seathatflowsinourveins · Oct 7, 2026 · Oct 7, 2026 · Oct 7, 2026
@@ -1,7 +1,6 @@
 {
   "$schema": "https://json.schemastore.org/claude-code-settings.json",
   "enableWorkflows": true,
-  "ultracode": true,
   "effortLevel": "xhigh",
   "workflowSizeGuideline": "unrestricted",
   "switchModelsOnFlag": false,

@@ -545,11 +545,13 @@ GitHub-hosted macOS runner; see
    [jCodeMunch recipe](../recipes/README.md#focused-jcodemunch-retrieval)
    covers indexing and the native statistics.
 
-   Then apply the settings template itself (model, effort, ultracode,
+   Then apply the settings template itself (model, effort,
    workflow env, hooks, and the credential deny rules plus the `PreToolUse`
    secret-guard hook from [`docs/secret-storage.md`](../docs/secret-storage.md#user-level-guards-deployed-by-the-claude-profile)) into the live `~/.claude/settings.json` with
    [`tools/adoption/apply_claude_settings.py`](../tools/adoption/apply_claude_settings.py),
    after rendering it for this host with step 4's `render_config.py --out`:
+   The baseline omits the standing Ultracode key; task/launch opt-in uses the
+   [native workflow profile](../recipes/claude-native-ultracode.md) separately.
    ```sh
    python3 tools/adoption/apply_claude_settings.py --template "$RUN_DIR/rendered/settings.json"
    python3 tools/adoption/apply_claude_settings.py --template "$RUN_DIR/rendered/settings.json" --dry-run

@@ -357,7 +357,6 @@
   "skipDangerousModePermissionPrompt": true,
   "theme": "dark",
   "verbose": true,
-  "ultracode": true,
   "enableWorkflows": true,
   "showThinkingSummaries": true,
   "cleanupPeriodDays": 3650,

@@ -562,8 +562,8 @@
   {
    "key": "claude:env:CLAUDE_CODE_WORKFLOW_PREFIX_STAGGER_MS",
    "disposition": "declined",
-   "reason": "off on both hosts (read 2026-10-04)); the repository record at examples/claude-native/workflows/README.md:1083 names it",
-   "source": "examples/claude-native/workflows/README.md:1083",
+   "reason": "off on both hosts (read 2026-10-04)); the repository record at examples/claude-native/workflows/README.md:1089 names it",
+   "source": "examples/claude-native/workflows/README.md:1089",
    "carrier": null,
    "value": null,
    "scope": null,

@@ -297,3 +297,82 @@ Sonnet 5.5. The 2.1.281 findings above stay as history for that release. On 2.1.
   Opus level.
 - **The `child-usage.mjs` mirror statement under "Vendored lane" is history.** The file here has since grown well beyond agent-lab
   `b31f640`, and a local agent-lab checkout holds a different, smaller copy.
+
+## Amendment (2026-10-07): select orchestration per task
+
+This amendment supersedes the standing-Ultracode choice in decision item 1.
+It leaves the recorded probes and the effort/model rules unchanged.
+
+**Decision.** The repository and portable user-settings template omit the
+standing `ultracode` key. The coordinator chooses a dispatch mode for the task:
+direct execution or a deterministic script for suitable work, normally a
+bounded GPT lane or SDK worker for substantial research and builds, and a
+subagent, background agent, isolated builder, workflow or team when its
+coordination contract calls for it. Workflow availability stays enabled; this
+change selects when to use orchestration rather than changing effort.
+
+Claude Code 2.1.292 starts Ultracode off when the key is unset or false. A true
+value selects standing orchestration, subject to workflow/model availability.
+A shared project setting outranks user settings, which is why a committed true
+value could restore the standing choice in other checkouts. Removing the
+portable template's value also stops a subsequent profile apply from replacing
+an existing user false with true; native merge semantics preserve explicit user
+choices and unrelated settings.
+
+**Task opt-in.** A direct interactive prompt can request a workflow or use the
+native keyword. The session toggle is `/effort ultracode on|off`; interactive
+toggles do not persist. A launch can pass
+`claude --settings '{"ultracode":true}'` or the retained explicit opt-in example.
+Keep model and effort choices separate. Before opting in, specify finite stage
+and agent counts, concurrency, output limits, allowed effects, and verification.
+The size guideline is advisory, and Ultracode suppresses native large-workflow
+warnings and the session subagent concurrency limit; an unrestricted guideline
+does not supply a task budget. Existing project effort, workflow-size and
+environment settings are unchanged.
+
+**Observed motivation, not a controlled comparison.** On October 7, the
+command center recorded one task requesting five pieces of read-only evidence
+without naming a tool, once in each headless client. The Claude run with the
+standing switch on reported a workflow launch. The following are different
+counter scopes, not a token or cost ratio:
+
+| Observation | Scope | Reported value |
+| --- | --- | --- |
+| Codex Sol at max | Client session usage | 1,390,078 input tokens, including 1,264,640 cached input as a subset; 18,609 output tokens |
+| Codex execution | Runner start/end stamps for the call | 356 seconds |
+| Claude Opus 5.5 | Client model usage, cache-read input | 25,503,839 tokens |
+| Claude Sonnet 5.5 | Client model usage, cache-read input | 17,108,897 tokens |
+| Claude Fable 5.1 advisor | Client model usage | 297,092 input and 34,949 output tokens; 4.72 USD within the session estimate |
+| Claude session | Client result | 45 turns, approximately 536 seconds, 26.38 USD estimated across all three models |
+
+Cached input is not added to Codex input. The Claude figure is the client's
+estimate, not provider billing; advisor cost is included, not an additional
+charge to sum. One run per client establishes neither a general dispatch
+advantage nor equivalent task quality. No run was repeated for this amendment.
+The sanitized revision-2 extract has SHA256
+`7f5d26c423134ee3c11c4265375d5bdaf1c49280639f4bb7ebc231585c04190d`.
+The runner log's SHA256 is
+`e41bffad6830f3f25ef922a4140009bcbedd6b03c8bd26ebcc731141a5bb31b9`;
+its start/end stamps are 05:11:57Z and 05:17:53Z on October 7. No raw client
+record, session identifier or host path is included here.
+
+**Alternative and revisit condition.** Keeping standing orchestration would
+continue letting checkout settings select it for every eligible task. Keeping
+it as an explicit task choice preserves the native feature while avoiding that
+automatic dispatch. Revisit the choice when a changed native mode, a named
+workload or better upstream evidence establishes a different appropriate
+default; keep dispatch, information quality and complete usage scopes distinct.
+
+**Sources.** Installed Claude Code 2.1.292's optional Boolean schema is at byte
+204350508 of the Linux executable, SHA256
+`a967e7b1d8b4e47ee421d5433027880347952b0c0857abf880e2c942a4ec93b3`;
+`claude --help` exposes `--settings`. The native schema notes that interactive
+toggles do not persist. Primary documentation read October 7:
+[setting/default](https://code.claude.com/docs/en/settings-reference.md#L1362-L1379),
+[scope precedence](https://code.claude.com/docs/en/settings.md#L631-L635),
+[workflow opt-ins and effects](https://code.claude.com/docs/en/workflows.md#L115-L162),
+[advisory sizing](https://code.claude.com/docs/en/workflows.md#L416-L447),
+[subagents](https://code.claude.com/docs/en/sub-agents.md),
+and [experimental teams](https://code.claude.com/docs/en/agent-teams.md).
+The [2.1.284 changelog entry at the installed release](https://github.com/anthropics/claude-code/blob/v2.1.292/CHANGELOG.md#L821)
+keeps the toggle independent of effort.
@@ -254,3 +254,11 @@ Revisit this record when any of these happens:
 ## Repository rule relocated verbatim (2026-10-05)
 
 - This repository commits `.claude/settings.json` with Ultracode on and `effortLevel: xhigh`, the saved fallback for any model. A terminal session started through the ecosystem `claude` launcher runs the coordinator at `max` (the launcher adds `--effort max` only when nothing chose an effort and the client is 2.1.284 or newer; `claude --effort xhigh` opts out). On Claude Code 2.1.284 Ultracode stays on at any effort level and the `ultracode` setting sets none, so a `max` session keeps its workflow orchestration on; the `max` default rests on the user's requirement, not on a measured gain here. Headless `-p` runs pass `--effort` per call site, and `CLAUDE_CODE_EFFORT_LEVEL` stays unset at every scope (any value overrides every child's effort). Pass `effort: 'max'` with an explicit task-matched `model` on every ad-hoc workflow `agent()` call: a stage that names no effort runs at its agent's frontmatter effort, else at the effort the session was given explicitly (`--effort`, `/effort`, the model picker), else at its model's saved level or default, and one that names no model takes its definition's model, else `CLAUDE_CODE_SUBAGENT_MODEL` (`opus`), else the lead's. `opus` takes judgment; `sonnet` (Sonnet 5.5) takes fan-out units that an executable oracle or a later Opus stage checks (`examples/claude-native/workflows/README.md`, "Sonnet 5.5 fan-out units"). Probes and overturn conditions: `docs/decisions/2026-09-29-max-default-effort.md`, `docs/decisions/2026-09-29-sonnet-5-5-dispatch.md` and `docs/decisions/2026-09-23-max-effort-default.md`.
+
+## Amendment (2026-10-07): orchestration is selected per task
+
+The [September 23 record's dated amendment](2026-09-23-max-effort-default.md#amendment-2026-10-07-select-orchestration-per-task)
+supersedes the standing-Ultracode setting retained above. The committed baseline
+and portable settings template omit that key; explicit task/session/launch
+opt-ins remain available. This amendment changes no model or effort rule, and
+does not alter the original measurements or relocated passage contract.
@@ -1107,3 +1107,16 @@ The [local-time decision](2026-10-05-user-facing-local-time.md) adds one sentenc
 
 ```text
 ```
+
+## Amendment (2026-10-07): omit standing Ultracode from the baseline
+
+The [task-dispatch amendment](2026-09-23-max-effort-default.md#amendment-2026-10-07-select-orchestration-per-task)
+removes the optional standing `ultracode` input from the portable settings
+template. Its native capability-map entry remains valid for an explicit opt-in;
+the baseline no longer renders that practice setting.
+
+Today: 397 pieces, 357 wired (206 practice, 151 through a slot), 24 not wired
+(0 through a slot that does not install, 24 by their own entry) and 16 authorization
+pieces. This is the current repository-input projection from
+`python3 tools/adoption/new_wsl_client_config.py --check`, not a new host
+installation or client acceptance run. The prior dated counts remain unchanged.
@@ -52,6 +52,12 @@ the Codex side.
    section) and, when the project keeps usage receipts, `usage_receipts_dir`,
    `routing_doc` and `task_record`. A configured path that does not exist fails
    the suites; nothing is skipped silently.
+   **Amendment (2026-10-07).** Use the retained settings example as an explicit
+   task/launch profile, including for the contract's `settings` path. Persistent
+   baseline settings omit `ultracode`; do not merge that opt-in key into them.
+   Follow the [dated dispatch amendment](../../../docs/decisions/2026-09-23-max-effort-default.md#amendment-2026-10-07-select-orchestration-per-task)
+   when choosing the mode. Other project settings and permission requirements
+   retain their own scope.
 3. Give the session a permission source that covers the Workflow tool. A headless
    `claude -p` run of a saved workflow is refused ("Review dynamic workflow before
    running", observed in the 2026-09-21 qualification, see
@@ -1086,6 +1092,18 @@ Applies to the coordinator and every Workflow/Agent child (project agents and `.
 - **Usage accounting.** Count each client separately: native `/usage`, `claude agents`/`/workflows` journals, `ccusage` offline reports and `ecosystem-token-report refresh`. Never sum RTK, Context Mode, jCodeMunch, Headroom and provider counters, and never state a savings percentage from a fixture. After a Workflow run, `node .claude/workflows/child-usage.mjs <Transcript dir printed by the Workflow tool>` (or `--latest`) returns each child's requested and resolved model, effort, provider-returned usage and first-prompt size, and exits 1 when a child is null, substituted, named no model or ran an older model than its alias documents (the client resolves a family alias to the lead's exact model when the lead belongs to that family, so a `sonnet` child under a lead pinned to an older Sonnet is flagged too; keep the lead on the alias). An attempt that returned nothing and that the runtime re-ran under the same journal key (a Workflow pauses at a usage limit and re-runs its waiting agents after the reset) is listed under `superseded_attempts`, not as a lost child, and its usage still counts in `by_resolved_model`. Usage such an attempt holds that cannot be counted (an assistant message without provider usage or without a resolved model, or no transcript at all) is its `usage_issues`, and it leaves the run incomplete. Advisor usage lies outside these totals (2026-09-28): subagents inherit the configured advisor ([advisor](https://code.claude.com/docs/en/advisor)), the API reports each advisor call as a `usage.iterations[]` entry of type `advisor_message` and keeps top-level usage executor-only ([advisor tool usage](https://platform.claude.com/docs/en/agents-and-tools/tool-use/advisor-tool#usage-and-billing)), and a run-mode child's `usage` and `by_resolved_model` read only the top-level counters (since 2026-09-29 `lanes.measurement.usage` also reports the advisor iterations, as `advisor_iterations` and `totals_including_advisor`). A count-only scan of one host found 1,018 advisor iterations in 660 of 2,912 subagent transcripts ([receipt](../../../evidence/receipts/claude-advisor-usage-scan-20260928.json)). Take complete session usage from `/usage`, which includes advisor usage, and hold the advisor state equal across comparison arms with `CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1` or `/advisor off`.
 - **Cross-family review.** Explicit `/codex:review` or `/codex:adversarial-review --background` at integration points (see `recipes/claude-codex-cooperation-lanes.md`); findings are verified against source, not accepted by agreement.
 - **Opt-in and limits.** The `ultracode` keyword starts a workflow only from a prompt typed in the session; it is inert from `-p`, an unstamped SDK prompt, a scheduled task or a relayed comment, so a headless run invokes a saved workflow by name under a settings source whose permission mode or allow rule (`Workflow` or `Workflow(<name>)`) covers the tool. Scripts take no mid-run user input, no `import()` and no `Date.now()`, `Math.random()` or argless `new Date()` (pass timestamps through `args`; run a stage that needs sign-off as its own workflow); one `parallel()`/`pipeline()` call takes at most 4,096 items and a run at most 1,000 agents.
+  **Amendment (2026-10-07).** Standing Ultracode is off in the portable baseline.
+  The coordinator selects it per task through the direct interactive keyword,
+  the nonpersistent `/effort ultracode on|off` session toggle, or launch-scoped
+  `--settings` (for example `'{"ultracode":true}'`). Workflows stay available and
+  model/effort choices stay explicit. Before opting in, bound the number of
+  stages and agents, concurrency, outputs and permitted effects, and name the
+  verification/recovery steps. Size guidance is advisory; Ultracode suppresses
+  native large-workflow warnings and the session subagent concurrency limit,
+  so a task brief must supply its own finite scope. The existing project size
+  guideline and environment values are unchanged. See the
+  [decision and observed scopes](../../../docs/decisions/2026-09-23-max-effort-default.md#amendment-2026-10-07-select-orchestration-per-task)
+  and [native opt-ins](https://code.claude.com/docs/en/workflows.md#L115-L162).
 - **Failure and replay.** On resume a failed or stopped agent runs again together with every agent started after it, completed ones included, and a run with nothing cached has nothing to resume; native failure, cancel and recovery remain documented but unobserved (open gate in `docs/native-ultracode-20260921.md`).
 - **Agent allowlists and skill preloads.** Block a specific command with a `permissions.deny` Bash rule, never with a specifier inside an agent's `disallowedTools`, which removes the whole tool. A child `skills:` entry preloads the full skill text into its first prompt. Targeted builder and security-reviewer preloads are documented configurations with their own first-prompt sizes unmeasured and preregistered alongside the researcher/verifier rows in the [decision record](../../../docs/decisions/2026-09-26-stack-agents-role-dispatch.md). The [listing-state addendum](../../../docs/decisions/2026-09-25-skills-trial-and-usage.md#addendum-2026-09-26-listing-state-and-agent-preload) probed preload viability on Claude Code 2.1.283 for a pinned, table-listed skill; the builder's plugin-scoped `context-mode:context-mode` preload instead relies on its installed `SKILL.md` carrying no invocation-disabling frontmatter. Neither is a repeat native probe of these specific configurations. The separate Read/Glob/Grep reviewer probe of 2026-09-22 measured 15,059 tokens with one skill; that figure does not qualify these new configurations.
 - **Version gates.** A rule that depends on a client version names the version observed when it was recorded. This section was checked against `claude --version` 2.1.278 on 2026-09-22; the official docs fetched that day gate the settings-file size guideline at 2.1.219, `/workflow-authoring` at 2.1.248 and the concurrency setting at 2.1.269. The excerpts those fetches returned (including the stagger default above) are retained in `evidence/artifacts/harness-rules-convergence-20260922/official-doc-excerpts.json`.