allowlist grants #447 made the carrier role-matched:
  • Default block (adoption/hooks/claude/token-lanes-block.md line 13): the line that told children to use the verification-before-completion skill becomes an evidence sentence: "Show evidence before a success claim: the command and what it returned, or the file:line read". The upstream item it follows is named.
  • Builder block (token-lanes-block.builder.md): the line is removed. isolated-builder already preloads verification-before-completion in its frontmatter (adoption/agents/claude/isolated-builder.md L7-9), and its body tells it to use that skill, so the carrier line was a third copy.
• Base commit: 718e5620
• Lane: lane:foundation. manifests/evidence.json only re-registers hashes.
• Owned paths touched:
  • adoption/hooks/claude/{token-lanes-block.md,token-lanes-block.builder.md,SHA256SUMS}
  • docs/token-session-handbook.md (the carrier mirror)
  • tests/test_token_lanes_subagent_start.py
  • docs/decisions/2026-09-27-token-lanes-subagent-start.md (addendum)
  • docs/decisions/2026-09-25-skills-trial-and-usage.md (addendum, placed after Skills trial: cloudflare/security-audit-skill name-only for LLM-native automation; stale-upstream flags; sandbox-profile open gate #448's)
  • manifests/evidence.json

SOTA sources

Each page below was re-read at its cited line of the Markdown source on 2026-09-28:

• Prompting Claude Opus 5 L61, L74, L81: verification instructions cause over-verification and should be removed.
• Prompting Claude Opus 5.5 L9: the Opus 5 patterns are "a reasonable starting point".
• Prompting best practices L780: Opus 5 is the exception to the self-check advice ("remove these instructions rather than rewriting them").
• Claude Code best practices L52: show the command and what it returned. The added "or the file:line read" is tagged as a local extension.
• Sub-agents L600: the full content of each preloaded skill is injected at startup.
• obra/superpowers@8ca22db skills/verification-before-completion/SKILL.md L10, "Evidence before claims, always". This is the preloaded skill; the installed copy matches the manifest skill_md_sha256.

Evidence-class table

Claim	Evidence class	Command / receipt
The hook emits the new default block (4,088 bytes), the builder block (2,747), the scout block (1,089) and no block for blind roles, all within the 4,100-byte budget and each allowlist	local_integration (subprocess runs of the hook with SubagentStart payloads)	tests/test_token_lanes_subagent_start.py; Codex re-ran the hook for every shipped role
The new assertions fail on the old carrier and pass on the new	synthetic (the module run against the #447 blocks: 19 subtest failures; now 17 tests OK)	the addendum's failing-first control
No child start context on a host shows the new line yet	not run: needs the reinstall	see Residuals

Local commands run

$ cd adoption/hooks/claude && sha256sum -c SHA256SUMS
all 9 OK
$ TMPDIR=<fresh> python3 -m unittest tests.test_token_lanes_subagent_start tests.test_install_claude_profile tests.test_adoption_docs_consistency tests.test_context_mode_practice_docs tests.test_effort_default_guard tests.test_render_config tests.test_skills_manifest tests.test_token_e2e_preregistration tests.test_adoption_status
Ran 307 tests, OK (skipped=4)
$ python3 scripts/validate.py; python3 scripts/component_matrix.py --check; python3 scripts/new_host_grand_list.py --check; git diff --check origin/main...HEAD
exit 0; exit 0; exit 0; exit 0

Decision record

• "Addendum 2026-09-28: verification line" in docs/decisions/2026-09-27-token-lanes-subagent-start.md. It covers the need, the default change, the builder decision, sizes, tests, the non-Opus boundary, rejected alternatives, the PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381 freeze and the overturn condition.
• The carrier-condition log (C0-C3) and the review rule in docs/decisions/2026-09-25-skills-trial-and-usage.md.

Review

1. Claude workflow review.
   • A build, verify, review, repair and final-verify pipeline ran in an owned worktree.
   • Verify passed with 3 minor findings; the review passed with 7 minor findings and 9 verification gaps.
   • The repair fixed all findings and measured six of the gaps.
   • Final verify: pass, with integration residuals.
2. Rebuild onto main. The branch was rebuilt onto main 718e5620.
   • Skills trial: cloudflare/security-audit-skill name-only for LLM-native automation; stale-upstream flags; sandbox-profile open gate #448's skills-trial addendum and this one both append at the end of that record. The resolution keeps Skills trial: cloudflare/security-audit-skill name-only for LLM-native automation; stale-upstream flags; sandbox-profile open gate #448's first, as the final verifier tested.
   • The diff-of-diffs against the reviewed head b8f1205b is identical outside the manifest, and the record's added lines are identical.
3. Cross-family review. codex exec -s read-only -m gpt-6-astra -c model_reasoning_effort=max on the rebuilt head returned changes-needed with one minor finding: the "no Bash means no command output" rationale ignored ctx_execute, which security-reviewer has. Fixed by qualifying the rationale. Codex's hook runs, allowlist checks and mirror checks all passed.

Residuals

• After merge: reinstall with python3 tools/adoption/install_claude_profile.py --only guard, then check that one new Agent-tool child's start context shows "Show evidence before a success claim". I'll do this on this host.
• Timing with PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381 / Gate A: like Claude Code practice for Opus 5.5: 2026-09-27 review resolved, host read-back, weekly reference freshness lane #453, this should land before the organic Gate A run starts, never during one. The token-save-practice E2E owner confirmed at about 09:15Z that none is active.
• Handbook scope note: the L217 table row and the L286 sources paragraph change beside the mirrored lines. Review findings asked for both edits.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment (no floating tags). (none changed)
• New/changed workflows declare top-level permissions: contents: read (or a narrower, explicitly justified addition). (none changed)
• No secrets are printed, logged or committed; no new required secret was added without a documented owner.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved (not deleted, moved or overwritten).

🤖 Generated with Claude Code

[seathatflowsinourveins]

Post-merge acceptance on this host (2026-09-28, Claude Code 2.1.283, local observation):

• Install. python3 tools/adoption/install_claude_profile.py --only guard, run from a clean detached worktree at 85a604ac.
  • The dry run listed exactly two installs, token-lanes-block.md and token-lanes-block.builder.md. The other seven hook files already matched and were skipped.
  • The real run exited 0. The installed sha256 prefixes equal the repository's: d6c3c19d6e94d3f4 and c665c230cc81b451.
• Default child (Agent tool, general-purpose). Its injected TOKEN LANES block is present. The block's last line is the new "Show evidence before a success claim: …" line, and the block does not name verification-before-completion.
• isolated-builder child. Its injected block is present. It has no "Show evidence" line and does not name verification-before-completion, and its last line is the one-lane line ("Use one lane per artifact; …").
• Cross-family review. Codex returned changes-needed with one minor finding: the rationale ignored ctx_execute. That was fixed before merge (see the PR body).