[asto18089]

Summary

Teaches the reasoning-replay gate the K3.1 family spellings so a K3.1 tool round replays reasoning_content like bare K3 does.

The replay gate only learned the bare k3 ids through is_exact_direct_moonshot_k3_route / is_exact_kimi_code_k3_route. The k3.1-pro[-highspeed] Kimi Code membership spellings fell through to the generic heuristics, which never matched, so the sender omitted the assistant tool-call message's reasoning_content on the follow-up request — the Kimi Code error reference documents that missing field as HTTP 400. The kimi-k3.1-pro[-highspeed] direct-platform spellings did replay at default effort via the generic kimi- prefix heuristic, but were still stripped whenever a stored effort of off was present; the review-round arm (2026-11-04) makes all four spellings replay catalog-independently and effort-immune on both vendor routes, including the overseas api.kimi.ai/coding/v1 endpoint.

Contract

• models-overview: the K3.1 spellings carry kimi-k3's parameter constraints verbatim ("参数约束也与 kimi-k3 完全一致") — always-reasoning, so the replay rule applies with no effort gate.
• Kimi Code error reference (missing-reasoning-content-field): the replayed tool-call assistant message must carry reasoning_content.

Scope

Deliberately narrow: the gate gains its own K3.1 arm on the vendor's two official routes instead of extending the shared exact-K3 matchers, whose other consumers (effort shaping, route validation, pickers) carry different K3-only contracts. Regression pins both routes with a tool-call round (assistant thinking + tool use → tool result → assert reasoning_content replays).

Context

Surfaced by the pinvou-agent#825 review round 3 (the app's model-catalog refresh adds these presets); the consumer's wire-body harness is ready to pin the second request once this lands.

No-Issue: follow-up to the pinvoy-agent#825 model-catalog review; this repository has issues disabled.

[asto18089]

Rebased onto the current #131 head (0583e19cb): this branch was cut from the pinned candidate head f1eb10efc, which has since advanced by 29 commits, and #131 is still open — so this PR's chain carried a stale slice of the candidate lineage.

The single K3.1 commit was replayed onto 0583e19cb with an identical patch-id (64c64a9b); message and DCO unchanged. Gates on the rebased tree: kimi_k31 regression 1/1, kimi + replay_reasoning suite 126/0. New head: 1d9b74bec.

Also retargeted this PR's base from main to feat/host-armed-guardrail-prompts (#131's head branch): with main as base the commits tab carried the whole candidate lineage (2,300+ commits); stacked on #131 it now shows exactly the one K3.1 commit. When #131 merges, this PR needs a re-cut onto the squash and a retarget back.

[asto18089]

Pushed the overseas-endpoint follow-up for the review P2 (5765736266 / 5765853804): 5e4545615 teaches both Kimi Code route gates the official api.kimi.ai/coding/v1 host the host catalog registers and prefs preserves verbatim.

• Exact gate (is_exact_kimi_code_route) + loose gate (moonshot_base_url_uses_kimi_code) now accept the .ai host; neighboring paths, wrong scheme/port, and custom gateways stay outside (test matrix extended with .ai positives and .ai neighbors).
• The K3.1 tool-round regression now pins both bare ids (k3.1-pro / k3.1-pro-highspeed) on the overseas route, alongside the existing .com / direct-platform arms.
• Credential guidance strings updated to name both official hosts.
• Negative validation: removing the .ai arm turns the endpoint-gate test red; full codewhale-config 644 pass / 0 fail, kimi family 125 pass / 0 fail, cargo fmt --check clean.

Sequencing note for the parent pin: once this and the #131 stack land on pinvou3-clean, the codewhale-hq#825 fork-guard registration and gitlink flip to the public head and the parent fast-gate goes green — the current pin (1d9b74bec) will need one more bump to this SHA after the public advance.

[asto18089]

Fresh multi-lane review (7 lanes + 2 verification lanes: gate semantics, shared-matcher consumer audit, Kimi-family sweep, other-vendor sweep incl. all pinvoy-agent#825 families, mutation testing at both heads, smuggle/CI classification, claim verification). All findings verified at head 5e4545615. Verdict: the fix is real, root-caused, and well-scoped; requesting changes on three items before merge.

Verified-good

• Root cause confirmed: build_chat_messages_with_reasoning emits reasoning_content on the replayed tool-call assistant message only when should_replay_reasoning_content_for_provider_on_route returns true (chat.rs:2885,2918-2920); pre-PR the gate only learned bare K3 via the exact matchers. The new arm is catalog-independent, effort-immune (consistent with the always-reasoning contract), and correctly ordered (no shadowing with the K3 arms — spelling sets are disjoint, all matchers exact-compare).
• Dedicated arm (vs extending the shared matchers) is justified: ~23 production consumers of is_exact_direct_moonshot_k3_route/is_exact_kimi_code_k3_route gate K3-only contracts (effort dialects, route shapes, picker ladders, output limits, search timeout floor, two display labels); teaching the shared matchers K3.1 would have flipped all of them at once. PR body claim verified factually accurate.
• Commit 2's .ai gate arm is exactly as strict as the .com arm (is_exact_https_route reuse), with negative neighbors (case/path/scheme/port/query) in provider.rs:2226-2253. Loose-helper over-match safety verified (host+path pinned, no substring matching).
• Smuggle: clean — 4 files, every hunk serves the disclosed purpose; no deps/submodule/workflow changes. Both commits DCO-signed, conventional subjects.
• Mutations at 5e4545615: arm deletion, Kimi-Code URL, .ai URL, and 3 of 4 spellings all go red when removed. fmt rc=0; applicable clippy gates rc=0; kimi/replay suites 125/75 green; ubuntu CI passed the full suite with the PR applied.

Must fix

• M1 — doc-comment hijack (chat.rs:3485-3504): the helper is_kimi_k31_model_spelling was inserted between the gate fn's doc block and the gate fn, so the entire "Route-aware reasoning replay policy." block now rustdoc-attaches to the 4-line spelling helper; the gate fn is left undocumented; "see the arm inside" (3494) points backwards; sentences 3492-3494 vs 3495-3496 duplicate each other. The inherited sentence "only Kimi Code's exact membership-plan endpoint has this replay contract" (3489-3490) is contradicted by the direct-platform K3 arm (3513) and the new arm itself (3533-3534) as the block now stands. Compiles clean, so CI cannot catch it.
• M2 — commit 2 is incomplete on the config side (split-brain): the TUI exact gate and the TUI copy of the loose fallback learned api.kimi.ai/coding/v1, but the config-crate twins did not: moonshot_base_url_uses_kimi_code (crates/config/src/lib.rs:4619-4624, still .com-only) and provider_base_url_is_official(Moonshot, …) (lib.rs:4874-4882, delegates to it). Verified consequence: a Moonshot route on api.kimi.ai/coding/v1 with no model resolves the default to kimi-k2.7-code (the config-crate fallback misses), which validate_kimi_code_api_model_id then hard-rejects as a direct-platform id on the membership endpoint — a config-resolution failure on exactly the route commit 2 exists to support (the same config works on api.kimi.com/coding/v1). provider_base_url_is_official returning false also contradicts its own "so route canonicalization and credential scoping cannot disagree" doctrine (lib.rs:4852-4856); consumers include client/responses.rs:905, cli lib.rs:3117, app-server chat_completions.rs:153/255.
• M3 — required link gate red ×3, PR-caused: the body has no line-start Closes #N / No-Issue: anchor (checked pr-issue-link.yml's regex; the mid-sentence "pinvoy-agent#825" mentions don't satisfy it). Will add the line in the fix wave.

Minors (will fix in the fix wave unless noted)

• m — PR body overstates the pre-PR bug: the two kimi-k3.1-pro[-highspeed] spellings already replayed at default effort via the kimi- prefix heuristic (models.rs:447); the actual pre-PR failures were (a) bare k3.1-pro[-highspeed] on Kimi Code never replaying and (b) all four spellings stripped at stored effort=off. Will recalibrate the body wording.
• m — "the host catalog registers api.kimi.ai" (provider.rs:517-518, client.rs:9386-9387, commit 2 body): no host catalog in this repo registers it (docs still document api.kimi.com only); name the app-side catalog explicitly or update docs/PROVIDERS.md/docs/CONFIGURATION.md.
• m — typo "pinjou-agent#825" ×3 in code (provider.rs:518, client.rs:9328, client.rs:9387) + commit 2 body.
• m — stale adjacent strings: the membership-rejection error still names only api.kimi.com/coding/v1 (config.rs:10227); Moonshot guidance suggests only the .com URL (config.rs:7075).
• m — test-strength (mutation-verified): unpinned — the direct-platform matcher (masked by the generic kimi- prefix; an effort=off direct leg would pin it), the provider == Moonshot route restriction (no custom-gateway negative), case/trim normalization (no variant input), and the 4th spelling kimi-k3.1-pro-highspeed. Sibling K3 tests (chat.rs:5166-5191, 6228-6282) set the negative-test norm.
• m — commit 2's .ai widening legitimately extends K3 effort shaping / sampling / route shape / billing / pickers to the overseas endpoint, but only the gate + replay legs are tested on .ai.

Follow-ups disclosed (not for this PR)

• OpenRouter moonshotai/kimi-k3.1-pro (a codewhale-hq#825 row): not deterministically covered — the provider == Moonshot restriction is faithful to the aggregator precedent, but the generic heuristics miss the moonshotai/ prefix (only k2.x org ids enumerated, models.rs:473-476), so replay depends on the runtime models.dev cache (fails offline/fresh installs); the same id on Volcengine replays via the bare-kimi- prefix — a pure prefix artifact. Cheap fix: add the org-prefixed spellings to the recognition list.
• MiniMax M3.1-Flash (codewhale-hq#825-served, always-thinking): minimax-m3.1-flash is in no recognition list → no replay, no stream classification. Different vendor contract; genuine follow-up.
• Other codewhale-hq#825-family recognition staleness (pre-existing): zai-org/GLM-5.3 on SiliconFlow, minimax/minimax-m3.1-flash and anthropic/claude-opus-5.5 org-prefixes on OpenRouter, Qwen/Qwen3.8-27B, mimo v2.6 ids, gemini-4 thought-signature matcher, kimi-k2.5 preserve_thinking verification, and the app-side bridge-arm gap for Mistral/Together/Meta/Groq (belongs on pinvoy-agent#825, not here).
• CI reds other than link are base/environmental (verified against base 0583e19cb): Lint = pre-existing clippy::collapsible_if at turn_loop.rs:5611 (untouched file); cargo-audit/cargo-deny = RUSTSEC-2026-0430 in the untouched lockfile; Lint & Type Check = release-data drift; Version drift / Integrations / Test-windows = unrelated subsystems (the PR's own test passed on Windows).
• Upstream-first note: the gate is foundation code; fork-policy letter prefers upstream-first for all-embedder fixes — worth an upstream PR once the stack lands.
• Known pre-existing flake observed during verification: forkguard_model_operation_auxiliary_retry_keys (operation_tests.rs:379) is order-dependent, passes in isolation, unrelated to this PR.

Fix wave incoming for M1-M3 + the minors; follow-ups stay disclosed.

[asto18089]

Fix wave pushed as fd0aef4 + e11b90e (head e11b90ec5), addressing the review round:

M1 — doc-comment hijack (e11b90e): the spelling helper now carries its own doc line; the gate's "Route-aware reasoning replay policy." block is restored directly above should_replay_reasoning_content_for_provider_on_route with an accurate K3.1 paragraph (the duplicated sentence pair and the backwards "see the arm inside" pointer are gone).

M2 — config-crate .ai split-brain (fd0aef4): moonshot_base_url_uses_kimi_code in the config crate now recognizes both membership hosts, with the exact-gate-subset doctrine documented on the helper, and provider_base_url_is_official(Moonshot, …) follows automatically (it delegates). New test kimi_code_loose_route_recognition_covers_both_membership_hosts pins: both hosts exact↔loose↔official agreement, neighboring /coding paths (loose+official but deliberately not exact), case variants, lookalike hosts/scheme downgrades as negatives, and the direct platform staying official-but-not-membership.

M3 — link gate: body now carries a line-start No-Issue: line and the pre-PR claim is recalibrated (the direct-platform spellings did replay at default effort via the generic kimi- prefix; the true pre-PR failures were the bare Kimi Code spellings never replaying and the effort=off stripping — body now says exactly that).

Minors: route-gate comment premise reworded to "the app's model catalog registers api.kimi.ai" (provider.rs); membership-rejection and plan-key guidance strings now name both hosts; remaining pinjou typo in the overseas test comment fixed.

Test-strength (e11b90e): the K3.1 regression test now pins all four spellings across both hosts, plus — each verified red under its targeted mutation before revert —

• effort=off direct-platform leg → pins the direct-platform matcher (mutation: matcher removed → red at exactly this leg; the generic kimi- prefix masks this at default effort),
• custom-gateway + non-Moonshot-provider negatives → pin the route restriction (mutation: provider == Moonshot removed → red at the non-Moonshot leg),
• case and padding variants on the bare Kimi Code ids (mutations: case-sensitized / trim removed → red at exactly those legs).

Erratum to the review comment: the pinjou-agent#825 typo list included client.rs:9328 — that line was already correct; the typo existed only at client.rs:9388 (now fixed) plus provider.rs:518 (fixed) and commit 2's message (history, left as-is).

Local gates: cargo fmt --check rc=0; clippy --workspace --all-features --locked with the fork-ci allowance set, both --lib --bins and --all-targets, rc=0; codewhale-config --lib -- kimi 8/8; tui kimi 125/0, replay 75/0, kimi_code 46/0, upgraded kimi_k31 1/0.

Disclosed, not fixed here (follow-ups): OpenRouter moonshotai/kimi-k3.1-pro recognition (generic heuristics miss the moonshotai/ prefix; replay there depends on the runtime models.dev cache), MiniMax M3.1-Flash recognition (minimax-m3.1-flash in no recognition list), and the other codewhale-hq#825-family enumeration staleness items from the review comment. Upstream-first contribution of the gate change is also worth a follow-up per fork-policy.