Launcher: interactive terminal launches start at effort max (--effort max only when nothing chose an effort); receipt and decision record - #483

Scope

• What this PR changes, in one or two sentences: the claude launcher that install_native writes (both bootstrap scripts, byte-identical) now starts an interactive terminal launch at effort max by adding --effort max only when nothing else chose an effort (terminal on stdin and stdout, no -p/--print or --effort, CLAUDE_CODE_EFFORT_LEVEL unset, client 2.1.284 or newer), with its tests, docs, the decision record and a native receipt. The same launcher was applied to this host by hand with a backup and a read-back.
• Base commit: ba31dcd0 (main when rebased on 2026-09-29; the manifest commit was redone on main's copy; Skills path parity: install_skills, skills_status and skill_usage join global paths as skills 1.7.0 does #482 landed afterwards and GitHub reports the merge as clean)
• Lane: lane:foundation
• Owned paths touched: adoption/bootstrap-linux.sh, adoption/bootstrap-macos.sh, adoption/bootstrap.md, adoption/platforms/macos-arm64.md, tests/test_adoption_bootstrap.py, docs/decisions/2026-09-29-max-default-effort.md, evidence/receipts/claude-max-default-effort-20260929.json, and the last commit's manifests/evidence.json. Not touched: AGENTS.md, the portable and host CLAUDE.md, the recipes, the settings files and the effort guard (the session that owns the Sonnet 5.5 / Opus 5.5 dispatch text, 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, amends those in one lane:shared follow-up that cites this record; the record says how the guard treats a session the launcher starts at max).

SOTA sources

• Claude Code 2.1.284, the latest release (anthropics/claude-code v2.1.284, 2026-09-28; main at dec92bc8 when read on 2026-09-29): 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 docs pages settings reference, ultracode, model configuration, Adjust effort level (max is not accepted as a saved level; applies to the current session only) and environment variables, CLAUDE_CODE_EFFORT_LEVEL (takes precedence over --effort, /effort and the saved levels), fetched 2026-09-29; the installed 2.1.284 binary's settings schema and Ultracode gate, read with bounded windows (receipt binary_schema).
• Effort guidance and measurements: Anthropic's effort page (Sonnet 5.5 and Opus 5.5 sections), the Claude Sonnet 5.5 and Claude Opus 5.5 system cards (PDF sha256 89cfd6a6... and 7311c9c6..., sections 8.4, 8.5, 6.5.1), Cognition's FrontierCode v1.1 data file, the Artificial Analysis Terminal-Bench 4.0 page and Cursor's CursorBench page, all read 2026-09-29 and re-checked by an independent verifier (30 of 40 claims confirmed, 10 corrected as written in the record), then fact-checked once more by a read-only GPT-6 run (nine wording corrections, each re-fetched from Cognition's data.json, Anthropic's launch page footnote 2 and Zapier's leaderboard 1.0.6, all applied).
• The macOS pty behaviour behind the test-helper fix: apple-oss-distributions/xnu bsd/kern/tty_dev.c (ptsclose, last commit f6217f89) calls ttyclose in bsd/kern/tty.c (last commit 8d741a5d), whose ttyflush(tp, FREAD | FWRITE) empties the terminal's queues at the last slave close; read 2026-09-29.
• Prior art for a flag-adding wrapper: Ramsbaby/jarvis infra/bin/claude-xhigh.sh (blob 66e75106, repository pushed 2026-09-28, not archived).
• The generator this extends: this repository's own install_native (adoption/bootstrap-linux.sh, identical to the macOS copy, enforced by test_both_scripts_share_one_install_native).

Evidence-class table

Local commands run

$ TMPDIR=/var/tmp PYTHONDONTWRITEBYTECODE=1 python3 -m unittest tests.test_adoption_bootstrap tests.test_adoption_bootstrap_macos tests.test_adoption_docs_consistency tests.test_adoption_contract
Ran 290 tests in 74.681s
OK (skipped=33)
$ python3 scripts/validate.py
{"components": 69, "hashed_files": 7993, "profiles": 4, "receipts": 167, "status": "passed"}

Decision record

docs/decisions/2026-09-29-max-default-effort.md: the selection, six alternatives with why each was rejected, coverage limits, overturn conditions and limitations. It says, plainly, that the default rests on the user's requirement and not on a measured gain on this repository's work: for Sonnet 5.5 the direction of the effect differs between Terminal-Bench in --bare mode (max wins) and FrontierCode in the full client (max loses at about 12.6 times the tokens on Main; these results do not isolate the effect of skills or subagents), and for Opus 5.5 the gains vary by benchmark and come at higher cost, with their value for this repository unmeasured. claude --effort xhigh is the one-word opt-out.

Cross-family reviews: (1) one read-only GPT-6 round on the launcher (gpt-6-astra, effort max, live search), five findings, all handled in one repair round (short-flag clusters, --, a client-version floor, a leaking pty helper, a lossy argument recorder); the reviewer's suggestion to parse value-taking options was not adopted (the launcher cannot know the client's option table), and the limit is documented and tested in its safe direction. (2) A second read-only GPT-6 run fact-checked the record's vendor section: nine wording corrections, all applied after re-fetching the figures; one suggestion (naming Cognition's tokens field as output tokens) could not be confirmed from the fetched sources and was not adopted.

Host evidence

Not applicable: no file under evidence/hosts/ changes. The host launcher was replaced by hand (announced to the live sessions first): before 165 bytes, after 1,810 bytes, hashes in the receipt.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment (no floating tags): no workflow changes.
• New/changed workflows declare top-level permissions: contents: read: no workflow changes.
• 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