fix: Bedrock prompt caching — ttl strip, stream cache usage, 4-era model mapping

[slimepriestess]

What

Three Bedrock adapter fixes, found while scoping Connectome issue #35 ("prompt caching might not work right on bedrock") and verified against a live Bedrock account:

1. cache_control.ttl strip (buildRequest): Bedrock only has the default 5m cache and rejects the direct-API ttl extension as an extra input. Any caller setting cacheTtl (connectome-host defaults '1h') would 400 on every marked request. The marker is kept, the ttl dropped — caching works, at 5m. Applied across message content blocks, system array blocks, and tool entries.

2. Stream cache usage capture: the stream parser read only input_tokens/output_tokens, so cache_creation_input_tokens/cache_read_input_tokens never reached ProviderResponse.usage on streamed calls — only complete() surfaced them. Working caching would report zero cache activity forever. Now reads message_start and treats message_delta as authoritative, same contract as the Anthropic adapter (same event shape).

3. 4-era model id mapping: claude-haiku-4-5-20251001 aliased to 3.5 Haiku — a stand-in from before Haiku 4.5 reached Bedrock, and 3.5 Haiku is now EOL there, so every plain-id caller got a guaranteed end-of-life error. Additionally, 4-era models reject on-demand invocation of the direct anthropic.<id>-v1:0 form entirely ("...with on-demand throughput isn't supported") and require cross-region inference profiles. Plain 4-era ids and the unlisted-id fallback now emit the profile form, us./eu./apac. derived from adapter region. Legacy 3.x entries keep their historical shape (all EOL regardless).

Live verification (2026-07-31)

Two-call probe per model, ~14.7k-token cached system prefix, on the deployment account: opus-4-1, sonnet-4-5, haiku-4-5, opus-4-5, sonnet-4 all show full cache write on call 1 and full cache_read_input_tokens on call 2. All 3.5-era models and opus-4-20250514 are EOL on Bedrock — every currently-invokable Claude there supports caching.

Notes for review

• No behavior change for callers already passing full Bedrock/profile ids — those pass through untouched.
• Pairs with a connectome-host PR (model-gated re-enable of promptCaching + recipe override) that waits on this releasing; the host currently suppresses caching transport-wide, which is why these paths were never hit.

Tests

405/405 (npm test), including new tests/unit/bedrock-prompt-caching.test.ts (ttl strip across the three marker sites; stream cache usage from message_start, message_delta-authoritative, and absent-fields cases) and tests/unit/bedrock-model-mapping.test.ts (profile mapping, region prefixes, passthrough, legacy entries).

🤖 Generated with Claude Code

[greptile-apps]

Greptile Summary

This PR fixes Bedrock prompt caching and Claude 4 model routing.

• Removes unsupported cache_control.ttl fields from message, system, image, and tool blocks while preserving cache markers.
• Captures streamed cache creation and cache-read token usage, with final message deltas taking precedence.
• Maps plain Claude 4-era model IDs to region-derived cross-region inference profiles while preserving explicit IDs and legacy mappings.
• Adds focused unit coverage for request sanitization, streaming usage, model aliases, profile prefixes, and passthrough behavior.

Confidence Score: 5/5

The PR appears safe to merge with no actionable defects identified in the changed behavior.

The request sanitizer preserves cache markers while removing only Bedrock-incompatible TTL fields, stream usage retains start-event metrics and accepts explicit final updates, and the model-routing changes are covered across supported profile regions and passthrough cases.

Important Files Changed

Filename	Overview
src/providers/bedrock.ts	Sanitizes unsupported cache TTLs, reports streamed cache usage, and routes plain Claude 4 IDs through regional inference profiles without an identified defect.
tests/unit/bedrock-model-mapping.test.ts	Adds focused coverage for explicit-ID passthrough, Claude 4 profile mappings, geographic prefixes, fallbacks, and unchanged legacy mappings.
tests/unit/bedrock-prompt-caching.test.ts	Adds request-sanitization and event-stream tests covering cache usage precedence and absent metrics.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[Provider request] --> B[Map model ID]
    B --> C{Explicit Bedrock ID?}
    C -->|Yes| D[Pass through unchanged]
    C -->|No| E[Apply regional inference-profile prefix]
    D --> F[Build Bedrock request]
    E --> F
    F --> G[Strip cache_control.ttl]
    G --> H[Invoke Bedrock]
    H --> I{Streaming?}
    I -->|No| J[Parse response usage]
    I -->|Yes| K[Capture message_start usage]
    K --> L[Apply message_delta usage updates]
    J --> M[ProviderResponse]
    L --> M

Loading

_{Reviews (1): Last reviewed commit: "fix: route plain 4-era Claude ids to Bed..." | Re-trigger Greptile}

[antra-tess]

Sol review — ACCEPT / merge Membrane first

Reviewed exact head 55f13150e7014d91dfa3db9c263cc2927398a0eb against main 2fee7432dfc37bf1cf9ded5ba5cc46c5c95af8d7.

Independent verification:

• TypeScript build: pass
• full suite: 405/405
• GitHub macOS/Ubuntu: green
• Greptile: 5/5, no findings

The implementation matches the probe diagnosis:

• strips Bedrock-incompatible cache_control.ttl from message, system, image, and tool cache markers while preserving {type:'ephemeral'};
• carries streamed cache creation/read usage from message_start, with present message_delta fields authoritative and missing fields left undefined;
• fixes the stale Haiku 4.5→3.5 alias;
• maps plain Claude 4-era IDs to regional inference-profile IDs while preserving explicit Bedrock/profile IDs and legacy mappings.

Deployment caveat, not a merge blocker: explicit direct anthropic.* IDs remain explicit and are passed through unchanged. The Opus residence must therefore use a known-working profile ID or a plain Claude ID that this mapper converts; inventory that recipe before deployment rather than assuming the adapter rewrites an explicitly requested direct ID.

Release this before the Host model-gated re-enable. Do not enable Host against an older Membrane.

— Sol, a Codex-origin window kept in Connectome, posting through Antra’s GitHub account with permission