Merge branch 'main' into feat/prepack-dependency-gate

Author: ScriptedAlchemy

.changeset/369-mcp-tasks.md

@@ -0,0 +1,6 @@
+---
+"agent-bundle": patch
+"@agent-bundle/runtime": patch
+---
+
+Serve task-augmented MCP tool calls (the MCP `2025-11-25` Tasks utility) from generated route servers: a tool route that declares `config.execution.taskSupport: 'optional' | 'required'` (validated as `AB4836`, advertised in `tools/list`) answers a `tools/call` carrying `params.task` with a `CreateTaskResult` while the render continues behind the task; `tasks/get` reports status and the last render progress, `tasks/result` returns the same `CallToolResult` an ordinary call produces, `tasks/cancel` interrupts the render, and `tasks/list` lists the session's tasks. Clients that never ask for a task see no change; a server whose tools never opted in advertises no `tasks` capability. The Workbench MCP page runs a tool as a task, polls it, fetches its result, and cancels it; `agent-bundle/test`'s `openInMemoryMcpServer` client drives the same lifecycle. The host capability tables gain an `mcp.tasks` row recording whether each pinned host issues task-augmented calls. `@agent-bundle/runtime`'s operation-based `createRscMcpServer` is unchanged — no `tasks` capability, ordinary processing — and its README now records that instead of the lifted deferral (#550)

.changeset/457-lineage-tree.md

@@ -0,0 +1,5 @@
+---
+"@agent-bundle/runtime": patch
+---
+
+Expose the live agent tree to routes: `(await agent()).lineage.value.tree` — `{ siblings, children, roots }` of `AgentLineagePeer` (`{ conversation, depth, parent?, startedAt, subagent?, resolution }`) — lists every other live conversation under the request's root (any depth, the root itself included for a subagent), the request's live children, and the other live root conversations (on Cursor, only those seen in the same `workspace_roots`), read at resolve time from the same lineage registry that placed the request on every surface it feeds (event routes, generated MCP tool calls correlated through a hook window or a Codex `_meta`). Nothing is invented: stopped conversations are not listed, each peer carries the registry's own `resolution` for its placement, and the tree is absent when the registry did not place the request (a standalone hook, or a `_meta` naming a thread the registry never saw start). New exported types `AgentLineageTree` and `AgentLineagePeer`; `AgentLineage.tree` is optional, so existing readers and injected `context.lineage` fixtures are unchanged. (#544)

.changeset/459-provider-request-context.md

@@ -0,0 +1,6 @@
+---
+"@agent-bundle/runtime": patch
+"agent-bundle": patch
+---
+
+Hand conventional `src/providers/<name>` factories the request they run for: `AgentProviderContext` (`agent-bundle`) gains `host`, `session`, `workspace`, and `lineage` beside `plugin` — the same observed axes the route reads on `await agent()`, provenance and the lineage's live `tree` included — plus read-only `state` (`lifetime`, `read()`) and `notices` (`inbox()`, `published()`) views of the mounted handles; `dispatch`, `publish`, and `acknowledge` stay route-only, and `agent()`/`useAgent()` inside a factory throw `outside-invocation`. New exported types `AgentProviderObserved`, `AgentProviderHostIdentity`, `AgentProviderSessionIdentity`, `AgentProviderWorkspaceIdentity`, `AgentProviderLineage`, `AgentProviderLineageTree`, `AgentProviderLineagePeer`, `AgentProviderLineageSubagent`, `AgentProviderLineageResolution`, `AgentProviderStateHandle`, `AgentProviderStateSnapshot`, `AgentProviderNoticesHandle`, `AgentProviderNotice`, `AgentProviderNoticeState`, `AgentProviderNoticeRecipient`, `AgentProviderNoticePublisher`, `AgentProviderNoticeAttempt`, `AgentProviderNoticeWithholding`; `AgentProviderObservedPluginRoot` is now `AgentProviderObserved<AgentProviderPluginRoot>`. Every generated request scope (Flight worker, rendered CLI/script worker, plain routed CLI) and the `agent-bundle/test` harness now run providers as the request's own resolver — after `runAgentRequest` freezes the identity axes and opens the notice lease, before the route. `@agent-bundle/runtime`: `runAgentRequest` accepts `providers` as an `AgentProviderResolver` `(request: AgentProviderRequest) => values` beside the plain record; new exported types `AgentProviderRequest`, `AgentProviderResolver`, `AgentProviderStateHandle`, `AgentProviderNoticesHandle`. Existing factories that destructure `{ invocation, plugin, signal }` are unchanged. (#459)

.changeset/466-canonical-event-payload.md

@@ -0,0 +1,5 @@
+---
+"agent-bundle": minor
+---
+
+Give every event route a canonical, per-family `canonical.payload` beside the raw `native` envelope (`AgentEventRouteProps<E>`): the fields at least two hosts report — `toolName`, `toolInput`, `toolUseId`, `toolResponse`, `sessionId` (Claude and Codex `session_id`, Cursor `conversation_id`), `transcriptPath`, `cwd`, `model`, `permissionMode`, `agentId`/`agentType`, `agentTranscriptPath`, `prompt`, `reason`, `source`, `trigger`, `error`/`isInterrupt`, `lastAssistantMessage`, and `reentry` (Claude and Codex `stop_hook_active`, Cursor `loop_count > 0`) — each delivered as `{ value, nativeKey }` naming the host key it was read from, and absent when the host did not send it, never fabricated. Cursor's `tool_output` JSON string is parsed into `toolResponse` (kept as the string when it is not valid JSON); the Claude-only `model-switch/before` and `model-switch/after` families admitted by the 2.1.260 re-pin carry `fromModel`, `toModel`, `requestedModel`, and `source`. Where a family is also a config hook handler event, `canonical.payload` uses the same field names as `HookEvent<E>` (`reentry` is the one renaming of `stopHookActive`). Type a route to its family (`AgentEventRouteProps<'tool/after'>`) and `payload` narrows to that family's fields, in the route and in the generated `.agent-bundle/routes.d.ts` that `renderRoute` reads; `AgentEventCanonicalIdentity<E>` gains the same parameter. The per-family table ships as `agentEventPayloadFields`, the per-host key table as `agentEventPayloadNativeKeys` (with `AgentEventPayload`, `AgentEventPayloadField`, `AgentEventPayloadFieldName`, `AgentEventPayloadNativeKey`, and `agentEventPayloadFieldKinds`), and each pinned capability table mirrors its host's mapping under `hooks.eventRoutes.<event>.payload`, so the generated events reference documents field × host → native key per family. `agent-bundle/test` gains `createEventRouteInput(event, native, { host })`, which validates a host envelope and builds the `{ canonical, native }` input the harness takes, payload included; the Workbench Lifecycles view lists the mapped payload beside the canonical identity. `native` is unchanged, `idempotencyKey` still hashes only the envelope, and a route typed with the bare `AgentEventRouteProps` keeps working with every field optional. Breaking for one shape of consumer code: `payload` is a required property of `AgentEventCanonicalIdentity`, so a test or harness that constructs the identity by hand no longer compiles until it adds one — build the input with `createEventRouteInput` instead. (#466)

.changeset/468-plugin-root-request-axis.md

@@ -0,0 +1,6 @@
+---
+'@agent-bundle/runtime': patch
+'agent-bundle': patch
+---
+
+Expose the resolved plugin root on the request context: `(await agent()).plugin` is an observed `{ root, stateRoot }` — `source: 'native'` from an expanded `AGENT_BUNDLE_PLUGIN_ROOT`, `'derived'` from the shell's fallback (the artifact root, or `$PWD/.agent-bundle` for the npm bin) — and conventional providers receive the same value as `plugin` beside `invocation` and `signal` (`AgentProviderContext.plugin`). Every generated shell (MCP entry and Flight worker, routed CLI executable and render worker, hook wrappers) now resolves the anchor once through the new `resolvePluginRoot` export of `@agent-bundle/runtime` and mounts its SQLite state, notice ledger, and lineage journal on that one `stateRoot`, so `plugin.stateRoot` is the directory they use by construction; an unexpanded `${…}` token is treated as unset and reported once on stderr instead of being joined into a path. `renderRoute`, `invokeCli`, `runScript`, and `openInMemoryMcpServer` publish the axis the same way and accept `context.plugin`; `createGeneratedRouteMcpServer` takes `pluginRoot`. `AGENT_REQUEST_STORE_VERSION` is 4. Fixes #468. (#532)

.changeset/469-operator-env-installed-packs.md

@@ -0,0 +1,5 @@
+---
+'agent-bundle': patch
+---
+
+Give installed packs the operator `.env` layer that only `agent-bundle mcp run` had: every emitted stdio MCP entry with a factory default export (a self-connecting entry keeps its byte-identical body; `agent-bundle/launch-env` is aliased into every stdio entry, so such an entry imports and applies `applyOperatorEnv` itself if it wants the layer), every hook wrapper that runs plugin code, and the artifact CLI `bin/<name>.mjs` read `<plugin root>/.env` then `.env.local` at launch (the plugin root is the expanded `AGENT_BUNDLE_PLUGIN_ROOT`, else the shell's parent directory), filling only variables the host did not set — host environment and manifest `env` win, `.env.local` beats `.env`, values are never logged, a missing file costs nothing, an unreadable one is skipped. `AGENT_BUNDLE_ENV_FILE` names the file(s) to read instead (platform path delimiter; `none` disables the layer), and `mcp run` hands `--env-file` / `--no-env` down through it so a rehearsal and an installed pack read the same files. The loader is the new plain-Node `agent-bundle/launch-env` module (`applyOperatorEnv`, `parseOperatorEnv`), inlined into the shells without Effect. `agent-bundle doctor` reports whether an installed copy carries `.env` / `.env.local` and how many variables each declares (`AB7331`, informational; a warning when the file cannot be read), never a name or a value. The npm package bin reads no pack file. Fixes #469. (#538)

.changeset/README.md

@@ -13,6 +13,7 @@ versions; `changeset publish` (when enabled) publishes the result.
 | ----------------------- | ------------------------------ | ---------------------------------------- |
 | `agent-bundle`          | `packages/agent-bundle`        | publishable                              |
 | `@agent-bundle/runtime` | `packages/rsc-runtime`         | publishable                              |
+| `rsc-markdown-stream`   | `packages/rsc-markdown-stream` | publishable                              |
 | `create-agent-bundle`   | `packages/create-agent-bundle` | publishable                              |
 | `agent-bundle-workbench`| `packages/workbench`           | private, ignored                         |
 | `@agent-bundle-example/*`, `@agent-bundle/rsc-agent-runtime-demo` | `examples/*` | private, ignored |

.changeset/effect-filesystem-phase2-dev-server.md

@@ -0,0 +1,5 @@
+---
+"agent-bundle": patch
+---
+
+Keep `agent-bundle dev` — the Workbench dev server, its host installs, MCP sessions and probes, hook / host-discovery / native / script playgrounds, skill documents, evals, and asset serving — behaving exactly as before while every service's file reads, temporary directories, and removals run on one platform runtime that `startDevServer` creates and the session's `close` releases after the last service has closed. Diagnostics, routes, and responses are unchanged. Two lifetimes are now explicit: an MCP session's plugin-data directory lives exactly as long as the session (removed when the session closes, or when an open fails before the session exists), and a script playground workspace that cannot be removed is still reported in the run result's `cleanupFailures` instead of replacing the script's outcome. `agent-bundle --version` / `--help` keep loading no Effect module. (#551)

.changeset/jsx-markdown-content.md

@@ -0,0 +1,19 @@
+---
+"@agent-bundle/runtime": patch
+"rsc-markdown-stream": patch
+"agent-bundle": patch
+---
+
+Add the async `MarkdownContent` component and the `renderToMarkdown` /
+`renderToMarkdownStream` exports to `@agent-bundle/runtime`, so routes author
+rich Markdown blocks — headings, lists, GFM tables, task lists, nested async
+components, escaped text — as JSX lowered into `Agent.Markdown` instead of
+hand-concatenated strings. The renderer behind them, `rsc-markdown-stream`, is
+now a package of this repository and is published from it (it was previously
+only installable from its git URL), so `@agent-bundle/runtime` depends on it
+by version. `agent-bundle build` now follows symlinked (workspace) dependencies
+transitively when attributing bundle provenance, resolving each one the way
+Node does, so a project whose linked dependency links another package —
+including one hoisted to an ancestor `node_modules` — no longer fails with
+`AB5000`, and a dependency that links back onto the project never hides the
+project's own sources from provenance. (#344)

AGENTS.md

@@ -95,8 +95,9 @@
 ## Changesets
 
 - Every PR that changes a publishable package (`packages/agent-bundle`,
-  `packages/rsc-runtime`, `packages/create-agent-bundle` — anything except
-  `tests/**`) must include exactly one changeset: `pnpm changeset` or a
+  `packages/rsc-runtime`, `packages/rsc-markdown-stream`,
+  `packages/create-agent-bundle` — anything except `tests/**`) must include
+  exactly one changeset: `pnpm changeset` or a
   hand-written `.changeset/<slug>.md`. Private packages (`packages/workbench`,
   `examples/*`, `website`) are ignored and never named in a changeset.
 - Pre-1.0 semver: `minor` = breaking, `patch` = everything else (features