feat(ai): lower reasoning into bedrock requests

[zmaril]

Follow-up to #309. PR D of the per-driver reasoning-lowering series: wires the bedrock dialect so SimpleStreamOptions.reasoning / thinking_budgets reach the ConverseStream request, using mistral_backend.rs / anthropic_backend.rs (#309) as the override template.

Before / After

Before: BedrockBackend inherited the default Provider::stream_simple, which extracts .base and drops reasoning. A caller asking for a thinking level got a request with no thinking config — the ported BedrockOptions.reasoning / thinking_budgets fields were orphaned.

After: BedrockBackend overrides stream_simple, mapping the seam options into BedrockOptions and routing through a new driver::stream_simple that lowers reasoning into the request exactly as pi does.

How

New driver::stream_simple faithfully ports pi's bedrock-converse-stream.ts streamSimple (:392-438), which is model-family-aware:

• No reasoning (:398): the backend short-circuits to the raw stream, so the request is byte-identical to the pre-seam path (no thinking config added).
• Claude + adaptive (:403-408, gated on supportsAdaptiveThinking): pass reasoning + thinkingBudgets straight through; the driver's build_additional_model_request_fields emits thinking.type = "adaptive" + output_config.effort.
• Claude, non-adaptive (:413-430): fit thinking inside the output cap via the shared adjustMaxTokensForThinking (simple-options.ts:50), re-clamp to the context window with clampMaxTokensToContext (simple-options.ts:15), and override the clamped level's budget with min(adjusted.thinkingBudget, max(0, maxTokens - 1024)) (:428).
• Non-Claude (:433-437): reasoning passes through but the driver ignores it (build_additional_model_request_fields returns None for non-Claude), so no thinking config reaches the request — matching pi.

base.maxTokens in every branch is pi's buildBaseOptions output, i.e. clampMaxTokensToContext(model, context, options.maxTokens ?? model.maxTokens) (simple-options.ts:29).

Helpers

• Promoted clamp_reasoning (xhigh/max -> high) from private to pub(crate) in api/anthropic/simple_options.rs so Bedrock can override the clamped level's budget with the same collapse pi applies (:428). Behavior-preserving for Anthropic — it still calls the helper unchanged from adjust_max_tokens_for_thinking.
• Reused the existing pub adjust_max_tokens_for_thinking (api/anthropic/simple_options.rs).
• Ported a BedrockModel-typed clamp_max_tokens_to_context in api/bedrock.rs (the Anthropic one is typed to Model<AnthropicMessagesCompat>); it reuses the shared context-token estimator pi imports from utils/estimate.ts. Added a context_window field to the lean BedrockModel slice to feed it.
• Claude / adaptive detection (is_anthropic_claude_model, supports_adaptive_thinking) were already ported for the driver's build_additional_model_request_fields; promoted from private to pub(crate) so the sibling stream_simple can reuse them (no re-port).

Tests (param-exact; mirror #309 + pi bedrock-thinking-payload.test.ts)

• Adaptive Claude (anthropic.claude-opus-4-8) + high -> thinking.type = "adaptive" + output_config.effort = "high".
• Non-adaptive Claude (claude-3-5-sonnet) + medium, caller cap 2000, model cap 32000 -> inferenceConfig.maxTokens = 10192, thinking.budget_tokens = 8192.
• Non-adaptive Claude + custom medium budget 20000, model cap 16000 -> budget override bites: maxTokens = 16000, budget_tokens = 14976 (= min(14976, 16000-1024)).
• Non-Claude (meta.llama3) + reasoning -> no additionalModelRequestFields (driver ignores).
• No reasoning -> byte-identical to the raw stream path.

Verification

The mapping was verified directly against pi source (vendor/pi/packages/ai/src/api/bedrock-converse-stream.ts:392-438 and simple-options.ts:15,29,47-77) since the providers co-sign session is dormant. Each branch and helper is cited against those .ts lines in the code comments.

CI

cargo build, cargo test -p pidgin-ai (794 lib + 9 integration passing, incl. 5 new bedrock reasoning tests), cargo clippy --all-targets -- -D warnings, and cargo fmt --check all clean. No touched file crosses the 1500-line straitjacket file-size limit; existing duplication markers preserved.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GkuR3Xbx7yeqvEv6MPbFg7

---

Generated by Claude Code

[github-actions]

Conformance smoke: agent + ai + tui

• pi_sha: 3da591ab74ab9ab407e72ed882600b2c851fae21
• manifest_native_modules: 52

Headline is rust-backed: passing cases in files whose module-under-test is a native (Rust addon) module. Raw all-pass is shown secondary — it is inflated by unflipped TypeScript that passes without touching any Rust.

Package	rust-backed	Native modules	raw pass (secondary)	Failing	Skipped
agent	73/180 (40.6%)	7	180/180	0	0
ai	79/1296 (6.1%)	9	553/1296	3	740
tui	375/678 (55.3%)	14	678/678	0	0
smoke total	527/2154 (24.5%)	30	1411/2154	3	740

agent: 73/180 (40.6%) · ai: 79/1296 (6.1%) · tui: 375/678 (55.3%)

rust-backed = passing / total tests run, per the manifest's per-native-row tests lists. The Native modules column counts modules served by the Rust addon. This is the agent+ai+tui smoke subset; the full baseline lives in committed conformance.json.

Attribution

A pi test file is rust-backed iff the module it primarily exercises — its module-under-test — is status=native in conformance/manifest.json. A package's rust-backed count is the passing cases in its rust-backed files. Transitive or infrastructure use does not count (e.g. ~30 ai files construct the native faux provider but test other things — only faux-provider.test.ts, whose subject is faux, counts). A file that substantially tests both a native and an original module is excluded rather than counted, so the number under-reports rather than over-claims.

Per-file decisions (from each native manifest row's tests list):

Native module	Test file	Decision
ai/api/anthropic-messages.ts	test/anthropic-sse-parsing.test.ts	counted
ai/providers/faux.ts	test/faux-provider.test.ts	counted
coding-agent/utils/ansi.ts	test/ansi-utils.test.ts	counted
coding-agent/utils/changelog.ts	test/changelog.test.ts	counted
coding-agent/utils/git.ts	test/git-ssh-url.test.ts	counted
coding-agent/core/tools/path-utils.ts	test/path-utils.test.ts	counted
tui/keys.ts	test/keys.test.ts	counted
tui/utils.ts	test/truncate-to-width.test.ts, test/regression-regional-indicator-width.test.ts	counted
coding-agent/utils/mime.ts	test/image-process.test.ts	mixed — excluded (2 of 3 cases test processImage/original)
coding-agent/utils/version-check.ts	test/version-check.test.ts	mixed — excluded (fetch cases mock the original; not separable per-file)
coding-agent/core/export-html/ansi-to-html.ts	test/export-html-whitespace.test.ts	mixed — excluded (asset-grep + tool-renderer/original dominate)
coding-agent/core/tools/truncate.ts	—	no dedicated test (exercised via tools.test.ts, subject is the read/edit factory/original)
coding-agent/core/tools/edit-diff.ts	—	no dedicated test (edit-tool tests' subject is the edit tool/original)

CLI conformance (black-box, against the pidgin binary)

• CLI conformance: 15/15 pass against target/release/pidgin (pass delta ±0, fail delta ±0).

File	Passing	Failing	Skipped
packages/coding-agent/test/session-file-invalid.test.ts	1	0	0
packages/coding-agent/test/session-id-readonly.test.ts	7	0	0
packages/coding-agent/test/startup-session-name.test.ts	2	0	0
packages/coding-agent/test/stdout-cleanliness.test.ts	5	0	0

The four repointed coding-agent CLI test files spawn the compiled pidgin binary via $PIDGIN_BIN instead of pi's own cli.ts. This is a separate signal from the module smoke table and is never folded into the per-package Native count.