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

…he measured effort rule

The portable instructions and the workflows README state which units may run on Sonnet 5.5
(oracle- or Opus-checked fan-out), keep every judgment on Opus, and record the 2.1.284
effort and inheritance findings. The instruction word budget is re-baselined to 1,310.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

… and the default child model

Adds docs/decisions/2026-09-29-sonnet-5-5-dispatch.md, the probe receipt
claude-model-effort-probes-20260929, a dated addendum to the 2026-09-23 record and two
anti-pattern rows.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…ent version

ALIAS_RESOLUTION had rows for opus only, so a child that requested sonnet but
ran an older Sonnet than its client version documents was not flagged as
substituted. Add sonnet (Sonnet 5 from v2.1.197, Sonnet 5.5 from v2.1.284) and
fable (Fable 5.1 from v2.1.257) from the model-config doc's version history
table (fetched 2026-09-29). haiku and any alias without rows still have no
expectation.

test-child-usage.mjs: the "alias without rows" case moves from sonnet to
haiku; the opus table's sonnet clause takes the documented value
(claude-sonnet-5 on 2.1.281) and gains a haiku-is-null clause; new cases cover
the sonnet and fable version tables, a whole child on claude-sonnet-5 under
2.1.284 (older than documented), claude-sonnet-5-5 under 2.1.284 and
claude-sonnet-5 under 2.1.283 (complete), and a change from claude-sonnet-5-5
to claude-sonnet-5 within one child. 95 -> 101 cases.

SHA256SUMS lines for both files regenerated. manifests/evidence.json is left
to the integrating coordinator.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

… change resolves

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Claude Code 2.1.284 (measured 2026-09-29): ultracode neither sets nor overrides
a per-model effort level. An unsaved claude-sonnet-5-5 session ran at medium under
the host settings and the guard stayed silent, because resolve() returned xhigh for
any settings file with `ultracode: true`. That also kept the SessionEnd healer from
saving a level, since it only saves when resolve() gives none.

- resolve(): drop the ultracode branch. The level now comes only from
  CLAUDE_CODE_EFFORT_LEVEL, the highest-precedence saved modelSettings or applicable
  top-level effortLevel, and the maxEffortLevel caps. No other logic changed.
- Docstring: record the 2.1.284 measurement and that 2.1.281 behaved differently
  (docs/decisions/2026-09-23-max-effort-default.md stays as history).
- Tests: an unsaved Sonnet 5.5 under ultracode now warns; a saved xhigh stays silent;
  a saved lower level still warns; under ultracode the healer saves xhigh when nothing
  is saved, and respects a saved low.
- SHA256SUMS: regenerate the effort-default-guard.py line.

The installed host copy (~/.claude/hooks/effort-default-guard.py) is not touched here.
Refresh it with tools/adoption/install_claude_profile.py --only guard; until then
test_shipped_guard_is_verbatim fails on a host that has the old copy installed.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…ement

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…om the transcripts

The 2026-09-28 model-currency addendum counted the Sonnet child's 82 transcript rows as
messages; it made 38 (all claude-sonnet-5-5 at max), and the first receipt drafts carried the
same kind of row counts (299 for the coordinator). The receipt is regenerated by a script that
asserts every row (model, effort, reminder lines, message count, client version) against the
native transcripts before writing, takes the run times from them, and embeds one extractor per
kind (headless cases, interactive cases) with a per-script digest. Its claim now separates the
headless findings from the interactive one (the lead's max reached every child, the Opus ones
included), scopes the effort findings to the setting form and to levels passed with --settings,
and counts 27 probes and two session observations. The addendum's count is corrected and the
mistake goes in the anti-pattern log.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…cument the same-family alias

The SessionStart warning said a user-settings top-level effortLevel does not apply to Opus 5.5
and later; it applies only to Opus 5, Fable 5.1 and earlier models (Sonnet 5.5 ran at medium
under a top-level xhigh, receipt case A1). The guard's docstring cites the documentation of the
2.1.284 change and scopes its ultracode rule to 2.1.284 and later. child-usage.mjs no longer
says a stage without effort inherits the coordinator's xhigh, and its alias table notes the
documented same-family exception (a family alias resolves to the lead's exact model when the
lead belongs to that family). test-envelope.mjs states the documented order for a stage that
names no model. Both SHA256SUMS files are regenerated and the host guard copy reinstalled.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

… Sonnet rule across files

Repairs the findings of one GPT-6 and three Opus 5.5 reviews of this branch. The workflow
contract, the routing table, the recipes, the portable CLAUDE.md and the dated addendum now
state the Ultracode and effort behaviour with its version (2.1.284: the setting leaves effort
unchanged, --effort ultracode still sets xhigh, Ultracode stays on at other levels; 2.1.281
statements are history). The portable Quality bullet lists the same Sonnet classes and
conditions as the README (a test oracle alone never clears a Sonnet build); the coordinator
model is stated once (Opus 5.5 by default, a Sonnet 5.5 coordinator sends each judgment to an
opus stage). The record quotes the system card and the benchmark pages with their scope (the
three Terminal-Bench 4.0 evaluations disagree and differ in more than the harness, the
FrontierCode result for max, the Vals cost and latency figures, the welfare-section figure it
no longer leans on), names the floor change's branch, records the guard and CLAUDE.md host
steps, and corrects the sealed-run and RUNBOOK statements. The instruction word budget is
re-baselined to 1,356.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

…ions explain, fix the child-usage and coordinator statements

Applies the second-round review of this branch. The interactive inheritance finding is limited
to the children that named no effort (C1, C3 to C6); the project agents of C2 and C8 declare
effort max, so they show the definition, and the receipt, the record, the README, the recipe
and the addendum now say so. The record no longer says child-usage.mjs flags a Sonnet 5 child
on a client older than 2.1.284 (its alias table records that as the documented resolution) and
tells the operator to check `claude --version` before routing a stage to `sonnet`; it names the
floor pull request (#477). The portable CLAUDE.md gains the coordinator condition (a Sonnet 5.5
session hands each judgment to an opus-named stage), one Opus list and one build condition
across README, CLAUDE.md and the record, and the teammate effort follows the lead's session
effort. The recipe's remaining 2.1.281 statements are dated, the non-orchestrated max session
uses the 2.1.284 form (documented, not probed together with --effort max), the settings
example's description lists the two added keys, test comments no longer carry the retracted
rationales, the shell row drops the speed advantage (Vals' Terminal-Bench 4.0 data lists higher
latency and output tokens for Sonnet 5.5), the overturn conditions name measurements this
repository can make, and the anti-pattern row stops citing a generator the repository does
not retain. The instruction word budget is 1,372.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

A read-only GPT-6 Astra review at max of head 52d11856 returned four medium and one low finding
and no high one; four are confirmed against the repository text or the documentation and are
one-line edits, and the fifth (the receipt not yet registered) is the deferred hot-file step.
The recipe states the documented model order for a stage that names no model (definition, then
CLAUDE_CODE_SUBAGENT_MODEL, then the lead) instead of the lead alone. A Sonnet coordinator no
longer verifies teammate claims itself: the lead collects results and has any claim verified on
Opus, in the portable CLAUDE.md and the record's teammate row. The two test comments scope the
headless observation and state frontmatter precedence and the interactive case. The README
carries the same Haiku statement as the portable instructions. The record notes that the floor
pull request has merged.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>

Registers the dispatch receipt and the hashes of the files this branch changes on the
current main copy of manifests/evidence.json, the only shared hot file this lane touches.

Co-Authored-By: Claude Sonnet 5.5 <[email protected]>