Make standing Ultracode opt-in and record task dispatch scope

[seathatflowsinourveins]

Scope

Shared project settings currently enable Ultracode for every checkout, overriding a user's lower-precedence choice. The portable template also overwrites an existing false with true at the next apply. Remove the standing key from both inputs, preserve explicit task opt-in and generic merge behavior, and clarify the adoption directions that would copy the opt-in profile into persistent settings.

• Base commit: e712bae606e8d27f010af8e3c13ded6b7a7c207a.
• Lane: lane:foundation.
• Owned paths: the two shared JSON inputs, two settings tests, adoption/recipe directions, the workflow README, three dated decision amendments, one displaced catalog citation, and the evidence registry.
• The September 23 record actually decided the standing switch. The September 29 effort record gains a pointer amendment; effort, models, workflow availability, environment and size guidance retain their values. The October 2 record gains a dated projection read-back for the omitted settings piece; its old counts remain.

Landing path and overlaps

Draft for an independent GPT read, the command center's ACK, then the owner's own direction on the settings change before 5f's landing cue. Requested landing path follows #803, which has landed at b593372; #802 and #813 keep their existing landing turns. The draft and checks are authorized; landing remains held for that owner direction and the relayed 5f cue. One rebase at the landing turn, with main's registry and this PR's registrations last.

#823 adds claudeMdExcludes near the removed project key. #813 changes the same README. Each later lander rebases once on its cue. This PR does not push or rebase either records PR. Instruction files, host/user settings and vendored workflow scripts are outside this change.

SOTA sources

• Installed Claude Code 2.1.292, native claude --version and claude --help (--settings <file-or-json>, --setting-sources <sources>). Public executable SHA256 a967e7b1d8b4e47ee421d5433027880347952b0c0857abf880e2c942a4ec93b3: optional-boolean ultracode schema at byte 204350508; independent keyword-trigger default at bytes 224707120–224707650. This is installed-client source review, not a new workflow execution.
• Anthropic settings reference, Ultracode and keyword trigger and settings scope precedence, reviewed 2026-10-07. CLI and project settings outrank user settings; omission leaves explicit user choices intact.
• Anthropic workflow invocation and Ultracode, especially lines 115–162 and 416–447 in the retrieved Markdown snapshot. Session /effort ultracode on|off, interactive keyword and launch --settings are distinct opt-ins. Ultracode suppresses the large-workflow warning and session concurrency limit; a finite brief is required rather than treating the size setting as an enforced cap.
• Anthropic subagents and experimental agent teams, reviewed 2026-10-07, for choosing a bounded dispatch mode instead of a standing workflow switch.
• Anthropic Claude Code changelog at v2.1.292, 2.1.284 effort behavior, line 821 in the reviewed snapshot: Ultracode and effort remain independent.
• Existing repository merge implementation: tools/adoption/apply_claude_settings.py:297-340,402-443 at the base above. Incoming scalars win; omitted fields and unrelated user choices are preserved. Existing repository decision addendum convention and scripts/host_receipts.py registration govern the append-only records and final registry commit.

Writer and pin audit

Locators in this inventory refer to the base commit above. Complete literal and case-insensitive scans preceded edits: 222 matching files, including historical evidence. The writer denominator is nine distinct setting inputs, generic write paths and adoption directions. Four changed: two standing JSON inputs and two persistent-merge directions. The four generic paths consume the corrected template unchanged; the explicit launch input remains true.

Writer/input	Disposition
.claude/settings.json:4	Remove only the standing key.
adoption/templates/claude.settings.template.json:360	Remove only the standing key; fresh apply omits it and existing false or true survives.
examples/claude-native/ultracode.settings.json:4	Retain true as the explicitly selected per-task launch profile.
tools/adoption/apply_claude_settings.py:402-433 (default template :38,438)	Retain generic atomic merge; corrected default input prevents re-enabling false. Test omission, preserved choices, idempotence and deliberate opt-in.
tools/adoption/render_config.py:255-270 (template :68)	Retain; rendered default settings inherit omission.
tools/adoption/new_wsl_client_config.py:2188-2237 (source :131)	Retain; shared-template-derived pieces omit the setting. Read back the resulting piece count in a dated amendment.
adoption/bootstrap-linux.sh:1382-1387	Retain full-profile wrapper; it applies the corrected rendered template.
recipes/claude-native-ultracode.md:58-59	Persistent merge excludes Ultracode; task launches select the explicit profile.
examples/claude-native/workflows/README.md:47-48	Append adjacent dated clarification of explicit profile versus persistent baseline, and task opt-in under “Opt-in and limits.”

Other pin/carrier	Disposition
catalogs/foundation/upstream-surface-dispositions.json:2267-2276	Open follow-up for currency/CC: the standing-adoption row still describes the October 4 user setting. Separate standing-off from per-task availability; no row/schema change here.
recipes/claude-native-ultracode.md:181	Correct current starting-choice wording; retain the observed qualification cell.
recipes/claude-native-ultracode.md:109-117	Retain the dated September 23 standing-default paragraph and append a local October 7 supersession linked to the dispatch amendment. This ninth pin was missed in the initial inventory and found by the exact-head read.
catalogs/foundation/decisions.json:2662	Open follow-up for its catalog owner/CC: qualify current activation text that assumes the coordinator is under Ultracode; preserve measured evidence.
adoption/bootstrap.md:548-555	Remove Ultracode from the persistent baseline property list and describe task-scoped selection.
examples/claude-native/workflows/contract.config.json:2	Retain explicit opt-in contract input; writes no live setting.
tools/token-e2e/freeze_snapshot.py:133	Retain boolean observation/drift schema; writes no default.
catalogs/foundation/upstream-surface-baseline.json:280	Retain supported upstream capability name.
adoption/new-wsl/client-config-map.json:299	Retain capability match, which supplies no boolean and creates no piece without an input key.
recipes/sota-convergence-practice.md:36	Open follow-up for its recipe owner/CC: its conditional “with Ultracode on” reference still points to the project file; redirect to explicit opt-in. It does not write a setting.

scripts/build_ecosystem.py:117-125,820,985-987 indexes paths and renders guide content; it contains no independent true-setting writer. It is unchanged; the supported --write refresh was run. tools/adoption/scaffold_repo.py does not emit a Claude settings file (the six-output contract is tested).

Test and historical-record dispositions

The nine test files in the brief are not nine defaults. The complete scan also found the resolver metadata fixture and four generic-writer modules:

Test module	Disposition
test_install_claude_profile.py	Add omission guard for project/template and retain explicit opt-in; preserve effort/model assertions.
test_apply_claude_settings.py	Add actual apply regressions for fresh omission, existing false/true with unrelated settings, idempotence and explicit opt-in.
test_render_config.py, test_new_wsl_client_config.py	Retain generic render/apply contracts; run against corrected template.
test_effort_default_guard.py	Retain true fixtures that test selected-task compatibility and effort behavior.
test_adoption_bootstrap.py	Retain explicit effort selection and historical RTK probe context.
test_adoption_status.py	Retain historical E1 tools/profile relationship.
test_freeze_snapshot.py	Retain true/false observation and drift fixtures.
test_runtime_worker_openhands.py, test_runtime_worker_openhands_resolver.py	Retain source=ultracode SDK serialization metadata fixtures; these are not settings.
test_token_e2e_receipt_checks.py, test_token_e2e_grader.py, test_token_e2e_preregistration.py	Retain immutable receipt, grading and source-resolution fixtures.
test_scaffold_repo.py	Retain positive six-file output enumeration.

The brief's eight-record count was a lead. Fourteen decision files contain lowercase ultracode; ten more contain a case-insensitive mention. All original text/observations stay intact:

Decision file under docs/decisions/	Disposition
2026-09-23-max-effort-default.md	Append the dated task-selection amendment to the actual standing decision.
2026-09-29-max-default-effort.md	Append a pointer amendment; preserve the relocated frozen rule and effort history.
2026-10-02-new-wsl-client-configuration.md	Append the current projection count; preserve old read-back.
2026-09-25-model-fallback-guard.md	Retain historical recipe scope/reference.
2026-09-26-harness-rules-cleanup.md	Retain installed-source observation.
2026-09-26-stack-agents-role-dispatch.md	Retain historical dispatch reference.
2026-09-27-claude-harness-settings.md	Retain settings/source context.
2026-09-27-model-currency.md	Retain model/template context.
2026-09-27-token-lanes-subagent-start.md	Retain historical recipe reference.
2026-09-28-community-sweep.md	Retain sweep/candidate observations.
2026-09-28-delegated-decisions.md	Retain source/reference.
2026-09-29-sonnet-5-5-dispatch.md	Retain model/effort and dispatch observations.
2026-09-30-task-model-routing.md	Retain research/source reference.
2026-10-04-new-wsl-token-layer-default.md	Retain historical recipe reference.
2026-10-04-token-full-stack-owner-default.md	Retain historical recipe reference.
2026-09-23-claude-user-profile.md, 2026-09-25-codex-mcp-scope.md, 2026-09-25-skills-trial-and-usage.md, 2026-09-26-telemetry-writer-identity.md, 2026-09-28-ecosystem-roadmap.md	Retain historical references and observations.
2026-09-30-rule-text-every-layer.md, 2026-09-30-sol-primary-quality-defaults.md, 2026-10-02-mac-claude-resolution.md, 2026-10-03-compaction-window-ab-retirement.md, 2026-10-05-harness-context-budget.md	Retain historical references and observations.

The README's insertions preserve all original text. All eleven frozen workflow fixtures and their exact passages remain unchanged. One current catalog citation is adjusted to the sentence's new line; historical source locators are retained as recorded.

Evidence-class table

Claim	Evidence class	Command / receipt
Optional standing setting, separate opt-in mechanisms and scope precedence	source_review	Installed 2.1.292 schema/help plus the official documents above.
Default inputs omit the key without changing other settings; apply preserves user choices	local_integration	Real file apply tests and parsed before/after comparison.
One read-only five-point task, one session per client on 2026-10-07	native_proven	CC/co-op sanitized revision-2 extract SHA256 7f5d26c423134ee3c11c4265375d5bdaf1c49280639f4bb7ebc231585c04190d; original/raw records are not committed.
Codex Sol max input 1,390,078, including cached subset 1,264,640; output 18,609; runner wall time 356 s	native_proven	Client usage block and runner start/end stamps, runner-log SHA256 e41bffad6830f3f25ef922a4140009bcbedd6b03c8bd26ebcc731141a5bb31b9. Cached input is not added again.
Claude standing-on session launched one workflow, 45 turns, about 536 s; cache-read Opus 5.5 25,503,839 and Sonnet 5.5 17,108,897; Fable 5.1 advisor input 297,092/output 34,949	native_proven	Client result block in the same sanitized extract. Advisor estimate 4.72 USD is included in total 26.38 USD.
Cost and comparison limits	source_review	Total is a client estimate, not billing. Different counter scopes and one run each establish no token/cost ratio or quality-equivalent head-to-head.

Local commands run

All commands run under nice -n 19 ionice -c3; durable private logs retain the original output. No live provider run was added.

python3 -m unittest tests.test_install_claude_profile tests.test_apply_claude_settings
rc 0 — 164 tests, 1 skip

python3 -m unittest tests.test_upstream_surface_watch
rc 0 — 114 tests, 6 skips; displaced README citation corrected

python3 -m unittest tests.test_install_claude_profile tests.test_apply_claude_settings tests.test_upstream_surface_watch
rc 0 after J829 amendment — 278 tests, 7 skips

python3 scripts/validate.py
rc 0 — 69 components, 10419 hashed files, 4 profiles, 224 receipts

sha256sum --check --strict SHA256SUMS
node test-envelope.mjs
node test-contract-mutations.mjs
all rc 0 in examples/claude-native/workflows

python3 tools/adoption/new_wsl_client_config.py --check
rc 0 — 397 pieces, 357 wired, 24 not wired, 16 authorization-gated

python3 scripts/build_ecosystem.py --write
rc 0 — refreshed ignored offline guide; generator unchanged

The first full unittest attempt used an interpreter without CI's calendar dependencies. It was stopped with rc 130 after 8678 completed events; that partial run is not a pass. The corrected complete native run at 695d344 returned rc 1: 11501 tests, 15 failures, 1087 skips, 1742.178 s, with CI's hash-locked .github/requirements-ci.txt in an isolated suite venv, separate promotion-gate venv, pinned tree-sitter-bash 0.25.1 and yaml 2.9.0. Invocation matches CI's PYTHONFAULTHANDLER=1 timeout --signal=ABRT --kill-after=60s 55m python3 -m unittest -v --durations 50, with private interpreter paths. Native Python is 3.13.16 and Node 24.21.0; CI uses Python 3.12.3 and Node 22.23.2. This returned native result is a baseline comparison with the environment limitations below; hosted landing CI supplies acceptance of record.

Verification classes (A8, 2026-10-07)

1. Checks covering this change: validate rc 0; amended apply/install modules 164 tests/1 skip rc 0; citation module 114/6 skips rc 0; workflow checksums/contracts rc 0. Independent source critics verified the nine writer dispositions, preserved historical prefixes and frozen fixtures, and corrected the displaced current citation.

2. Acceptance of record: hosted validate at the eventual landing head. At initial head 695d344abc61aadda16bff54a3e73a576da1e56a, CodeQL passed; validate did not run because the conflicting PR had no merge ref. Per the co-op's J829 ruling, validate follows the single landing rebase. The exact-head delta read and CC ACK remain conditional on that landing CI and the owner's own settings-change direction.

   CI at 44f77fd: CodeQL green; validate not run (DIRTY, no merge ref)

3. NativeStack2604 baseline comparison only: the complete local suite returned the failure result above, not a passing result or a gate for opening this draft. native-baseline-failures.json, SHA256 ce14378c52469415de2479eb4b27c680a72d2f62e89aff49da7543888e794eca, retains exact native micro-check failures and positive byte comparisons against unmodified base e712bae and observed main b593372. The co-op checked that the affected files are unchanged and routes the findings separately. The remaining full-run failures are retained and classified separately; they are not silently excluded or all attributed to that four-case receipt.

Reproduced native baseline test IDs:

tests.test_adoption_bootstrap.SystemPackagesBeforePrerequisiteCheckTests.test_skip_system_packages_with_missing_tool_lists_it_and_exits_4
tests.test_adoption_launchd.LaunchdAgentsScriptBehaviorTests.test_lint_falls_back_to_plistlib_when_plutil_is_absent
tests.test_adoption_launchd.LaunchdAgentsScriptBehaviorTests.test_lint_prefers_plutil_when_present
tests.test_adoption_guarded_runners.BoundedRunContainmentTests.test_containment_or_refusal

Three restricted-PATH fixtures omit dirname; the guarded-runner case assumes a positional order for systemd property values and receives running first. The adaptive-paper mutation helper additionally selects a deficient existing host-pinned interpreter rather than the isolated suite interpreter, with no supported override; its five mutation self-test failures remain visible in the full run. CI's fresh runner uses the provisioned interpreter fallback. The co-op routes the shell/property-order findings to the foundation defect owner and the pinned runtime gap to the trading owner through the command center. No unrelated test, host runtime or configuration is changed to bypass these failures. The complete returned result and exact-head CI will be reported before a landing request.

A separate runner correction applies to two additional Codex role-directory failures: setting umask 077 for the local suite's private logs also changed its tests' directory creation modes. Same-head paired controls pass both tests under the ordinary umask 022 (rc 0) and reproduce both failures under 077 (rc 1). These are invocation differences, not inherited defects or PR regressions. The original full output is retained with that limitation; hosted CI remains the acceptance of record.

The completed classification receipt full-baseline-result.json has SHA256 f6c8f557812719fdc6a3620af7a01b21f3ac6c8aeffca92d78ffae178d1ae262. Its fifteen failures are: four restricted-fixture/property-order cases, five pinned-runtime mutation cases, three runner-umask cases (including the dashboard control), one venv user-site skip/mutant-expectation mismatch, one forced local-clone hardlink failure across filesystems, and one installed Claude 2.1.292 notification name (auth_storage_failure) missing from the decision inventory. Related source bytes match the base and observed main; the notification's policy/introduction release is unassigned. Controls/source locators and original failed output remain separate from the change-covering checks and hosted landing acceptance.

Contract changes

contract-changes=none. This PR adds assertions for the default omission and explicit opt-in without reversing or narrowing an existing main assertion. Main's unedited versions of both modified main-carried test files pass on the prepared composition; they will be checked again against the relayed landing base. Any undeclared main-version failure blocks landing.

Decision record

Append-only Amendment (2026-10-07) in docs/decisions/2026-09-23-max-effort-default.md, with alternatives and the upstream-evidence comparison that would overturn the choice. The September 29 pointer amendment directs current readers to it without rewriting decided text. Current new-WSL projection is 397 pieces: 357 wired, 24 not wired, 16 authorization-gated; the earlier recorded 398 remains.

Host evidence

No files under evidence/hosts/ change. This PR records attributed measurements and source review, not a fresh host acceptance or provider run.

Checklist

• No GitHub Actions changes; existing pinned workflow is used.
• No secrets printed, logged or committed; no new required secret.
• No new paid hosting, subscription or billing surface.
• Peer-owned files and worktrees preserved.
• Historical observations and decided text retained; amendments are dated additions.
• Validate, amended modules and workflow contracts pass; final registry-only commit last.
• Complete suite/remote CI and inherited native-blocker disposition confirmed before landing.