feat(models): add DeepSeek V4.1 Flash (catalog sync to 1.53.0 + pricing)

[myohei]

Summary

DeepSeek V4.1 Flash (deepseek/deepseek-v4.1-flash) is now available on Command Code. This PR:

1. Syncs the generated catalog from command-code@1.44.0 → command-code@1.53.0 (npm run sync:commandcode-catalog):
   • Adds deepseek/deepseek-v4.1-flash — reasoning true, efforts low/high/max, text+image input, 1M context.
   • Also picks up gpt-6-astra, xai/grok-4.6, inclusionai/ling-3.0-flash-sante:free, and publishes upstream's own Muse Spark effort levels.
2. Drops the manual Muse Spark effort overrides — upstream 1.53.0 ships them now; the override test requires removal once upstream publishes levels.
3. Adds pricing for the 2026-09-10 catalog snapshot (src/pricing.ts + test fixtures):
   • deepseek/deepseek-v4.1-flash at the documented off-peak rate ($0.15/$0.60, cache $0.003).
   • The other six models present in the refreshed snapshot (Qwen 3.8 Max 0902, Gemini 3.8 Flash, Muse Spark 1.3 / 1.3 Contributor, LongCat 2.0 free, Ling 3.0 Flash Sante free).
   • Corrects deepseek/deepseek-v4-flash to the documented off-peak rate ($0.15/$0.60, cache $0.003).

Verification

• npm run typecheck ✅
• npm run test:unit ✅ (all suites pass, including the catalog/pricing fixtures)
• Live API probe: deepseek/deepseek-v4.1-flash accepts reasoning_effort (low/high/max) and tool calling against https://api.commandcode.ai/provider/v1/chat/completions ✅

Note: V4.1 Flash carries a boosted per-model allowance on GOAT/Pro through 2026-09-17 ($60 GOAT / $70 Pro); per-token rates are unchanged and no temporary pricing entry is needed.

[zidanefaqih]

Independent verification of this PR — I applied the equivalent changes locally on top of 0.6.4 and tested against the live endpoint. It works:

• --thinking max on deepseek/deepseek-v4.1-flash reaches the wire. The stream now carries thinking_start / thinking_delta / thinking_end (19 deltas on a short prompt), which on 0.6.4 was unreachable: reasoning:false made core.ts drop reasoning_effort before serialization and thinkingMetadataForModel returned undefined, so nothing surfaced.
• Image input works: @red-square.png returns "Red". Without the MODEL_INPUT_MODALITIES entry the host declared the model text-only and the image was rejected before it ever hit the API.
• Local cost matches recorded usage exactly — (148×0.15 + 17×0.6 + 1152×0.003)/1e6 = 3.5856e-05, identical to usage.cost.total.

Three review notes:

1. This PR and #86 are near-duplicates — only one can land cleanly.

Both bump COMMAND_CODE_CLI_VERSION to 1.53.0 and both add the same seven MODEL_COSTS keys (deepseek/deepseek-v4.1-flash, Qwen/Qwen3.8-Max-0902, google/gemini-3.8-flash, meta/muse-spark-1.3, meta/muse-spark-1.3-contributor, meituan/LongCat-2.0:free, inclusionai/ling-3.0-flash-sante:free).

The actual differences:

	#86	#93
gpt-6-astra pricing + modality	yes	no
deepseek/deepseek-v4-flash 0.22 → 0.15	no	yes
fixture fetchedAt assertion	/^2026-09-/	/^2026-09-10T/

Worth deciding which one merges and closing the other, or folding the missing piece into whichever lands first.

2. Emptying MODEL_EFFORT_OVERRIDES is user-visible, not a no-op.

The comment says command-code@1.53.0 publishes efforts for every reasoning model tracked, which is true — but the published sets differ from the overrides being removed:

model	override (removed)	catalog 1.53.0 (added)
meta/muse-spark-1.1, 1.2, 1.2-contributor, 1.3-contributor	minimal, low, medium, high, xhigh	low, medium, high, xhigh
meta/muse-spark-1.3	minimal, low, medium, high, xhigh	low, medium, high, xhigh, max

So minimal disappears for all five, and 1.3 gains max. The new sets match upstream, so this looks correct — but anyone on /thinking minimal for those models silently loses the option, and it reads as a cleanup when it probably deserves its own CHANGELOG line.

3. Leaving deepseek/deepseek-v4-flash-vision-exp at 0.22 is correct — please keep it.

The CLI's bundled models.md lists $0.15/$0.6 · cache $0.003 for that model, which invites "fixing" it to match flash. The pricing page JSON at PRICING_SOURCE_URL disagrees, and that is the source pricing.ts documents:

• deepseek-v4-flash-vision-exp — off-peak 0.22/0.66/0.007, peak 0.44/1.32/0.014
• deepseek-v4-flash — off-peak 0.15/0.6/0.003, peak 0.3/1.2/0.006

Both #86 and #93 keep 0.22, matching PRICING_SOURCE_URL. A note in the PR or an inline comment in pricing.ts would stop a future contributor or catalog sync from "correcting" it back to 0.15.

Also worth noting for sequencing: #91, #92, and #93 all modify src/commandcode-catalog-overrides.ts, so they will need rebasing in some order.

[patlux]

Consolidated this into pricing on current main; kept the newer 1.54.0 capability snapshot from #98. Re-fetched the live 69-model catalog and reviewed the official pricing page on 2026-09-15. Qwen 0902 and Gemini 3.8 cache writes are zero (unlike #85); Vision Exp now explicitly shows 0.15/0.60/0.003, so its older 0.22 rate was corrected too. Added explicit regression assertions and removed fixture-wide formatting noise. Full npm test passes including Pi and OMP, format and diff checks clean. Unique #86 Astra pricing was reviewed but intentionally not added: Astra remains absent from the live API, and preserving exact live-catalog pricing coverage avoids a special-case test exception. Its current long-context cache-write price is 25, not 12.5 as proposed in #86; add it with verified rates when API discovery exposes it. Other overlapping #85/#86 pricing is covered here. Waiting for CI.

[patlux]

Shipped in pi-commandcode-provider@0.7.0: https://github.com/patlux/pi-commandcode-provider/releases/tag/v0.7.0 . Public npm package model discovery and GOAT live smoke verified.