fix(ai,coding-agent): address post-merge review of Anthropic fallback credit

[H4vC]

What

folowup to pr #13157

Fixes the post-merge review findings on the Anthropic fallback-credit redemption:

• Only a refusal stop captures a credit handle, so the frozen request body and fct_ token stay out of session files (refusal turns are never persisted).
• The beta-header check now skips adding fallback-credit-2026-06-01 when any fallback-credit-* beta is already present (the duplicated -06-01 check was dead code).
• formatEchoedRefusalContent returns typed ContentBlockParam[] and drops blocks with no wire form (including unsigned thinking) instead of passing internal blocks through.
• The redemption ladder triggers on AIError.status(error) === 400 instead of any error text containing 400.
• AnthropicFallbackCreditHandle.params is MessageCreateParamsStreaming; the as / as unknown as casts and options!.fallbackCreditRedemption! are gone. A fallbackCreditRequestBody helper replaces the two hand-built frozen bodies.
• The client-side redemption window is the named FALLBACK_CREDIT_TTL_MS, with a link to the docs.
• TurnRecovery queues the redemption only when edit mode, system prompt, and tool list are all unchanged after the fallback switch, not just edit mode. The fallbackCreditHandle! assertion is removed.
• The stored redemption turn keeps the refused model's signed thinking, now documented: a tool loop needs the turn to open with its thinking block, and a later signature rejection is recovered by the existing invalid-signature retry. The test now pins that the replayed turn opens with the signed thinking.

Why

Post-merge review of #13157. On the open question about replaying refused-model signatures: an earlier revision of this PR stored that thinking unsigned, but review pointed out that a redemption turn ending in tool_use would then replay demoted text ahead of the thinking, which Anthropic rejects with a 400 nothing retries. Keeping the signature leaves only the failure the code already recovers from.

Not addressed: packages/catalog/CHANGELOG.md has no entry for #13157's catalog changes, but that release section is already published and immutable.

Known remaining gap (pre-existing in #13157): a candidate that can redeem credit skips the same-provider signed-thinking check. If redemption is then cancelled because the switch rebuilt the prompt or tools, that request replays the refused model's signatures without a token, and the existing signature-rejection retry recovers in one extra round trip. The broader guard makes this path more common than the edit-mode-only check did.

Testing

• bun run check:ts passes (oxlint, oxfmt, types in all packages). No Rust changed.
• packages/ai: bun test anthropic, 639 pass.
• packages/coding-agent: agent-session-retry-fallback.test.ts, 125 pass.
• New and changed tests fail on main and pass with this change:
  • a 529 error whose text mentions 400 keeps the continuation shape and token instead of stepping down;
  • a frozen fallback-credit-2026-07-01 beta header is sent without -06-01;
  • unsigned thinking is omitted from the echoed continuation prefix;
  • redemption is not queued when the fallback switch rebuilds the system prompt (parameterized alongside the existing redemption test).
• No live Anthropic run; a real Fable → Opus refusal is still unverified.

---

• bun check passes
• Tested locally
• CHANGELOG updated with the required attribution (if user-facing; internal issue fixes use issue links, external contributions add the PR link and contributor credit after creation)

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-14T13:13:35.875548Z	1217b23	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[roboomp]

Thanks @H4vC, this is a tidy follow-up to #13157. Ranked P1: one fix, scoped to that feature, and the new tests fail on main and pass here. I ran the three changed packages/ai test files locally: 86 pass. The coding-agent test could not run here because the generated tool-views.generated.js is missing.

• Correct as claimed: AIError.status(...) === 400 replaces the /\b400\b/ text match; the fallback-credit-* prefix guard; the typed AnthropicFallbackCreditHandle.params with the casts removed; the fallbackCreditRequestBody dedupe; token capture only when stopDetails.type === "refusal" (refusals are already persist-skipped); and the system-prompt/tools/edit-mode guard in TurnRecovery.
• Should-fix: the changelog entries credit #13157 rather than this PR and contributor (inline comment).
• Maintainer call: storedRefusalPrefix stores the prefix thinking unsigned. That trades a signature rejection the code already recovers from for a possible leading-text 400 in tool loops, which nothing retries (inline comment, unverified live). @can1357 should confirm which failure mode is preferable.

[roboomp]

Question for maintainer (behaviour trade-off, non-blocking): storedRefusalPrefix turns a known, recoverable failure into an unknown one.

• Before: the stored turn replayed the refused model's signed thinking. If the server rejected it on a tokenless turn, the error matched INVALID_THINKING_SIGNATURE_PATTERN (anthropic.ts:1952), and the existing signature retry recovered in one round trip.
• After: the prefix thinking is stored with no signature. convertAnthropicMessages (anthropic.ts:5842-5871) then demotes it to text, or drops it.

If the redemption response continues into a tool_use, the next request (the one carrying the tool_result) replays an assistant turn that starts with demoted text, not with a thinking block. With thinking enabled, Anthropic documents a 400 for that case: "a final assistant message must start with a thinking block". [INFERENCE: not verified live for Opus 4.8 / adaptive thinking.] No retry pattern in this file matches that error, so the PR body's "cost is a prompt-cache miss" may understate the risk in a tool loop. The added test covers only a text continuation (anthropic-stream-envelope.test.ts, the "stores the refused thinking unsigned" case), not one that ends in tool_use.

Your PR body already raises this as an open question to @can1357. A maintainer should pick a side before merge.

[roboomp]

should-fix (convention): This is an external contribution, so per AGENTS.md § Changelog these entries should be attributed to this PR, ([#15702](https://github.com/can1357/oh-my-pi/pull/15702) by [@H4vC](https://github.com/H4vC)), not #13157. Line 10 also goes into root cause ("which the fallback model could reject without the credit token; the echoed thinking is now kept unsigned..."). Changelog entries should say what the user sees; put the mechanism in the commit or PR. The same attribution fix applies to packages/coding-agent/CHANGELOG.md:22.