> ## Documentation Index
> Fetch the complete documentation index at: https://lab.pollack.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# What's New

> Weekly release highlights across all Pollack AI Lab projects

## Week of June 29, 2026

### Agent Journal 1.6.0

**First-class capture primitives + a cost-metering correction.** Slice 1 of the cross-repo journal-capture contract ships the §4 primitives downstream repos import, alongside a fix that roughly halves a long-run cost-accounting error.

* New primitives: `PhaseCapture.stepCosts()`, `JournalSteps.fromEvents()`, a production fail-loud `RunRecorder`, and per-turn usage in the immutable log
* Cost-metering fix: the headline `LLMCallEvent.tokenUsage` is now the cost-bearing Σ-per-turn-by-type aggregate (incl. cache) instead of the last-`ResultMessage` snapshot, which under-counted \~2× on long runs — same field, corrected value, no schema change
* A per-file `schemaVersion` header on `events.jsonl` / `analysis.jsonl` (loaders skip it)
* Additive on the frozen capture contract — a 1.5.0 consumer keeps working; 482 tests green
* [Project](/projects/agent-journal) | [What's New](/docs/agent-journal/whats-new) | [Source](https://github.com/markpollack/agent-journal)

### AgentWorks BOM 1.13.0

**Repins Agent Journal to 1.6.0.** A focused, journal-only bump behind the usual gate.

* All three journal artifacts (`journal-core`, `claude-code-capture`, `gemini-cli-capture`) → 1.6.0; every other member unchanged
* Published behind the CI-enforced `bom-verification` gate (plain / Boot 3.5 / Boot 4), re-verified from Central post-publish
* [Version Table](/projects/agentworks-bom) | [Source](https://github.com/markpollack/agentworks)

***

## Week of June 15, 2026

### Agent Journal 1.5.0

**First-class Gemini CLI capture.** The new `gemini-cli-capture` module projects Gemini runs into the same portable trace + cost schema as Claude, so one analysis layer reads both runtimes — plus per-step cost attribution.

* New `gemini-cli-capture` module — cross-runtime capture into one trace + cost format
* Per-step cost attribution (`StepCostEvent`, `analysis.jsonl` sidecar); portable `TraceWriter` moved into `journal-core`
* [Project](/projects/agent-journal) | [What's New](/docs/agent-journal/whats-new) | [Source](https://github.com/markpollack/agent-journal)

### AgentWorks BOM 1.10.0 → 1.12.0

**CVE remediation locked in, then the Gemini capture cut.** Three BOM releases close out the Spring AI 2.0 security work and add the new journal member.

* 1.10.0 / 1.11.0 (Jun 15): repin the source-fixed members to lock in the CVE remediation — claude-code-sdk 1.4.0, workflow 0.10.0, Spring AI BOM 2.0.0 backstop
* 1.12.0 (Jun 17): repin Agent Journal to 1.5.0 and add `gemini-cli-capture` as a managed member — the first BOM release through the CI-gated `release.yml` (pre-publish gate + post-publish FINAL\_PROOF)
* [Version Table](/projects/agentworks-bom) | [Source](https://github.com/markpollack/agentworks)

***

## Week of June 8, 2026

### AgentWorks BOM 1.9.0

**Suite-wide security release: Spring AI 2.0 GA convergence, four CVEs cleared.** The BOM moves the suite onto Spring AI 2.0.0 GA and Spring Boot 4.0.7, and adds Jackson 3 and Reactor management so consumers resolve a vulnerability-free, converged dependency set — proven end-to-end against the published artifacts by the `bom-verification` gate.

* Clears Spring Boot CVE-2026-40976 (CRITICAL) and -40973 (HIGH), Spring AI CVE-2026-41712 (HIGH), and Jackson CVE-2026-29062 (HIGH)
* New managed import `tools.jackson:jackson-bom 3.1.4` — converges Spring AI 2.0's transitive Jackson 3 (its jsonschema-generator pulls `jackson-core 3.0.3`) up to a CVE-clear 3.1.4
* New managed `reactor-core 3.8.6` — reconciles claude-code-sdk's 3.8.5 with Spring AI 2.0's 3.8.6, a `RequireUpperBoundDeps` break the gate caught
* Preceded by 1.8.0 (Jun 11), which cleared the long-deferred ACP bump to 0.14.0
* [Version Table](/projects/agentworks-bom) | [Source](https://github.com/markpollack/agentworks)

### Agent Client 0.22.0

**Spring AI 2.0.0 GA + Spring Boot 4.0.7.** The upgrade behind BOM 1.9.0 — primarily a security and dependency release, with one user-facing behavior change.

* spring-ai 2.0.0-M2 → 2.0.0 GA and spring-boot 4.0.1 → 4.0.7, clearing the transitive Boot / Spring AI CVEs above
* Spring AI 2.0 GA moved to Jackson 3, so `agent-model`, `agent-claude`, `agent-codex`, and `agent-client-core` now declare their Jackson 2 usage explicitly (governed by `jackson-bom 2.21.2`)
* Default Gemini model upgraded to `gemini-3.5-flash`; override with `agent-client.gemini.model`
* [Project](/projects/agent-client) | [Source](https://github.com/markpollack/agent-client)

***

## Week of June 1, 2026

### AgentWorks BOM 1.6.0

**The `bom-verification` quality gate lands, with Jackson 2.21.2 convergence.** The largest coordinated release of the week — nearly every member moved — and the first to publish behind an automated gate that proves consumers get a single, converged dependency set.

* New `bom-verification/` harness: plain, Spring Boot 3.5, and Spring Boot 4 consumer scenarios with a `JacksonConvergenceTest` and the enforcer `RequireUpperBoundDeps` rule; releases now publish only behind a green gate
* Imports `jackson-bom 2.21.2`, converging every member and shielding consumers from stray third-party Jackson transitives
* The gate is load-bearing — re-running it against 1.5.0 fails on the exact Jackson divergence 1.6.0 fixes
* Bookended by 1.5.0 (Jun 4, a managed test stack — Mockito / JUnit / AssertJ / Byte Buddy) and 1.7.0 (Jun 7, agent-client 0.21.0)
* Coordinated bump: claude-code-sdk 1.3.0, agent-client / agent-starter 0.20.0, agent-workflow 0.9.0, agent-judge 0.12.0, agent-memory 0.2.0, agent-journal 1.4.0, agent-hooks 0.6.3, agent-sandbox 0.9.3, agent-bench 0.4.0, experiment-core 0.5.0
* [Version Table](/projects/agentworks-bom) | [Source](https://github.com/markpollack/agentworks)

### Agent Client 0.21.0

**Portable reasoning-effort control across providers.** A new provider-agnostic `effort` option (low / medium / high), with provider-native overrides — and the Provider Parity TCK now actually runs (it had been reporting green without ever executing).

* `AgentOptions.getEffort()` — the cross-provider low/medium/high intersection, covered by the TCK
* Provider-native overrides win when set: Claude `ClaudeAgentOptions.effort` (→ `--effort`), Codex `CodexAgentOptions.reasoningEffort` up to `xhigh`
* 0.20.0 (Jun 6) shipped first: Claude trace content modes — `traceContentMode(FULL | TRUNCATED | LENGTHS)` — and automatic session-transcript archival, powered by agent-journal's TraceWriter v2
* [Portable Options](/docs/agent-client/reference/portable-options) | [Source](https://github.com/markpollack/agent-client)

### Claude Agent SDK 1.3.0

**Aligned to Claude CLI 2.1.162, with a lossless `rawJson` escape hatch.** A compatibility-and-hardening release that pins the wire contract and exposes wire fields the typed API doesn't model yet.

* `ParsedMessage.RegularMessage` gains a `rawJson` component carrying the exact stdout line each message was parsed from — backward compatible
* All six `HookInput` subtypes now ignore unknown properties, fixing a hook hang under CLI 2.1.162's added PostToolUse fields
* New `WireFixtureTest` pins parsing against real 2.1.162 stream-json fixtures
* [Project](/projects/claude-agent-sdk) | [Source](https://github.com/markpollack/claude-agent-sdk-java)

### Agent Workflow 0.9.0

**A tool-agnostic `Question` / `Option` callback record.** The `AgentCallback.onQuestion` SPI no longer leaks a tool-library type, and the judge integration completes its move to the `io.github.markpollack` namespace.

* New `workflow.callback.Question` record (with nested `Option`) owned by workflow-api, replacing the leaked `AskUserQuestionTool.Question`; the `spring-ai-agent-utils` dependency drops out of workflow-api
* Judge integration migrated to `io.github.markpollack.judge` (agent-judge 0.12.0) across `JudgeGate`, `TieredGate`, and the jury loops
* [Agent Loop](/docs/agent-workflow/agent-loop) | [Source](https://github.com/markpollack/agent-workflow)

***

## Week of May 26, 2026

### Agent Client 0.19.0

**JSONL trace file support.** `ClaudeAgentModel` now writes durable trace files during agent execution — tool calls, thinking blocks, text, and result metrics, flushed per event. Set `traceDir` on the builder or `agent-client.claude.trace-dir` in Spring Boot config. Each invocation writes one uniquely-named JSONL file, and the path is exposed via `providerFields.get("tracePath")`.

* `TraceTarget` record pairs run ID with trace file path
* Fail-fast on invalid trace directory (before client setup)
* Full UUID in filenames for collision safety
* [Claude Reference](/docs/agent-client/reference/claude-reference#trace-files) | [Source](https://github.com/markpollack/agent-client)

### Agent Workflow 0.8.0

**Trace file capture through the workflow journal.** `AgentClientStep` propagates trace paths from the agent model into `StepTransition` and journal events. Every tool call in a multi-step workflow is now traceable back to a specific JSONL file.

* `ExecutionResult(text, tracePath)` — new return type from `AgentClient.executeForResult()`
* `AgentContext.TRACE_PATH` — well-known context key, cleared per step
* `JdbcTraceRecorder` — `trace_path` column in `step_transitions` table
* [What's New](/docs/agent-workflow/whats-new) | [Trace Capture Guide](/docs/agent-workflow/trace-capture) | [Source](https://github.com/markpollack/agent-workflow)

### ACP Java SDK 0.12.0

**7 new protocol methods.** Session lifecycle (list, close, resume) is now stable. Elicitation, fork, and config options land as unstable with the new `@UnstableAcpApi` marker.

* [Details](/projects/acp-java-sdk) | [Source](https://github.com/markpollack/acp-java-sdk)

### AgentWorks BOM 1.2.0

Coordinated version bump: agent-client 0.19.0, agent-workflow 0.8.0, agent-journal 1.2.0, ACP Java SDK 0.12.0, and all other managed artifacts.

* [Version Table](/projects/agentworks-bom) | [Source](https://github.com/markpollack/agentworks)

***

## Week of May 19, 2026

### Agent Workflow 0.7.0

**Managed Agents as a step runtime.** `ManagedAgentStep` delegates workflow steps to Anthropic's Managed Agents API. The workflow graph stays in charge — specific steps run in Anthropic's cloud sandbox.

* [What's New](/docs/agent-workflow/whats-new#070) | [Source](https://github.com/markpollack/agent-workflow)

### Agent Client 0.18.0

**Migration to markpollack org.** Package rename from `org.springaicommunity.agents` to `io.github.markpollack`, BSL license, standalone POM.

* [Migration Guide](/docs/migration/spring-ai-community-to-markpollack) | [Source](https://github.com/markpollack/agent-client)