@bastani/atomic 0.9.39 → 0.9.40-alpha.1

package/CHANGELOG.md

@@ -2,6 +2,34 @@
 
 ## [Unreleased]
 
+## [0.9.40-alpha.1] - 2026-10-20
+
+### Breaking Changes
+
+- Binding or switching a session no longer replays deliveries that a previous process left unfinished. SDK hosts call `session.startPendingRecovery()` and RPC clients send `start_pending_recovery` once they show the session. Until then, interrupted work stays pending ([#3760](https://github.com/bastani-inc/atomic/issues/3760)).
+- Removed the bundled `subagent` tool's `kill` action. Replace `subagent({ action: "kill", id })` with `subagent({ action: "quit", id })`, which stops the child gracefully and keeps it resumable. Calls with `action: "kill"` or the earlier `action: "interrupt"` are rejected with an error that names `quit` as the replacement ([#3760](https://github.com/bastani-inc/atomic/issues/3760)).
+
+### Added
+
+- Added `AgentSession.startPendingRecovery()`, the RPC `start_pending_recovery` command, and the `session_recovery_restore` and `session_recovery_replay` extension events. Extensions can now restore background work after a restart, decline to replay deliveries whose peer is gone, and defer deliveries they cannot decide on yet to the next resume without starting a turn ([#3760](https://github.com/bastani-inc/atomic/issues/3760)).
+- Added the `typesafe/jev-1.14.0` classifier. Both direct TypeSafe classifiers, `typesafe/jev-latest` and `typesafe/jev-1.14.0`, now accept up to 262K tokens and PNG or JPEG images, for example through `models.classify()` in codemode. See [TypeSafe Jev](docs/providers.md#typesafe-jev).
+- The bundled `subagent` tool now has workflow-style run control: `pause`, `resume` (with an optional `message`), and `quit`, with `all: true` for `pause` and `quit`. Pausing or quitting a parallel child still queued behind its concurrency limit keeps it from starting until you resume it. Paused, quit, interrupted, completed, and failed children resume from their own saved session under the same id and Intercom name, including children of a session you closed or switched away from (exit or crash, `/new`, `/resume` to another session, `/fork`). Subagents of a workflow stage that is paused or quit, including by Atomic exiting or a session switch, stay resumable too, and resuming the run after a restart restores them with the stage. With a `message`, a paused, quit, or interrupted child receives it once and then continues where it left off, and a completed or failed child receives it as one fresh prompt; without one, the child is asked to continue where it left off. Children cancelled for good (owner or batch cancellation, a `/tasks` stop), every subagent of a workflow stage that finished or whose run ended or was cancelled, including ones you paused or quit, and children reported gone after a restart are terminal, and `resume` on them returns an error that says why ([#3760](https://github.com/bastani-inc/atomic/issues/3760)).
+- Background subagent launch cards show the child's latest state (running, paused, quit, interrupted, completed, failed, or terminal), also after you reopen the session (`/resume` or `atomic --session <id>`) or view a workflow stage's transcript. Expanding a card shows the end of the child's saved transcript. A resumed child continues in its original card, and replayed `pause`, `resume`, and `quit` calls render as their results ([#3760](https://github.com/bastani-inc/atomic/issues/3760)).
+- Added per-process storage selection. `ATOMIC_STORAGE`, `ATOMIC_SESSIONS_STORAGE`, `ATOMIC_WORKFLOWS_JOURNAL`, `--storage '<json>'`, and `--storage-file <path>` take the same JSON selections as the storage settings, adapters and `secrets` included, so a host can start `atomic --mode rpc` children on their own store without editing settings. A process without them keeps using its settings, and `/resume` in each lists only its own store. Invalid or conflicting selections stop startup with a `StorageConfigError` naming the variable or flag. See [Selecting storage per process](docs/storage.md#selecting-storage-per-process) ([#3783](https://github.com/bastani-inc/atomic/issues/3783)).
+- Added multi-node routing: Intercom brokers on different machines that select the same shared `storage` (in settings, `ATOMIC_STORAGE`, or `--storage`) now deliver messages to each other through that database. A message to a session on another machine arrives exactly like a local one, with the same sender identity, replies and asks, group rules, supervisor routes and save-before-delivered receipts. A send to a session that is offline, including one that quit cleanly in the last 7 days, is reported as queued for an offline session and delivered when it reconnects on any broker. Brokers coordinate only through the shared database and a polled wake-up signal; there is no broker-to-broker network connection. Without a shared `storage` selection, Intercom behaves as before.
+
+### Fixed
+
+- Resuming a session with `atomic --session <id>` or `/resume` now shows the transcript before any recovery work runs, instead of leaving the screen blank while hidden agent turns replay messages from before the restart. Recovery runs afterwards and is visible in the transcript. A "Replaying N interrupted deliveries" line precedes any replayed turn, and a status line summarizes the result ([#3760](https://github.com/bastani-inc/atomic/issues/3760)).
+- Resuming no longer starts agent turns on its own for stale queued messages or background task results. Results that an earlier process never delivered appear as cards without starting a turn. Messages from peers that no longer exist are not replayed; one notice tells you and the agent which peers are gone. Messages from a background subagent that was relaunched but has not reconnected yet are deferred instead, and the notice says it is not gone ([#3760](https://github.com/bastani-inc/atomic/issues/3760)).
+- Intercom messages sent to a busy session are no longer lost when that session is killed or crashes before it goes idle: the recipient saves each message when it arrives, the sender sees `delivered` only after that save, and resuming the session with `atomic --session <id>` or `/resume` shows each saved message once. Like interrupted deliveries, a saved message starts a turn only when its sender is still connected; one from a gone sender is added once, in full, to the session history without starting a turn, and the recovery notice says the sender is gone; one whose sender cannot be checked is deferred to the next resume ([#3761](https://github.com/bastani-inc/atomic/issues/3761)).
+- Fixed Windows rejecting Bash and PowerShell executables installed as Windows Store app aliases, such as those in `%LOCALAPPDATA%\Microsoft\WindowsApps`, as missing. Shell discovery and a configured `shellPath` now accept them ([upstream commit](https://github.com/earendil-works/pi/commit/28bea006be4adb950dcb4f86f9ee5be857bf74ff)).
+- Fixed Windows PowerShell discovery when PowerShell is not on `PATH`: Atomic now also checks `%ProgramFiles%\PowerShell\7\pwsh.exe` and `%SystemRoot%\System32\WindowsPowerShell\v1.0\powershell.exe`. The legacy WSL `bash.exe` launcher is now recognized under the actual Windows directory, including installs on another drive or a renamed Windows directory ([#9490](https://github.com/earendil-works/pi/issues/9490)).
+- JSONL sessions no longer stall when many sessions share a directory. Each append now rereads only its own session header instead of every session header in the directory ([#3780](https://github.com/bastani-inc/atomic/pull/3780) by [@darionco](https://github.com/darionco)).
+- Opening or loading a session in a directory with many JSONL sessions no longer reads every session header, including the first lookup in a process and sessions created by another process.
+- A resumed session no longer splits an extension tool call into an arguments-only card and a separate result card when the extension saved its own entries between them, and tool cards in an attached workflow stage chat now redraw when their renderer updates them ([#3760](https://github.com/bastani-inc/atomic/issues/3760)).
+- Extension tool cards, including subagent cards, now use their extension renderers and current state when a session is opened with `atomic --session <id>`, instead of showing raw arguments. They also redraw as soon as their state changes, so an expanded subagent card shows the child's saved transcript and a resumed child updates in its original card without waiting for another screen redraw ([#3760](https://github.com/bastani-inc/atomic/issues/3760)).
+
 ## [0.9.39] - 2026-10-19
 
 ### Added

package/dist/builtin/intercom/CHANGELOG.md

@@ -4,6 +4,20 @@ All notable changes to the `pi-intercom` extension will be documented in this fi
 
 ## [Unreleased]
 
+## [0.9.40-alpha.1] - 2026-10-20
+
+### Added
+
+- Added multi-node routing: Intercom brokers on different machines that select the same shared `storage` (in settings, `ATOMIC_STORAGE`, or `--storage`) now deliver messages to each other through that database. A message to a session on another machine arrives exactly like a local one, with the same sender identity, replies and asks, group rules, supervisor routes and save-before-delivered receipts. A send to a session that is offline, including one that quit cleanly in the last 7 days, is reported as queued for an offline session and delivered when it reconnects on any broker. Brokers coordinate only through the shared database and a polled wake-up signal; there is no broker-to-broker network connection. Without a shared `storage` selection, Intercom behaves as before.
+
+### Fixed
+
+- On resume, Intercom messages interrupted by an exit are replayed only if their sender is connected again. Messages from a peer that is gone no longer start a hidden turn: they appear once as history in a visible notice that tells the agent the peer is gone. A resumed session that used Intercom re-registers right away. A resumed session or restored background subagent also keeps the Intercom session ID it had before the restart (unless a live session holds it), so peers can reach it by its name, by that ID, or by its Atomic session ID instead of getting `Session not found` ([#3760](https://github.com/bastani-inc/atomic/issues/3760)).
+- Intercom now keeps an interrupted workflow run's targets with the restarted session that launched the run, so messages sent before `/workflow resume` are queued instead of failing with `Session not found`. If another session resumes the run, it takes over those targets, and the restarted session stops answering for the run ([#3760](https://github.com/bastani-inc/atomic/issues/3760)).
+- A `send` or `ask` to a background subagent that its restarted owner has not restored yet now waits for it, whoever sends it: the session that owns it, a sibling subagent, a workflow stage, or another session in the subagent's group. This applies whether the message is addressed by the subagent's name, its Atomic session ID, or the Intercom session ID it had before the restart. The wait starts when the owning session starts, not only once its recovery has restored the subagent, and recovery takes the wait over without dropping or repeating a message. For the subagents of an interrupted workflow run's stages, the session that launched the run holds them while the run waits for `/workflow resume`, and the resumed stage takes the hold over, so a message sent in that interval is delivered once after the subagent reconnects. Waiting messages are delivered once each, in the order each sender sent them, as soon as the subagent registers again, and each sender gets its usual confirmation. A message still waiting after 20 seconds gets a "still being restored, try again" error; the owning session is told instead when the subagent was relaunched but has not reconnected. If the subagent did not survive the restart, waiting and later messages get an explicit "did not survive the restart and is gone" error instead of `Session not found`; senders other than the owner get it while the owning session stays connected to Intercom. If another Atomic session resumes the run, the messages waiting for its subagents get an error saying so. If the owning session disconnects, waiting messages get `Session not found`; a wait that nothing restores also ends after 60 seconds. Messages to a subagent while no Atomic process of its owner is running are not queued and get `Session not found` ([#3760](https://github.com/bastani-inc/atomic/issues/3760)).
+- If Intercom cannot be reached when a session resumes, its interrupted Intercom messages are no longer discarded or attributed to peers that are gone. They stay pending without starting a turn, one notice says that Intercom was unreachable and the deliveries were deferred, and the next resume decides about them again. The same applies to messages from a background subagent that was relaunched but has not re-registered with Intercom yet: they are deferred, not dropped, and the subagent is not reported as gone ([#3760](https://github.com/bastani-inc/atomic/issues/3760)).
+- Intercom messages that reach a busy session are no longer lost if that session is killed or crashes before it goes idle. The recipient saves each message to its session when it arrives, and the sender sees `delivered` only after that save. Resuming the session with `atomic --session <id>` or `/resume` shows every saved message it never showed or otherwise handled (a reply already returned to a waiting `ask` or a message it refused is not shown again), once each and in arrival order, after the transcript appears. Saved messages pass the same sender check as interrupted Intercom deliveries: only a message whose sender is connected again starts a turn; a message from a sender that is gone is added once, in full, to the session history without starting a turn, and the recovery notice says the sender is gone. A message whose sender cannot be checked is deferred until the session is opened again: Intercom was unreachable, or the sender is a relaunched subagent that has not reconnected yet. When the recipient does not confirm the save in time, the sender gets an explicit failure saying the outcome is uncertain instead of a false `delivered`, and a recipient that reconnects before confirming receives the message again without showing it twice, and its answer to an `ask` still reaches the waiting sender. Recipients from earlier Atomic versions keep the previous behavior, and so does an Intercom broker from an earlier version that is still running after an upgrade: upgraded sessions keep working with it and still save each message on arrival. Sessions started with `--no-session` are still not saved ([#3761](https://github.com/bastani-inc/atomic/issues/3761)).
+
 ## [0.9.39] - 2026-10-19
 
 ### Added