Skip to content

v2026.11.5-1

Choose a tag to compare

@github-actions github-actions released this 05 Nov 20:11
· 2292 commits to main since this release

@code-yeongyu/senpi

Breaking Changes

Added

  • Typed Windows presence guard bindings for the public socket-secret reservation dependency slice (desktop#4199).
  • Support SIWC and legacy ChatGPT accounts in the same credential pool without changing rotation policy or health storage (part of #5003).
  • Show ChatGPT account routes and plan permission in terminal and shared RPC account listings.
  • Added hold_host on observing RPC reads: an observer that sends hold_host: true keeps the host from idling out until it sends hold_host: false or disconnects, so a remote-control daemon can hold the host while a remote viewer is attached; plain observing reads still never count as activity (#5628).
  • Groundwork for offline pause and resume in the headless engine: a session whose owning machine or hand goes offline pauses and resumes exactly once on reconnect, and a re-attaching client sees it; nothing selects it yet, so existing hosts are unchanged (#5652)
  • Two-sided extensions on a split session: an extension that registers loop events and tool_call/tool_result without runs declarations keeps one state. Loop-side means any event the placement table puts on the loop or split side. On the hand, each undeclared tool handler of such an extension is forwarded to the loop that owns the call (the call's nested extension-events plane), and the reply is applied as a local handler's result would be. Its runs: "hand" handlers stay on the hand in their registration position. The forward fails closed when the loop's process lacks that extension's handlers. Builtins and fused/headless sessions are unchanged (#5664).
  • Split sessions present a hand's extension UI on the attached client: widgets, notices, status, titles and editor text that an extension produces on the hand while a call runs go in order to the loop that owns the call (its ui-presentation plane) and are presented there exactly as on a fused session, before the call ends. Presentation never changes a call's result or error: when the loop cannot take one, it is presented on the hand's own UI and reported on the runner's error channel. Dialogs stay off the port, and the hand still reports no UI to extensions. Fused sessions are unchanged (#5740).
  • A hand serving a loop that runs elsewhere can be the home of two-sided extensions (twoSidedHome: "hand" on the hand's scopes): their tool hooks stay with the tool on the hand instead of being forwarded to the loop, and the hand runs their loop hooks and the user's shell hooks on the same runtime (createLooplessHand), bound to the loop session's id for that call only. A runner without a session now runs extensions' before_agent_start. The default ("loop") is unchanged (#5741).
  • A same-account presenter can enroll on an RPC host's MCP presenter control endpoint with account_enroll (an issuer attestation bound to its installation key, then a signed host transcript) and extend it with account_reattest; every refusal is a typed account_refused reason. An account principal never becomes the host's launch owner (#4249)
  • createAgentSession({ executionPipelineWitness }): a session's execution-pipeline receipts go to the sink it is given, never into its entries. Without one, SENPI_EXECUTION_PIPELINE_WITNESS=<path> appends one JSON line { pipeline, toolCallId } per call to that file (#5771).
  • RPC: get_state reports the running turn as currentTurnId, and a prompt steer or follow-up, steer or follow_up may name it as expectedTurnId. The name is checked where the input enters the turn's queue, so a steer that names a turn which is not running then is refused with errorCode: "turn_conflict" instead of landing in another turn (#5779).
  • An embedding distribution can register a PresenterAccountAuthority from @code-yeongyu/senpi/presenter-authority before it loads the CLI; main() seals the registration first, so no extension or later code can supply or replace it (#4249)
  • Added the threads.chainCountLimit setting: the number of hops a thread chain may make before the next thread call is refused; unset (the default) means unlimited (#5882).

Changed

  • Sessions run their tool calls on the prepared execution pipeline (batch coordinator + prepared executor) by default, with today's events, results and order (#5733).
  • Sessions run model-issued tool calls on the prepared execution pipeline, with the same events, results and order as before. SENPI_EXECUTION_PIPELINE=legacy (a test and QA setting) is no longer accepted (#5738).
  • A Cursor exec call with invalid arguments now gets tool_execution_start / tool_execution_end events with the error result, as a model-issued call does; before, the bridge returned the error without lifecycle events (#5738, #5780).
  • The tool_result hook no longer runs for a Cursor exec call that preflight blocked or that lost its owning run, matching model-issued calls; it still runs for every call that executed (#5738, #5780).
  • An account-enrolled MCP presenter's mcp_presenter_assign and mcp_presenter_acquire now require the host to still be signed into its account and its attestation to be current; a sign-out, an account switch or reaching attestedUntil revokes it with a typed account_refused reason (#4249)
  • RPC: every extension_ui_progress now gets one reply keyed by its id (success, or question_already_resolved / unknown_extension_ui_request), after the question_updated it causes, so a client forwarding records can settle it like any other; RpcClient consumes the replies to its own unawaited sends quietly (#5769)
  • web_search, webfetch and the ask-user tools (request_user_input, ask_user_question) declare runs: "brain": on a split session they run where the agent loop runs, never on the hand. Fused sessions are unchanged (#5882).

Fixed

  • A fallback to a smaller-window model now reduces the context when the provider-reported usage does not fit, instead of skipping the reduction on a lower estimate and then refusing the switch; a refusal that remains is shown with the refused model's window and requirement (#3447).
  • A Claude subscription account whose Claude CLI login is signed out now shows as needing /login anthropic-subscription (in /claude-account list and the turn error) instead of being re-checked every 15 minutes forever. A turn whose Claude CLI refresh is refused marks the account the same way at once. It clears by itself once the login works again, and other accounts keep serving (#5350)
  • An RPC host never binds through a symlink planted at its socket path (macOS followed it and listened at the link's target), and its ownership checks read a symlink there as a foreign entry instead of the socket it names (#5319)
  • /resume no longer lists the session you are in while it has no conversation, so the first row and the default selection are the newest session you can actually resume (#5727)
  • A command or prompt submitted while the TUI is still starting now runs once startup finishes, instead of staying typed in the editor behind "Startup is still in progress" until a second Enter (#5731)
  • Tool calls that Cursor resolves over its exec channel run on the prepared execution pipeline, like model-issued calls, and write their execution-pipeline witness receipt (#5738, #5780).
  • /resume of a session that another live process is running no longer starts a second run of it: it asks whether to cancel or take over (the other process is asked to exit first) (#5744)
  • Prepared execution pipeline: an assistant message that repeats one tool call ID now runs each of those calls as its own call, as the legacy scheduler does, instead of refusing the whole batch (#5799).
  • A relocation request whose app leaves the session while the request is being delivered is now refused with app_disconnected under the prepared pipeline too, instead of ending as Tool execution aborted (#5805)
  • Under the prepared execution pipeline, a permission "Allow once" now covers the call that runs: the executor carries the approval to the frozen copy of the call's arguments (only while the input is unchanged), so the tool no longer asks a second time. The approved parent of a native file monitor is carried too, so the parent-swap check at registration runs under prepared as it does under the default pipeline (#5795)
  • A relocation request refused because its app left the session now always reaches the agent as that refusal (app_disconnected), never as Tool execution aborted: closing the session waits for the relocation tool calls to receive their answers before it aborts the run (#5884)
  • Sign in with ChatGPT compaction uses Responses v2 when available and otherwise local summarization, without calling the unsupported compact endpoint (#5101).
  • Closing a JS eval kernel whose WebView launch is still in flight now resolves only after that launch's Chrome has retired (#5945)
  • On Windows, sessions on a supervised RPC host register their control endpoint again, so messages sent to them start a turn instead of staying queued (#5985)
  • Retryable provider failures behind a gateway (a 429 JSON body, a 5xx HTML page) no longer print the raw body as a chat row on every retry: the turn shows one retry status with a short reason, and one final notice if every retry fails, with the raw bodies on expand (#6014)
  • An RPC host started with a short idle window no longer exits before the ensure that started it has connected, when its startup takes longer than that window (#6020)

Removed

@code-yeongyu/senpi-ai

Breaking Changes

Added

  • Add opt-in SIWC request shaping to the public Responses adapter, including supported tool loading and HTTP history replay (#5021). Provider registration remains separate.
  • Route saved SIWC ChatGPT accounts through public Responses, retain native web search, and keep legacy account requests unchanged (part of #5003).
  • Continue with ChatGPT signs in through SIWC on the existing ChatGPT subscription provider; signing in again with the same verified registration replaces its slot instead of duplicating it (part of #5003).
  • Offer Continue with ChatGPT alongside legacy sign-in through the shared subscription provider.

Changed

Fixed

  • Sign in with ChatGPT models are ineligible for prompt-cache prewarm even when their catalog enables explicit caching (#5101).

Removed

@code-yeongyu/senpi-agent-core

Breaking Changes

Added

  • The conformance entry exports the AdmissionLine type, so a writer outside the package can take a store's single admission line (#5652)
  • createRemoteToolHost({ filePlanes }): the planes a pending device file operation may host on its own lane, so a split session can serve a two-sided extension's forwarded tool events for a file operation. Absent, the default, a file operation hosts no nested request, as before. A hand passes each file operation its own nested requester (HandScope.file takes it) (#5664).
  • DeviceFiles.run: the memory component's git and lsof calls run on the device through an allowlisted operation, with no shell, a bounded environment, the hand's tool_call handlers, and the git-activity lock held as fd 3 (#5696)
  • Main-entry exports of createPreparedToolHost (the prepared pipeline as an in-process ToolHost), canonicalJson and toStoreValue (#5738).
  • AgentLoopConfig.batchStoreFactory / AgentOptions.batchStoreFactory, openMemoryBatchStore (a fresh in-memory store per tool batch) and BatchStoreMissingError (#5738).
  • AgentLoopConfig.afterToolFreeze / AgentOptions.afterToolFreeze (runs once per call after it is frozen, with the arguments object beforeToolCall saw and the frozen copy the tool receives), noPreflightCarry and PreflightCarryMissingError (#5795).
  • A hand fences a moved session's old owner with an epoch floor (src/execution/remote-epoch.ts, brain/hand S35, relay design 4.4): createHandEpochFence({ ownerOf, durable? }) keeps one floor per execution scope that only the scope's owner raises, one epoch at a time, that never lowers and remembers its owner; with durable it persists on the hand's store on the ledger's admission line. createRemoteExecutor({ fence }) answers floor frames and refuses an invoke below the floor (epoch_fenced) before anything of it is recorded, opened or run, also inside each step of a call's walk. createRemoteToolHost({ epoch }) dispatches a scope lane's calls only after the hand acked exactly its epoch; a late or duplicate floor answer cannot change a decided lane, and after a hand restart a pending call is re-sent only once the re-sent floor is acked. Without a ledger the hand re-checks the epoch right before the tool runs, and an acknowledgement is fenced inside its admission slot. A second link's executor on a live hand no longer re-runs restart recovery. Adds the EPOCH, MULTI-EPOCH and KILL-EPOCH conformance cases (#5739).
  • Loop-hook link for a loop that runs away from the user's machine: createLoopHookHost sends each hooked loop event to the user's machine (serveLoopHooks) and resolves with its result before the provider request. While that machine is offline the call waits and resumes exactly once. A call the previous process never started is sent again with the same ID; one it started is never re-run and ends LoopHookRestartedError. KILL-HOOK-01..05 conformance cases are registered and exported from ./conformance (#5741).
  • AgentLoopConfig.commitStep (optional): called once per model step after its provider tool results are emitted and before the loop ends the turn or dispatches tools; a rejection ends the run with that error. Plus the ModelEgress route types, StepCommitError and isStepCommitError (#5857)
  • AgentHarnessOptions.onExecutionPipeline: every harness tool body now runs on the in-process prepared ToolHost, and this sink receives one { pipeline: "prepared", toolCallId } receipt per call, as Agent.onExecutionPipeline does (#5859).
  • Added the session plane (src/execution/plane-session.ts): createSplitToolHost() runs each call to a tool declared runs: "brain" on the loop's own pipeline (createLoopToolHost()) and every other call on the hand, keeping the batch's mode and the model's source order; serveSessionPlane() / createSessionPlaneClient() let a hand-run tool ask the brain that owns its call for a loop-side tool call, and refuse any tool that is not loop-side. Added the thread chain count (src/execution/chain-count.ts): a chain call carries its hop count and is refused with chain_count_limit only past a count the user set; unset means unlimited, and there is no cost limit (#5882).

Changed

  • An agent loop with no toolHost runs model-issued tool batches on the prepared pipeline (batch coordinator + prepared executor) instead of its own scheduler. The package entry's Agent, agentLoop, agentLoopContinue, runAgentLoop and runAgentLoopContinue supply openMemoryBatchStore unless a factory is passed, so existing callers change nothing. A batch that repeats a tool call ID runs each block as its own call, as the removed scheduler did (#5796). The package entry also supplies noPreflightCarry unless an afterToolFreeze is passed. The loop core alone refuses a tool batch with BatchStoreMissingError when no factory is configured and with PreflightCarryMissingError when no afterToolFreeze is configured (#5738, #5795).

Fixed

Removed

  • The agent loop's own tool scheduler; every tool batch without a toolHost runs on the prepared pipeline (#5738).

@code-yeongyu/senpi-codemode

Breaking Changes

Added

Changed

Fixed

Removed

@code-yeongyu/senpi-tui

Breaking Changes

Added

Changed

Fixed

Removed