Commit 3f39dac

## [Unreleased]

## [5.0.0-beta.90] - 2026-09-24

### Changed

**A session that reattaches to another host generation keeps its memory.** Your memory identity was derived from the directory the host process happened to be started in, not from the session's own workspace. One shared host serves sessions from many projects, so a host ensured from somewhere else handed its own identity to every session that reattached to it: the session was told `memory identity conflict: session is bound to <workspace>-<hash>, but config resolved server-<hash>`, and its memory tools went away while the workspace had not moved at all. Identity now comes from the session's own working directory, and a reattach that still disagrees rebinds to the identity recorded in the session and notes it in the log instead of stopping. The error is kept for the case it was written for: you pointed `memory.agent` at a different identity yourself. ([#8556](https://github.com/code-yeongyu/oh-my-openagent/issues/8556))

## [5.0.0-beta.80] - 2026-09-20

### Changed

**RPC socket host: session-event credit on queue acceptance, 30 s dead-peer stall budget, observable cut notice.** (senpi [#1774](https://github.com/code-yeongyu/senpi/issues/1774))

### OmO

**Startup work that nobody waits for no longer runs before the first paint.** Three pieces of the omo-senpi plugin ran on the path the user waits through before the prompt appears, and none of them is observable before the first turn: the init-deep advisor spawned `git rev-parse` and stat'd the onboarding marker from its `session_start` handler just to decide it was ineligible (12.4 ms warm, 20.6 ms cold), the OmO-native session telemetry read the model inventory, loaded omo config twice and built its PostHog client on the same dispatch path (1.8 ms warm, 14.2 ms on the first session of a UTC day), and the LSP component built its mutation formatter - an omo-config read plus a formatter-marker scan - while it registered, for a step only a `tool_result` can reach. The advisor and the telemetry capture now run on the first post-paint edge (`input` / `before_agent_start`, or a 750 ms backstop), the formatter is constructed on the first mutation tool result, and the legacy telemetry product config no longer reads the package manifest at module scope. Tool, command, flag and hook registration is untouched: the engine's tables are complete before the first prompt exactly as before. Measured through a real senpi launch with the plugin (15 interleaved ABBA rounds): `interactiveMode.init` 97 ms to 84 ms median, 83 ms to 69 ms min; the plugin's `factory` row is unchanged at 37 ms. ([#8412](https://github.com/code-yeongyu/oh-my-openagent/issues/8412))

The changelog selection fix above lives in the engine and ships with this release. Also included: a cold-start fix for the RPC host, a goal-contract correction so a harness goal advises the loop instead of gating it, and Windows fixes for thread-socket discovery and a DAG race.

## [5.0.0-beta.1] - 2026-08-09

- 13db09a1a Merge pull request #6676 from code-yeongyu/release/v5.0.0-beta.1-source-state