Coordinator effort: terminal max through the launcher, saved xhigh fallback; AGENTS.md line 35 and the instruction text state the 2.1.284 rule - #487

Scope

• What this PR changes, in one or two sentences: states the coordinator-effort rule of the max-default record (Launcher: interactive terminal launches start at effort max (--effort max only when nothing chose an effort); receipt and decision record #483, session native-agent-stack-03) where the repository describes it: a terminal session started through the ecosystem claude launcher runs at max (claude --effort xhigh opts out), the saved per-model xhigh is the fallback for launches that skip the launcher, headless -p runs pass --effort per call site, CLAUDE_CODE_EFFORT_LEVEL stays unset at every scope, and Ultracode stays on at any effort on 2.1.284, so the 2.1.281 reason for keeping the coordinator at xhigh (AGENTS.md line 35) is retired. It also adds one receipt: an explicit --effort max defeats a committed project effortLevel: xhigh, which the launcher relies on inside this repository.
• Base commit: c26800f3 (origin/main at the last rebase; it includes Sonnet 5.5 and Opus 5.5 dispatch: fan-out units on Sonnet 5.5, judgment on Opus 5.5, effort and default child model saved in settings (2.1.284 probes) #479 and Launcher: interactive terminal launches start at effort max (--effort max only when nothing chose an effort); receipt and decision record #483)
• Lane: lane:shared. It changes what a shared hot file says (AGENTS.md line 35, the last commit, with the hash registration in manifests/evidence.json), so it needs the trading lane's acknowledgement; the rest is foundation-lane.
• Owned paths touched: examples/claude-native/CLAUDE.md, examples/claude-native/workflows/{README.md,test-envelope.mjs,test-contract-mutations.mjs,SHA256SUMS} (the coordinator row and the two tests that pin it), recipes/claude-native-{ultracode,profile}.md, docs/decisions/2026-09-23-max-effort-default.md (its addendum), docs/decisions/2026-09-29-sonnet-5-5-dispatch.md (coordinator row, effort paragraph, dated addendum), docs/harness-defaults.md (one cell), tools/sota-convergence/landscape-sweep/README.md (one word), evidence/receipts/claude-project-effort-flag-20260929.json (new); shared: AGENTS.md and manifests/evidence.json.
• Not touched: the launcher and its tests (Launcher: interactive terminal launches start at effort max (--effort max only when nothing chose an effort); receipt and decision record #483), the guard, every settings file, every agent definition and workflow, and the PR-H: E2E preregistration for the token-save-practice stack (frozen tasks, metrics, gate, procedure) #381 files. The host ~/.claude/CLAUDE.md takes the same text after this merges (backup, read-back), untracked.

SOTA sources

• The decision this text states: docs/decisions/2026-09-29-max-default-effort.md and receipt claude-max-default-effort-20260929 (Launcher: interactive terminal launches start at effort max (--effort max only when nothing chose an effort); receipt and decision record #483): its sources are the Claude Code CHANGELOG 2.1.284 ("Changed Ultracode into its own toggle in /effort: it no longer forces xhigh effort and stays on at any effort level"), the settings reference (ultracode, effortLevel), model-config (max is not accepted as a saved level; --effort sets a level for one session) and env-vars (CLAUDE_CODE_EFFORT_LEVEL takes precedence over --effort, /effort and the settings), all read 2026-09-29.
• The vendor and third-party effort evidence the caveat rests on is that record's "What the vendor and third-party measurements say about max" section (the Sonnet 5.5 system card section 8.4 FrontierCode figures are also in the dispatch record).
• The new receipt follows the probe method of this repository's earlier effort receipts (claude-model-effort-probes-20260929), on the same client.

Evidence-class table

Local commands run

$ python3 scripts/validate.py
exit 0  {"components": 69, "hashed_files": 7994, "profiles": 4, "receipts": 168, "status": "passed"}
$ python3 -m unittest tests.test_install_claude_profile tests.test_effort_default_guard tests.test_adoption_docs_consistency tests.test_child_usage_suite tests.test_foundation_catalog tests.test_apply_claude_settings tests.test_landscape_sweep_harness
exit 0  Ran 294 tests, OK (skipped=7)
$ node examples/claude-native/workflows/test-child-usage.mjs     SUMMARY passed=101 failed=0
$ node examples/claude-native/workflows/test-envelope.mjs        SUMMARY passed=254 failed=0
$ node examples/claude-native/workflows/test-contract-mutations.mjs   SUMMARY passed=74 failed=0
$ sha256sum --check --strict SHA256SUMS   (examples/claude-native/workflows)  exit 0
$ python3 scripts/component_matrix.py --check ; python3 scripts/new_host_grand_list.py --check ; python3 scripts/host_receipts.py validate   exit 0 each

The portable instruction file is 1,415 words (the test's ceiling is 1,440).

Decision record

No new record: the decision is docs/decisions/2026-09-29-max-default-effort.md (#483). This PR adds a dated addendum to docs/decisions/2026-09-29-sonnet-5-5-dispatch.md and closes the overturn condition that docs/decisions/2026-09-23-max-effort-default.md's addendum had left to the user. Reviews: the text was written from that record, its receipt and the installed launcher (1,810 bytes, read here). A read-only GPT-6 Astra review at max of this diff returned two medium findings and no high one: the rule for a child that names no effort was too narrow (it left out the effort the session was given explicitly, which the max-default receipt's cases F5 to F7 show a headless session passes to an unnamed subagent and to a Workflow stage that names none), and a limitation about headless --effort in the README was stale. Both were repaired in one commit (Follow-up repair), and by the repository's bounded-review rule the repair was not reviewed again.

Host evidence

Not applicable: no files under evidence/hosts/ change.

Checklist

• No new or changed GitHub Actions.
• No new or changed workflows.
• No secrets are printed, logged or committed; no new required secret was added; the receipt passes the private-content scan.
• No new paid hosting, subscription or billing surface was introduced.
• Peer-owned untracked files and worktrees were preserved.

🤖 Generated with Claude Code