@plumbus/core 0.6.15 → 0.6.17

package/CHANGELOG.md

@@ -1,5 +1,33 @@
 # @plumbus/core changelog
 
+## 0.6.17 — 2026-08-16 — `plumbus e2e` server lifecycle hardening
+
+### Changed
+
+- **`AGENT_WIRING_VERSION` bumped to 14.** `plumbus init` references `@plumbus/voice/instructions/continuous-sessions.md` (talk-over re-queue, stitched transcripts, sentence chunker, opt-in backchannel). Pointers are inert when `@plumbus/voice` is not installed. Run **`plumbus init --patch`** on existing projects to refresh the managed wiring block.
+
+### Fixed
+
+- **Empty provider API-key placeholders no longer crash worker boot.** A dotenv line like `AI_ANTHROPIC_API_KEY=` (present but empty, often left alongside a model var) used to produce a provider slot that 0.6.16's boot-time key validation then rejected — `AI provider "anthropic" requires apiKey` before the worker could start. Env discovery now treats an empty key as "slot not configured" and skips it, matching ≤0.6.15's effective behavior. Keyless Bedrock slots are unaffected (separate region-keyed reader).
+
+- **`plumbus e2e` can no longer orphan the frontend dev server on interruption.** The auto-started `next dev` runs as a detached process group, and cleanup previously lived only in a `finally` block — Ctrl-C/SIGTERM on the CLI (the detached group never receives the terminal's signal) or a SIGKILL/crash left an immortal dev server compiling and watching until reboot (reproduced 2026-08-16). Now: SIGINT/SIGTERM/SIGHUP handlers kill the server group before exiting; a tiny detached watchdog reaps the group if the CLI vanishes without cleanup (SIGTERM, then SIGKILL after a 5 s grace; it self-exits when the group is gone); and vitest runs via async `spawn` instead of `execFileSync` so a blocked event loop can't defer the handlers. Normal shutdown is unchanged and verified sufficient: the whole server tree stays in the spawn group and exits within ~3 s of the group SIGTERM.
+- **Port preflight** — the command fails fast with cleanup guidance when the target port is already in use (orphaned previous run or a concurrent checkout), instead of silently running the suite against a foreign server and reporting a bogus green (reproduced 2026-08-16 with two checkouts sharing one port).
+- **Dev-server heap cap** — the server spawns with `NODE_OPTIONS=--max-old-space-size=4096` (skipped when the caller already sets a cap) plus `NEXT_TELEMETRY_DISABLED=1`. A runaway compile now fails the run loudly instead of pushing the machine into memory-pressure thrash.
+
+## 0.6.16 — 2026-08-12 — Amazon Bedrock provider + agent wiring v13
+
+### Added
+
+- **Amazon Bedrock provider slot** — `createProviderAdapter('bedrock', …)` dynamically loads optional peer `@plumbus/ai-bedrock` (`0.1.x`). Env discovery: `AI_BEDROCK_REGION` (or `AI_BEDROCK_ENABLED` + `AWS_REGION`), `AI_BEDROCK_MODEL`, `AI_BEDROCK_EMBEDDING_MODEL`, `AI_BEDROCK_PRICING_FILE`, timeouts/limits.
+- **Provider-supplied AI cost** — optional `cost` on `ProviderResponse` and stream `done` / `usage` events. `createAIService` prefers adapter cost when set (Bedrock package-owned pricing); otherwise uses `calculateModelCost` against the OpenAI/Anthropic catalog.
+- **`AIProviderConfig`** — added `region`, `pricingFilePath`, `embeddingModel`, `pricingCacheTtlMs` for Bedrock-style providers. `apiKey` stays a required `string`, so `config.ai.apiKey` still assigns to `string` in existing app bootstrap.
+- **`AIProviderSlotConfig`** — new exported type accepted by `createProviderAdapter`: identical to `AIProviderConfig` but with optional `apiKey`, so keyless providers (Bedrock uses the AWS credential chain) need no placeholder key. An `AIProviderConfig` remains assignable to it, so existing calls are unaffected. Env discovery sets `apiKey: ''` on the Bedrock slot.
+
+### Changed
+
+- **`AGENT_WIRING_VERSION` bumped to 13.** `plumbus init` references `@plumbus/ai-bedrock/instructions/pricing.md` (AWS Price List URLs, curl, normalize, ConfigMap) alongside framework + README. Pointers are inert when the package is not installed. Run **`plumbus init --patch`** on existing projects to refresh the managed wiring block.
+- **OpenAI SSE stream parser** — when a chunk carries both `content` and `finish_reason` (Bedrock Mantle / some OpenAI-compatible proxies), emit the content delta before `done` so streamed text is not dropped.
+
 ## 0.6.15 — 2026-08-10 — migrate rollback hardening + model pricing sync
 
 ### Fixed