Practice gaps #6 and #8: truthful skills-install message; agent teams are an allowed dispatch mode - #388

What

This PR fixes two of the eleven Claude Code / Codex practice gaps from the 2026-09-27 GPT-6 layer discovery. An evidence-reviewer pass verified all eleven against their primary sources: 6 confirmed, 5 partly confirmed. The token-lane session takes #1, #4, #5, #7, #9, #10 and #11, and #2 and #3 are skipped as not recommended.

#8 — agent teams (CONFIRMED)

• examples/claude-native/CLAUDE.md:37 said: "agent teams stay off because each teammate is a separate instance whose results are not summarized back". That was wrong on two counts:
  • the Claude Code docs say a finishing teammate "automatically notifies the lead and includes its final answer in the notification" (https://code.claude.com/docs/en/agent-teams#context-and-communication);
  • the operator's 2026-09-27 instructions make agent teams an allowed dispatch mode for parallel exploration.
• The sentence is replaced word-neutrally, because the portable template has an enforced budget (881 words + 5%) and stays at 924:
  • new text: "Agent teams suit parallel exploration where teammates message each other; they cost more; the lead verifies their results."
  • The operator's full dispatch-mode section lives in their global instructions.
• scripts/adoption_status.py: the agent_teams_off wiring requirement, which flagged this host after the opt-in was set at the operator's request, becomes agent_teams_opt_in.
  • It is a 0/1 count reported as information.
  • A count, not a bool, never enters the all-booleans completeness rule, and the output keeps its flags-and-counts-only invariant.
• docs/token-efficiency-stack.md: the coverage-rule text now matches.
• docs/harness-rules-convergence-20260922.md WH-3: marked superseded, keeping the original text.

#6 — skills install message (CONFIRMED)

• tools/adoption/install_skills.py:191 said "installed and verified at tree". But classify_skill checks only SKILL.md's sha256 and the lock's recorded tree. skillFolderHash is the source tree hash recorded at install (vercel-labs/skills v1.7.0 src/skill-lock.ts:25-28), not a hash of the installed support files.
• The message now says what is actually checked. The historical receipt that quotes the old message is unchanged.

Out of scope, flagged for the operator. examples/claude-native/CLAUDE.md and the operator's global ~/.claude/CLAUDE.md also differ in the top-rule and token-practice wording. The repo has #371's approved procedure; the global file has older wording. Which one is canonical is the operator's call.

Tests

• New: test_agent_teams_opt_in_is_information_and_never_blocks_completeness.
• Updated: the fixtures and test_effort_and_agent_team_opt_ins_are_found_by_name_whatever_their_value.
• python3 -m unittest tests.test_adoption_status tests.test_adoption_docs_consistency tests.test_install_claude_profile tests.test_install_skills tests.test_skills_manifest: 213 run, OK (4 skipped). This includes the portable template's word-budget and negative tests.
• scripts/validate.py and scripts/build_ecosystem.py --check: passed.

SOTA sources

• Claude Code agent teams: https://code.claude.com/docs/en/agent-teams#context-and-communication
• vercel-labs/skills v1.7.0, src/skill-lock.ts (skillFolderHash semantics)

🤖 Generated with Claude Code

https://claude.ai/code/session_012qwXNtFyUhY5x2dtrm7tG5