feat(models): refresh Command Code catalog to 1.79.2 with claude-haiku-5-5

[sandexzx]

Adds claude-haiku-5-5 from the latest Command Code catalog and clears the drift that had accumulated since the pinned command-code@1.72.4 snapshot.

What changed:

• Synced the static capability snapshot to command-code@1.79.2: new claude-haiku-5-5, mistral/mistral-large-4 (image input, 262K output limit), and free stealth/glyph-cluster:free; retired stealth/pixel-canary and stealth/space-bunny-alpha.
• Recognized the off reasoning effort now published for the DeepSeek V4 and V4.1 models. The daily catalog check was failing on it (Unexpected reasoning efforts for deepseek/deepseek-v4-pro: off, high, max) because VALID_EFFORTS and the generated type union did not know off. The generate transport still never forwards off as reasoning_effort; the Provider API and Oh My Pi adapters translate the level on their own wire.
• Reviewed display pricing against the official page (verified 2026-10-09): claude-haiku-5-5 base plus its 100K long-context tier, mistral/mistral-large-4, and free stealth/glyph-cluster:free; dropped the retired stealth/space-bunny-alpha; corrected claude-sonnet-5-5 cache-read to $0.10/M. The snapshot now covers all 87 advertised models.
• Refreshed both fixtures, updated the affected test expectations, and replaced a stale README note that claimed MODEL_EFFORT_OVERRIDES still added manual levels (the map is empty).

Validation: npm test (exit 0), npm run check:commandcode-catalog (no drift), npm run check:commandcode-pricing (PASS), npm run typecheck, npm run format:check, git diff --check.

[pierreraby]

Thanks for catching this and refreshing the catalog. I reviewed exact head a2a19ef in an isolated checkout; a separate read-only review also found no introduced defect.

Local validation passed:

• Full npm test, including real pi 1.1.0 and OMP 18.4.2 against local mock APIs.
• npm run format:check and git diff --check.
• Read-only live catalog check: no drift against command-code@1.79.2.
• Read-only live pricing check: PASS for all 87 advertised API models.

I also independently checked the changed prices against the official pricing page, including Haiku 5.5's >100K tier, Mistral Large 4, Glyph's free preview, and Sonnet 5.5's $0.10/M cache-read rate. A local offline probe accepts off, high, max, still rejects an unknown effort, and confirms the separate pi/OMP thinking metadata. The generate off filtering already existed before this PR, so this is principally a catalog/parser repair rather than a new generate-wire behavior.

One related CI hardening item would help prevent a repeat: the catalog steps use ... | tee ... with the default shell. In the October 8 scheduled run, the parser printed Unexpected reasoning efforts for deepseek/deepseek-v4-pro: off, high, max, but the synchronization step was marked successful, and no sync PR was created. The run failed later on the independent pricing check.

The log shows /usr/bin/bash -e {0}, without pipefail. I reproduced the exit-code behavior locally: false | tee /dev/null exits 0 under that shell, but exits 1 with -e -o pipefail.

Could you add explicit shell: bash to the catalog comparison and synchronization steps in .github/workflows/model-metadata.yml (or defaults.run.shell: bash for the workflow)? GitHub's explicit Bash shell enables pipefail. This is a pre-existing workflow issue, not a regression introduced by your diff; a small follow-up is also reasonable if you prefer to keep it separate.

Two optional coverage improvements, not merge blockers:

• Build the off-omission stream test with thinkingMetadataForModel("deepseek/deepseek-v4-flash"), so it pins the canonical thinking.effortMap path as well as the older thinkingLevelMap path.
• Add a small offline pricing-page fixture for Haiku 5.5's labeled ≤100K / >100K bands; the existing HTML fixture predates the refreshed catalog, although the live checker passes today.

[pierreraby]

Thanks for addressing the three follow-ups on new head 8645883. I re-verified in an isolated checkout.

The shell: bash fix on both catalog steps resolves the masked | tee exit I reported: explicit Bash enables pipefail, so a parser failure can no longer pass silently and block the auto-sync PR. The Haiku fixture mirrors the live page row and its test compares against the real runtime MODEL_COSTS, so drift would actually fail. The stream test now pins the canonical thinking.effortMap path for off.

Local validation on the new head passed:

• Full npm test, including real pi 1.1.0 and OMP 18.4.2 against local mock APIs.
• npm run format:check and git diff --check.
• Read-only live catalog check: no drift against command-code@1.79.2.
• Read-only live pricing check: PASS for all 87 advertised API models.

Approving — CI is green on this head and the scope matches the reviewed catalog refresh.

[pierreraby]

Approved on exact head 8645883: catalog refresh to command-code@1.79.2, off recognition, and reviewed pricing verified locally (full npm test incl. pi 1.1.0 + OMP 18.4.2 against mocks, live catalog/pricing checks PASS, format/diff-check clean). Follow-up commits address the reported pipefail masking and add the effortMap + Haiku-band coverage.