v26.330.3

v26.330.3

Cerebellum Recovery Guidance

• New AssessTurnRecovery gRPC RPC — the Cerebellum receives full recovery context (progress ledger, checkpoints, browser state, partial content) and returns wait, retry, or stop with a tailored modelMessage for the Cerebrum
• Repetition detection: if the last 6 tool actions repeat the same 1-2 patterns, the Cerebellum returns stop to prevent infinite retry loops
• wait action for first-time stalls under 90s in waiting_model phase — avoids false nudges when the model is thinking
• stop action for repetitive completion retries — aborts early instead of exhausting the retry budget
• Fallback to local TypeScript logic if the Cerebellum is offline or the gRPC call fails

Recovery Architecture

• Both stall and completion retries now go through assessTurnRecovery()
• The Cerebellum's modelMessage replaces the hardcoded resume context
• operatorMessage shown to the user, diagnosis and nextStep logged
• streamDeferredUntil pauses the watchdog when the Cerebellum says wait

Docker Image

The Cerebellum Docker image needs a rebuild to include the new AssessTurnRecovery handler. Run bash scripts/publish-cerebellum.sh to push the updated image.

v26.330.2

v26.330.2

Browser Progress Ledger

• Structured progress tracking across retry attempts — each tool result records action, summary, URL, tab, and state-changing flag
• Progress entries survive retries via TurnContinuityState, giving the retry context a detailed record of what was already accomplished
• Read-only tools (browserGetText, readFile, etc.) are automatically classified as non-state-changing
• Ledger capped at 20 entries with smart eviction (read-only entries evicted first)

Task Checkpoints

• New task_checkpoint internal tool — Cerebrum marks milestones ("profile continuity checked", "engagement pass done") with evidence
• Checkpoints are upserted by step name and included in retry resume context
• Checkpoints are preserved across retries — the model sees a structured list of what's already been verified

Browser State Tracking

• BrowserStateSnapshot tracks current URL, active tab, and tab list
• Updated from result.metadata.resume on each browser tool execution
• Included in retry context: "Current URL: https://x.com/home", "Active tab: 134851321"

Structured Browser Tool Results

• All browser tools now return ToolExecutionValue with metadata.resume containing action summaries
• Enables the progress ledger to record rich context without parsing output strings

Unified Resume Context

• Single buildRetryContextMessage() handles both stall and completion retries
• Resume context includes: browser state, checkpoints, progress ledger entries ([done]/[seen]/[error]), partial text

v26.330.1

v26.330.1

Retry Message Cleanup

• Failed attempt tool/system messages are excluded from retry prompts using ID-based filtering (not positional cutoff)
• Failed attempt messages are deleted from the conversation store on ALL exit paths: success, exhaustion, error, and Cerebellum disconnect
• Successful retry's own tool results are preserved in the conversation store
• Stable message ordering via ORDER BY timestamp, rowid

Resume Context (No Repeat)

• Retried turns see only: user message + transient resume summary + nudge
• Failed attempt's raw tool call history is excluded — model cannot "follow the template" and repeat steps
• Resume prompt explicitly states: "tool call history has been removed, do NOT repeat these steps"

Task Completion Guard

• task_complete / task_blocked internal tools for structured turn endings
• Completion guard retries independently from stall retries (separate budgets)
• Completion retry context includes tool summaries and partial text

Stream Finish Metadata

• Provider captures finishReason, stepFinishReasons, chunkCount, toolCallCount, stepCount
• Completion guard uses this to detect turns ending on tool-calls without final answer

v26.329.28

v26.329.28

Completion Retry Context

• Completion retries now include transient resume context (recent tool results, partial text, finish reason)
• Matches the stall retry context pattern — injected into retry attempt, NOT persisted to conversation

Non-Debug TUI

• Default view shows only message content (user=blue, cerebrum=green) without role labels or tool outputs
• --debug flag restores the full verbose view with role labels, tool details, and streaming header
• Cleaner experience for normal users

Event-Driven Stall Detection

• Removed 250ms polling interval for stall state
• TUI now updates stall indicator purely from events (cerebrum:stall, watchdog, tool:end, etc.)

Presentation Utilities

• formatLocalTimestamp() — timezone-aware formatting via Intl.DateTimeFormat
• filterTranscriptMessages() — filters to user+cerebrum for non-debug display

v26.329.27

v26.329.27

Stall Retry Continuity

• Retried turns now receive a transient resume context with recent tool results and partial text from the interrupted attempt
• Prevents retried turns from restarting from scratch — instructs the model to resume from the most advanced confirmed state
• Resume context is NOT persisted to conversation history (transient injection only)

Adaptive Stall Threshold

• waiting_model phase uses 3x base threshold (models legitimately take longer to think)
• Each retry adds 1x base threshold (linear backoff)
• waiting_tool phase keeps the original threshold (tools should complete promptly)
• Reduces false-positive nudges during genuine model reasoning

v26.329.26

v26.329.26

Task Completion Guard

• Internal task_complete and task_blocked tools — Cerebrum must call one before ending a tool-driven turn
• If the turn ends without a completion signal, the orchestrator retries with a prompt to continue
• Internal tools are hidden from conversation history and TUI

Split Retry Accounting

• Stall retries (maxNudgeRetries) and completion retries (maxCompletionRetries) now have independent budgets
• Total max attempts = 1 + stall retries + completion retries (default 5)
• Each retry tracks its cause (stall or completion) for clear diagnostics
• Safety limit prevents infinite loops

Stream Finish Metadata

• Provider captures finishReason, rawFinishReason, stepFinishReasons, chunkCount, textChars, toolCallCount, stepCount
• Completion guard uses this to detect turns that ended on tool-calls without a final answer

v26.329.25

v26.329.25

• Watchdog diagnostics in TUI — shows stage/message trace (stalled, nudge_requested, abort_issued, retry_started, retry_recovered, retry_failed, teardown_timeout)
• Abort teardown fallback for stuck iterators
• Stream diagnostic events with phase context

v26.329.24

v26.329.24

Consolidates watchdog recovery fixes and shared abort helpers:

• Watchdog recovery for stuck tool calls — abort signal propagates through provider → tool-runtime → browser backend. Stuck browser operations (navigate, getPageText, etc.) are now interrupted by the watchdog.
• Stream phase diagnostics — StreamPhase (idle/waiting_model/waiting_tool) with active tool name in stall events
• Shared abort helpers — raceWithAbort, throwIfAborted, createAbortError extracted to @cereworker/core (no more duplication)
• Iterator teardown — iterator.return() called on abort to stop background tool execution
• AbortSignal threading — wired through orchestrator → tool-runtime → browser tools → extension relay

v26.329.22

v26.329.22

Republish of v26.329.21 — the gateway package tarball was missing on npm CDN due to a partial publish failure. All packages verified and installable.

Includes all stream watchdog end-to-end fixes:

• gRPC 5s deadline on verifyToolResult
• Cerebellum-gated nudge with error/null fallback
• Cerebellum required: sendMessage blocks when offline
• Mid-stream disconnect aborts cleanly
• Provider abort race with iterator.return() teardown
• Single abort listener (no per-part leak)
• Duplicate error suppression + discovery error surfacing
• Stream phase tracking (idle/waiting_model/waiting_tool)
• Abort signal wired through sub-agent manager

v26.329.21

Stream Watchdog End-to-End Fix

This release consolidates all stream watchdog fixes from v26.329.17–v26.329.21 into a working end-to-end stall recovery system.

What was broken

The Cerebrum would silently stop mid-response during complex browser automation workflows. The watchdog detected the stall but couldn't recover because:

1. gRPC calls to the Cerebellum hung forever (no deadline)
2. The AI SDK's async iterator didn't break on abort during multi-step tool execution
3. Error events were duplicated or swallowed

What's fixed

• gRPC 5s deadline on verifyToolResult — prevents hung Cerebellum calls
• Cerebellum authority — nudge decision goes through Cerebellum; errors/timeouts trigger nudge as fallback
• Cerebellum required — sendMessage blocks when Cerebellum is offline
• Mid-stream disconnect — aborts the stream cleanly with clear error message
• Provider abort race — Promise.race with iterator.return() teardown ensures the abort actually breaks the stream and stops background tool execution
• Single abort listener — no per-part listener leak on long streams
• TUI error surfacing — offline errors show in chat UI for both normal messages and first-run discovery
• Duplicate error suppression — intentional aborts (nudge + disconnect) only emit one error