Haiku 5.5 takes cheap exact extraction: record the 2026-10-07 repeat trial (land after 2026-10-08T17:18Z) - #840

Scope

• What this PR changes, in one or two sentences: records the 2026-10-07 Claude Haiku 5.5 repeat trial as a sanitized receipt and, because it meets the overturn condition this repository recorded for "Cheap exact extraction | Haiku | Not routed", routes cheap exact extraction (inventories and line-cited extraction that relay no command output) to Haiku 5.5 through a per-call model: 'haiku' on source-scout at effort max. Command running and verbatim output relay stay on source-scout's Sonnet default and judgment stays on Opus. The background-haiku currency row moves to claude-haiku-5-5, and child-usage.mjs gains a haiku alias row.
• Land no earlier than 2026-10-08T17:18Z (1:18 PM EDT on October 8). That is 24 hours after Claude Code 2.1.293 reached npm (2026-10-07T17:18:04.464Z), the release that made Haiku 5.5 the default Haiku. The command center lands it on cue; it is open for review, not for an earlier merge.
• Base commit: 36654a8174b5cad1380bd13ed1fffc0f135aac2c (origin/main when the branch was cut).
• Lane: lane:foundation. This PR also changes what manifests/stack.json says (the Haiku model row moves from claude-haiku-4-5-20251001 to claude-haiku-5-5) and adds a receipts[] entry to manifests/evidence.json. Enforce dated model currency and propose daily source reviews #826 changed the same rows under lane:foundation, so the label follows it; relabel to lane:shared if docs/lanes.md's "changes what a shared hot file says" is read to cover this.
• Owned paths touched: examples/claude-native/workflows/{README.md,child-usage.mjs,test-child-usage.mjs,SHA256SUMS}, examples/claude-native/CLAUDE.md and its generated copy adoption/new-wsl/claude-user-instructions.md (regenerated with tools/adoption/new_wsl_client_config.py --write-blocks), recipes/claude-native-ultracode.md, recipes/claude-native-profile.md, catalogs/foundation/{decisions.json,model-currency.json}, docs/model-currency.md, docs/decisions/2026-09-30-task-model-routing.md (one row, one overturn clause), docs/decisions/2026-09-29-sonnet-5-5-dispatch.md and docs/decisions/2026-10-07-model-currency-enforcement.md (one dated note each), evidence/receipts/haiku55-extraction-trial-20261007.json (new). The hot files manifests/stack.json and manifests/evidence.json are in the last commit, per the hot-file protocol.

SOTA sources

• Claude Code 2.1.293 (anthropics/claude-code; npm @anthropic-ai/[email protected], published 2026-10-07T17:18:04.464Z, dist.shasum d653d8b46b676742dfd6dea808bfcc82f2fc0fd8, integrity sha512-PfPSvsd0zH1R1NhLKLebYdEXg/KcGCXc7LBntE186EWuZtBXa7sMnBqs0/uW/BvzkBEfHKqww/WbK/kMAk0TXg==): CHANGELOG.md line 5 at 79babc372d64101f981bd2b52c3dbe588596dc56 reads "Added Claude Haiku 5.5 (claude-haiku-5-5), now the default Haiku model on the Anthropic API — 1M context, $0.10/$0.50 per Mtok ($0.50/$2.50 for prompts over 100K)". That repository publishes no CI run at the tag.

• Claude Code model configuration (version history, read 2026-10-07T23:58Z; the page carries no version):

  • "v2.1.293 | haiku resolves to Haiku 5.5 on the Anthropic API";
  • the "Earlier" row, where haiku resolves to Haiku 4.5;
  • the provider table: Bedrock, Google Cloud's Agent Platform, Microsoft Foundry and Claude Platform on AWS still resolve haiku to Haiku 4.5;
  • "Use v2.1.293 or later with Haiku 5.5".

  These are the source of the child-usage.mjs row, the README's version check and the Anthropic-API-only scope.

• Claude Code sub-agents (choose a model): the per-invocation model parameter comes first in the model order, and haiku is an accepted alias. This is why the route is a per-call override, not a new agent type.

• Claude Code advisor (advisor): subagents inherit the configured advisor, which is the source of the trial's advisor disclosure.

• Claude Platform: the model-currency and stack.json rows rest on

  • the release-note entry of October 7, 2026, the RSS permalink (the plain #october-7-2026 section is an unrelated Sonnet 5.5 cache-price change);
  • the Haiku 5.5 model page: "Released October 7, 2026", dateless claude-haiku-5-5, no dated snapshot;
  • the models overview.

• Effort (effort doc): Haiku 5.5 supports all five levels, max included, and defaults to medium. Anthropic advises xhigh or max only where evals show a gain. This is the basis for the recipe's effort sentence, the "effort is the stage rule" note, and leaving out a claude-haiku-5-5 settings pin.

• Trial design: the overturn condition recorded in docs/ultracode-token-routing-20260921.md and the workflows README: "a repeat trial in which Haiku returns no unanchored citation on two distinct extraction packets". It ran on the native Claude Code Workflow tool with this repository's source-scout and evidence-reviewer agents and child-usage.mjs.

• Prices: LiteLLM model_prices_and_context_window.json (BerriAI/litellm main, ETag W/"35442afac19b8c20555fa17a4f487c7998c99b658444f04820628f1010f40a85", the copy ccusage caches). This is a third-party table and is labelled as one; its Haiku 5.5 base rate equals the CHANGELOG's $0.10/$0.50.

Evidence-class table

Local commands run

All commands ran on the final head e72e070f unless noted.

$ python3 scripts/validate.py
{"components": 69, "hashed_files": 10557, "profiles": 4, "receipts": 227, "status": "passed"}   exit 0
$ python3 scripts/validate.py --scan-file evidence/receipts/haiku55-extraction-trial-20261007.json
{"scanned_files": 1, "status": "passed"}   exit 0
$ python3 scripts/host_receipts.py validate                                     exit 0
$ python3 scripts/release_due.py --strict-if-repinned origin/main               exit 0
$ python3 scripts/validate_catalogs.py                                          exit 0
$ python3 scripts/validate_foundation.py --root . --json                        exit 0
$ python3 scripts/landscape.py --root .                                         exit 0
$ python3 blueprints/blind-catalog-convergence/audit_reports.py --check         exit 0
$ python3 scripts/validate_convergence.py --all-recorded --root . --json        exit 0
$ python3 tools/sota-convergence/build_verdicts.py --check                      exit 0
$ python3 scripts/component_matrix.py --check     {"rows": 32, "status": "checked"}             exit 0
$ python3 scripts/new_host_grand_list.py --check  {"status": "passed", "layers": 32, "winners": 66}   exit 0
$ python3 tools/sota-convergence/gap_crosswalk.py build --check                 exit 0
$ python3 tools/sota-convergence/gap_wave_ledger.py --wave gap-wave2-20260923 --wave gap-wave3-20260923 --owner gap-resolution --check   exit 0
$ python3 scripts/build_ecosystem.py --check                                    exit 0
$ python3 scripts/evidence_manifest.py --check                                  exit 0
$ python3 tools/adoption/new_wsl_client_config.py --check                       exit 0 ("check passed"; one pre-existing mcp-inspector warning)
$ (cd examples/claude-native/workflows && sha256sum --check --strict SHA256SUMS)   exit 0
$ node test-envelope.mjs            SUMMARY passed=254 failed=0 total=254       exit 0
$ node test-contract-mutations.mjs  SUMMARY passed=74 failed=0 total=74         exit 0
$ node test-child-usage.mjs         SUMMARY passed=174 failed=0 total=174       exit 0   (2 SKIP lines: no local tree-sitter-bash install; CI provisions it)
$ node test-usage-receipts.mjs      SUMMARY passed=18 failed=0 total=18         exit 0
$ node test-codex-envelope.mjs                                                  exit 0
$ node --test blueprints/native-skill-practice/test-contract.cjs   pass 6 fail 0   exit 0
$ python3 -m unittest   (full suite) started on e72e070f; its result was not recorded before this session hit its usage limit; CI runs it

The trial ran on 2026-10-07 from 21:31Z to 23:48Z (5:31 PM to 7:48 PM EDT). An independent three-stage Opus review covered the diff, the receipt against the private evidence, and a re-run of the checks. Its findings are folded in: the advisor disclosure and split costs, both rounds' verdicts, the portable instructions' stale line, the settings-pin and client-minimum notes, and the wording and number fixes. A second reviewer confirmed all 11 fixes and raised two items left open here: (A) catalogs/foundation/upstream-surface-dispositions.json:73-74 and :313-314 cite child-usage.mjs:67 and :1925, which the new comment line moved to :68 and :1926; (B) the receipt's P2 Haiku executor mean ($0.0180) and P1 Haiku advisor-inclusive mean ($0.2407) are means of unrounded per-child costs, so the rounded per-run values give $0.0181 and $0.2408, and provenance.cost_method should say so. A wording nit: README:1084 says "relays no command output verbatim" where CLAUDE.md says "relay none".

Decision record

The routing decision is recorded in these places:

• The new "Haiku 5.5 for cheap exact extraction (2026-10-07)" section of examples/claude-native/workflows/README.md: evidence, boundaries, advisor, cost basis, settings pin, client minimum and overturn.
• docs/decisions/2026-09-30-task-model-routing.md, the task-to-model routing record: one row and one overturn clause.
• Decision lean-workflow-child-routing in catalogs/foundation/decisions.json.
• Dated notes in docs/decisions/2026-09-29-sonnet-5-5-dispatch.md (the pin is left out on purpose) and docs/decisions/2026-10-07-model-currency-enforcement.md (the currency row).

Alternatives considered:

• Keep every extraction on Sonnet. Rejected: the trial shows equal accuracy at under a tenth of the executor cost per child.
• A new Haiku agent type. Rejected: the per-call override comes first in the documented model order, and it leaves the agent and role tables, which the envelope suite pins, unchanged.
• Route command relay too. Rejected on the 5 of 6 relay result.
• Add a claude-haiku-5-5 settings pin. Deferred: it would lift every Haiku session that names no effort without an eval.

Overturn: send extraction back to Sonnet when a routed Haiku stage returns an unanchored or misattributed citation that a later check finds, or when a haiku child resolves a model older than Haiku 5.5.

Host evidence

Not applicable: no file under evidence/hosts/ changed.

Merge interactions, follow-ups and limits

• Align code graph MCP server key with native client registrations #834 also edits examples/claude-native/workflows/child-usage.mjs and test-child-usage.mjs. Whichever lands second rebases and regenerates that directory's SHA256SUMS with sha256sum -- *.mjs *.js *.json.

• Make standing Ultracode opt-in and record task dispatch scope #829 and Instruction core: trim the always-loaded text of both clients; trading rules load only where they apply #836 edit the workflows README, the Ultracode recipe or the portable instructions; Instruction core: trim the always-loaded text of both clients; trading rules load only where they apply #836 trims examples/claude-native/CLAUDE.md, which this PR touches in one clause. Document retrieval-first local-model choices and supersede foundation sources #810, Pin Grafana MCP 2.0.1 and AgentsView 0.44.0 for monitoring #807 and Qualify CLI transfers and retain applied client-tools evidence #776 edit manifests/stack.json. Rebase through the hot-file protocol (docs/lanes.md). After a CLAUDE.md change, rerun tools/adoption/new_wsl_client_config.py --write-blocks.

• Follow-up: once its 24-hour gate passes, raise the claude-code minimum from 2.1.284 to 2.1.293 together in manifests/stack.json and both pin files, as a separate currency change. Until then the README's claude --version check is the route's only guard.

• Follow-up (optional): rerun the comparison with CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1 to hold the advisor equal across arms. The route does not rest on it, because the 6 unassisted Haiku runs were exact.

• Limits:

  • one Mac, one account, one client version and two packets;
  • the costs are not provider billing, and the advisor share depends on rate limits;
  • no lower Haiku effort was compared;
  • the background slot was not read back.

  The receipt's limitations list each one.

Checklist

• New/changed GitHub Actions are pinned to a full commit SHA with a version comment (no floating tags). Not applicable: no workflow file changed.
• New/changed workflows declare top-level permissions: {} and grant each job only what it needs. Not applicable: no workflow file changed.
• No secrets are printed, logged or committed; no new required secret was added. The gitleaks pre-commit found no leaks in either commit, and 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 (not deleted, moved or overwritten).

🤖 Generated with Claude Code

https://claude.ai/code/session_01WhN8fzZs47tp6KunqRQ9gs