@bastani/atomic 0.9.63 → 0.9.64

package/CHANGELOG.md

@@ -2,6 +2,30 @@
 
 ## [Unreleased]
 
+## [0.9.64] - 2026-11-02
+
+A quiet chat when workflows log or are quit in the terminal UI, and Claude sessions that survive a cross-provider fallback.
+
+- **Workflow console output stays out of the chat:** in the terminal UI, console output from a workflow's own code, including its abort listeners and the cancellations a quit causes, goes to a per-process workflow log instead of one chat line per message; the chat shows one line per run naming the log. Quitting a run with 120 pending logging steps used to print 120 lines into the chat; now it prints none ([#4176](https://github.com/bastani-inc/atomic/pull/4176))
+- **Claude keeps its signed thinking across a fallback:** after a fallback from another provider, the thinking of the latest Claude turn is sent back exactly as Claude produced it, so the guard that protects it no longer loses track after a reasoning-only turn or a server-side fallback marker, and signed thinking is no longer altered by surrogate sanitizing ([#4189](https://github.com/bastani-inc/atomic/pull/4189))
+
+### Fixed
+
+- In the terminal UI, console output from a workflow's own code no longer floods the chat with one status line per message. Output of runs a terminal-UI session starts or resumes goes to the process's workflow log under `~/.atomic/agent/workflow-logs/`, one line each tagged with its run and written as it happens, and the chat shows one line per run naming that file. Cancellations from quitting a run never produce that line, so quitting a run whose code logs each failed step no longer prints a line per pending step. Each run gets up to 8 MiB of the log and each log up to 64 MiB. Logs of exited processes not written for 14 days, or beyond 1 GiB in total, are deleted when the next terminal-UI session starts. If the log cannot be written, the chat gets one line saying so and later output is dropped. Runs started or resumed by another owner, and print, JSON and RPC modes, SDK hosts and pooled worker processes, still write console output to stderr. ([#4175](https://github.com/bastani-inc/atomic/issues/4175))
+- Fixed two causes of `400 invalid_request_error: ... thinking or redacted_thinking blocks in the latest assistant message cannot be modified` when Claude continues a session after a model fallback, for example a subagent launched on `github-copilot/gpt-6.1-sol` whose fallback chain moves it to Claude Opus 5.5. Atomic restores the exact bytes of Claude's reasoning before each request, but it gave up for the whole request when an earlier turn sent Claude nothing, such as a GPT turn that held only reasoning, or sent only a Claude server-side fallback marker. The latest Claude turn could then go out with its thinking changed by a `before_provider_request` extension hook. Separately, the Anthropic converter stripped unpaired surrogates from signed thinking; it now sends signed thinking unchanged. GPT reasoning is still never sent to Claude. The same error was also reported in a long Claude-only session with no fallback; that case has not been reproduced and is not covered by this fix ([#4185](https://github.com/bastani-inc/atomic/issues/4185))
+
+## [0.9.64-alpha.1] - 2026-11-02
+
+A quiet chat when workflows log or are quit in the terminal UI, and Claude sessions that survive a cross-provider fallback.
+
+- **Workflow console output stays out of the chat:** in the terminal UI, console output from a workflow's own code, including its abort listeners and the cancellations a quit causes, goes to a per-process workflow log instead of one chat line per message; the chat shows one line per run naming the log. Quitting a run with 120 pending logging steps used to print 120 lines into the chat; now it prints none ([#4176](https://github.com/bastani-inc/atomic/pull/4176))
+- **Claude keeps its signed thinking across a fallback:** after a fallback from another provider, the thinking of the latest Claude turn is sent back exactly as Claude produced it, so the guard that protects it no longer loses track after a reasoning-only turn or a server-side fallback marker, and signed thinking is no longer altered by surrogate sanitizing ([#4189](https://github.com/bastani-inc/atomic/pull/4189))
+
+### Fixed
+
+- In the terminal UI, console output from a workflow's own code no longer floods the chat with one status line per message. Output of runs a terminal-UI session starts or resumes goes to the process's workflow log under `~/.atomic/agent/workflow-logs/`, one line each tagged with its run and written as it happens, and the chat shows one line per run naming that file. Cancellations from quitting a run never produce that line, so quitting a run whose code logs each failed step no longer prints a line per pending step. Each run gets up to 8 MiB of the log and each log up to 64 MiB. Logs of exited processes not written for 14 days, or beyond 1 GiB in total, are deleted when the next terminal-UI session starts. If the log cannot be written, the chat gets one line saying so and later output is dropped. Runs started or resumed by another owner, and print, JSON and RPC modes, SDK hosts and pooled worker processes, still write console output to stderr. ([#4175](https://github.com/bastani-inc/atomic/issues/4175))
+- Fixed two causes of `400 invalid_request_error: ... thinking or redacted_thinking blocks in the latest assistant message cannot be modified` when Claude continues a session after a model fallback, for example a subagent launched on `github-copilot/gpt-6.1-sol` whose fallback chain moves it to Claude Opus 5.5. Atomic restores the exact bytes of Claude's reasoning before each request, but it gave up for the whole request when an earlier turn sent Claude nothing, such as a GPT turn that held only reasoning, or sent only a Claude server-side fallback marker. The latest Claude turn could then go out with its thinking changed by a `before_provider_request` extension hook. Separately, the Anthropic converter stripped unpaired surrogates from signed thinking; it now sends signed thinking unchanged. GPT reasoning is still never sent to Claude. The same error was also reported in a long Claude-only session with no fallback; that case has not been reproduced and is not covered by this fix ([#4185](https://github.com/bastani-inc/atomic/issues/4185))
+
 ## [0.9.63] - 2026-11-02
 
 Workflow journals that survive a stalled lock holder, and pooled roots that stay responsive with thousands of finished runs.

package/dist/builtin/intercom/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/intercom",
-  "version": "0.9.63",
+  "version": "0.9.64",
   "private": true,
   "description": "Atomic extension providing a private coordination channel between parent and child agent sessions. Fork of: https://github.com/nicobailon/pi-intercom",
   "contributors": [

package/dist/builtin/mcp/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/mcp",
-  "version": "0.9.63",
+  "version": "0.9.64",
   "private": true,
   "description": "Atomic's pi-compatible MCP integration for server connections, OAuth, tools, resources, and terminal management.",
   "contributors": [

package/dist/builtin/subagents/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/subagents",
-  "version": "0.9.63",
+  "version": "0.9.64",
   "private": true,
   "description": "Atomic extension for delegating tasks to subagents with parallel execution. Fork of: https://github.com/nicobailon/pi-subagents",
   "contributors": [

package/dist/builtin/web-access/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/web-access",
-  "version": "0.9.63",
+  "version": "0.9.64",
   "private": true,
   "description": "Atomic extension for web search, URL fetching, GitHub repo cloning, PDF/video extraction. Fork of: https://github.com/nicobailon/pi-web-access",
   "contributors": [

package/dist/builtin/workflows/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/workflows",
-  "version": "0.9.63",
+  "version": "0.9.64",
   "private": true,
   "description": "Atomic extension for multi-stage workflow authoring and execution.",
   "contributors": [