fix(tui): handle fixed-sampling compaction and report post-compaction context

[qiuYliangM]

Background

Compaction requests use an explicit temperature: 0.3. The Kimi Code membership route rejected that value for kimi-for-coding, so manual and automatic compaction could fail while normal chat continued to work. Kimi Code's current official roster also includes the K3 IDs k3 and k3-256k, which share the membership route's fixed-sampling contract.

The Pinvou host also needs a post-compaction context estimate immediately; otherwise its usage chip keeps showing the stale pre-compaction value until another provider usage event arrives.

Changes

• On the exact Kimi Code membership endpoint, omit client-supplied temperature and top_p for the four documented membership IDs: k3, k3-256k, kimi-for-coding, and kimi-for-coding-highspeed. Custom gateways and unknown/future IDs remain untouched.
• Register k3-256k as a first-class Kimi Code model with the K3 nested reasoning contract, a 262,144-token static context window, and a conservative 131,072-token output ceiling.
• Keep bare k3 separate for membership-plan context handling: only bare k3 can enter the existing 1M plan-tier/provider-metadata path; k3-256k uses the ordinary fixed model metadata path.
• Resolve exact known model metadata before numeric name hints, so k3-256k is not misread as 256,000 tokens.
• Keep the observed DeepSeek workaround only on the exact deepseek-v4-flash Responses route. DeepSeek Chat preserves the documented temperature field, including on custom and lookalike hosts.
• Add CompactionCompleted.post_input_tokens using the engine canonical complete input estimate, including the system prompt and accumulated compaction summary.
• Add regression coverage for route lookalikes, unknown/future model IDs, K3 route boundaries, fixed sampling, model metadata, and the emitted host token value.

Validation

• cargo fmt --all -- --check
• cargo test -p codewhale-tui --lib --locked kimi_code — 48 passed
• cargo test -p codewhale-tui --lib --locked forkguard_ — 27 passed
• cargo test -p codewhale-tui --lib --locked models::tests — 27 passed
• exact K3 request-boundary tests — 2 passed
• exact K3 product-pairing regression — passed
• cargo check -p codewhale-tui --locked — passed
• Original contributor validation: live Kimi Code manual compaction succeeded after the sampling fix

Known risks

• Current DeepSeek documentation advertises a 0..=2 temperature range, and live testing on 2026-08-25 could not reproduce the original DeepSeek 400. The compatibility shim remains only for the exact V4-Flash Responses path; Chat and every other model preserve their sampling contract.
• k3-256k gets the documented 262,144-token static model fact and does not inherit bare k3's automatic 1M entitlement path. The existing explicit context_window operator override remains authoritative; this PR deliberately does not add a model-specific clamp or new limit-provenance machinery.
• The host consumption and fork register updates are stacked in fix(codex): integrate compaction usage and fixed-sampling compatibility pinvou-agent#322 and should pin the final commit produced when this PR lands.

No-Issue: repository Issues are disabled (creation returns 410); the reproduction and validation evidence are recorded above.

[github-actions]

Thanks @qiuYliangM for taking the time to contribute.

This repository is observing a maintainer-managed PR intake gate in dry-run mode, so this pull request is staying open. This note helps maintainers prepare the allowlist before any enforcement is considered.

Please read CONTRIBUTING.md for the expected contribution shape. A maintainer can grant recurring PR access by commenting /lgtm on a pull request.

[h3c-hexin]

I pushed two corrective commits after reviewing the full route and host contracts:

• 2e9eb9c338517fb830effac6161e9ca6dd7274d0 tightens the original compaction and usage contracts.
• 7f1a592a993d0a3cddc504df046d854a5dfff3ec adds the missing current Kimi Code K3 model variants.

The resulting implementation:

• preserves the documented DeepSeek Chat sampling contract and limits the remaining Responses workaround to exact deepseek-v4-flash;
• recognizes all four model IDs in the current official Kimi Code membership roster: k3, k3-256k, kimi-for-coding, and kimi-for-coding-highspeed;
• omits temperature and top_p only for those exact IDs on the exact first-party membership endpoint;
• gives both K3 IDs the same nested reasoning dialect while keeping bare k3 alone in the existing 1M plan-entitlement path;
• records k3-256k as 262,144 context tokens and prevents the generic name-hint parser from misreading it as 256,000;
• reports the engine's complete post-compaction input estimate, including the system prompt.

I initially explored a larger model-specific context-clamp/provenance change, but removed it before committing. The final supplement reuses the existing roster, route predicate, model metadata, and operator override paths; no new foundation abstraction was added.

Validation completed:

• cargo fmt --all -- --check
• Kimi Code suite — 48 passed
• fork-guard suite — 27 passed
• model metadata suite — 27 passed
• exact request-boundary tests — 2 passed
• cargo check -p codewhale-tui --locked

The dependent Pinvou/pinvou-agent#322 gitlink should be finalized only after this PR lands, using the resulting commit that is reachable from Pinvou/CodeWhale.

[qiuYliangM]

Verified the fixed-sampling claims against the live endpoints with real first-party credentials (results from today, 2026-08-25):

DeepSeek official endpoint — 400 not reproducible anywhere

Request	Result
Chat deepseek-v4-pro, temperature: 0.3	200
Chat deepseek-v4-pro, temperature omitted	200
Chat deepseek-v4-flash, temperature: 0.3	200
Responses deepseek-v4-flash, temperature: 0.3	200
Responses deepseek-v4-pro, temperature: 0.3	200
Responses flash, streamed + instructions + input items + top_p: 0.95, temperature: 0.3	200

So the original commit message's claim that the official DeepSeek v4 routes reject non-1 temperature does not hold today. Either the provider relaxed the constraint after 2026-08-19, or the observed 400 was misattributed. Consequences for this revision:

• Dropping the Chat-side apply_deepseek_v4_official_fixed_sampling is safe — v4-pro (the model actually bound to the Chat dialect) accepts 0.3.
• The Responses shim narrowed to exact deepseek-v4-flash no longer fixes a reproducible failure either; it is harmless insurance for compaction-only traffic. Keep or drop, both defensible.

Kimi Code membership route — 400 reproduced exactly

• POST https://api.kimi.com/coding/v1/chat/completions, model: kimi-for-coding, temperature: 0.3 → 400 {"error":{"message":"invalid temperature: only 1 is allowed for this model"}} (verbatim the error in the original commit message).
• Same request with temperature omitted → 200.

So the real fixed-sampling route is the Kimi Code membership route, and the exact-ID allowlist (kimi-for-coding, kimi-for-coding-highspeed) covers the model IDs actually configured on that route. The narrowing from prefix match to the two documented IDs is fine.

Minor: the api.deepseek.com.evil.example case in forkguard_deepseek_v4_chat_preserves_documented_temperature is now a dead assertion — with the host-matching helper removed, no Chat-side code path distinguishes lookalike hosts at all. Consider dropping that case or rewording the comment.

Cross-repo heads-up: the host wiring in pinvou-agent PR Hmbown#322 currently pins the CodeWhale gitlink to 6062fde85 (before this revision) and adds fork-guard fingerprints referencing apply_deepseek_v4_official_fixed_sampling( / forkguard_deepseek_v4_official_route_strips_fixed_temperature / forkguard_deepseek_v4_responses_drops_non_one_temperature. I'll update Hmbown#322 to match this revision: gitlink → 2e9eb9c33, fingerprints → the new symbol names, and the fork-modifications.md entry → the complete-estimate (system prompt included) semantics of post_input_tokens.

[qiuYliangM]

Follow-up: the branch was based on a36e6cd53 (#13) and no longer contained #15/#16/#17 — any consumer pinning the gitlink to the old head 2e9eb9c33 would silently drop the r9 work (steer, edit-last-turn, Shell cancellation). I rebased the branch onto the current pinvou3-clean head 07d183e35; the rebase was conflict-free and now includes the latest K3-variant commit as well. New head: ad3f26b39 (force-pushed with lease). pinvou-agent#322 will pin this SHA.

One note on the K3 commit: the membership allowlist grew to four IDs (k3, k3-256k, kimi-for-coding, kimi-for-coding-highspeed) behind is_kimi_code_membership_model — verified the forkguard_kimi_code_coding_plan_strips_non_one_temperature / forkguard_deepseek_v4_chat_preserves_documented_temperature names are unchanged, so the parent fork-guard fingerprints stay valid.

[h3c-hexin]

/lgtm

[github-actions]

@qiuYliangM is already approved for pr contributions in .github/APPROVED_CONTRIBUTORS.