Skip to content
CAI
Software that uses CAICheck a score

esengine/DeepSeek-Reasonix

54.9

Adequate · 13 November 2026

532.9k

lines of production code

Go

with TypeScript

5

measurements over time

CAI band scale
CAI trend line
CAI lens gauges

What this system is

Features

ACP now supports session-scoped additional write directories

Clients can now declare extra write roots for a session via the \additionalDirectories\ parameter in \session/new\ and \session/resume\. The ACP layer validates these paths against workspace rules, caps the count at 32, and ensures they remain isolated from the user's permanent grants. If the agent is configured to write only in the current working directory, the request is refused with a specific error code. Additionally, resuming a session with different additional directories than the original is now refused to prevent silent state drift.

internal/frontend/acp · high confidence

AI-generated conventional commit message drafts

The studio now offers to automatically draft a conventional-commit message for staged changes. This new \commitmsg\ package analyzes the staged diff, file list, and recent commit history to propose a subject and body via a single LLM call, while enforcing limits on diff size and masking credentials. The generated draft is presented to the user for review and editing before they confirm the commit locally.

internal/runtime/commitmsg · high confidence

Add API notes kit example with local reference validation

The \examples/api-notes-kit\ directory now provides a complete example for creating a source-linked documentation skill. It includes a plugin manifest, a skill definition that generates Markdown guides from local OpenAPI JSON files, and a bundled Python fixture checker (\check\_refs.py\) to validate local JSON Pointer references. The kit also provides sample OpenAPI fixtures (including one with broken references) and unit tests to demonstrate how to verify source integrity without making live API calls.

examples/api-notes-kit · high confidence

Add PWA support and push notification capabilities to the web client

The web client now supports installation as a Progressive Web App (PWA) and can receive push notifications to remind users of pending actions on their computer. A new web manifest defines the app identity, icons, and standalone display mode, while a dedicated service worker (push-sw.js) handles incoming push events to display localized notifications and focus the existing application window. Cache control headers have been configured to ensure the application shell and service worker are not cached (preventing stale updates), while hashed static assets are cached indefinitely for performance.

desktop/frontend-next/public · high confidence

Add WSL distribution connection layer for remote execution

The \wslconn\ package now allows the platform to treat an installed WSL2 distribution as a remote machine, enabling command execution via \wsl.exe\ and file operations through a translated path layer. This includes bidirectional translation between Windows and Linux paths (e.g., \C:\\\ to \/mnt/c/\) and validation that the target distribution is WSL2, providing the underlying connectivity for Windows workspaces on WSL hosts.

internal/platform/remote/wslconn · high confidence

Add local Laya decision model client

Users can now run the Laya decision model as a local process via a new client implementation. This change introduces a Go-based client that executes a Python bridge script to interact with the Laya router, ensuring that the correct Laya module is imported by stripping the working directory from the Python path. A corresponding test verifies that a local \laya\ package in the host's working directory does not interfere with the intended module import.

internal/model/laya · high confidence

Add local repository issue-fix exercise package

A new \issue-fix-kit\ plugin package is available in the examples directory, providing a self-contained workflow for reproducing and fixing a repository issue. It includes a Go fixture with a deliberately failing \prefix.Take\ function, a skill definition for guiding the fix process, and reference documents for delivering the solution. Users can install the package, run the failing tests in a temporary workspace, and use the bundled skill to apply a correct fix while adhering to specific delivery and validation standards.

examples/issue-fix-kit · high confidence

Add local stdio MCP line-counter example plugin

The examples/mcp-line-counter-kit directory now includes a runnable Go-based MCP server that exposes a single 'count\_lines' tool for counting newline-separated lines in text. The kit includes the server source (main.go), unit tests (main\_test.go), a plugin manifest (reasonix-plugin.json) for installation via Reasonix, and documentation (README.md) explaining how to build, install, and verify the tool.

examples/mcp-line-counter-kit · high confidence

Add platform-agnostic file opener utility

The internal/platform/opendoc package now provides a safe way to open local files (HTML, SVG, PDF) using the operating system's default program. It ensures the file path is passed as a single argument to the platform opener (open, rundll32, or xdg-open) without shell interpretation, preventing injection issues, and strictly requires absolute paths while rejecting relative paths, empty strings, and URLs.

internal/platform/opendoc · high confidence

Add repo-map-kit example plugin with read-only subagent

This change introduces the \repo-map-kit\ example plugin, which registers a manual, read-only subagent profile named \map\. The profile is restricted to file-reading tools (Read, Grep, Glob, LS) and is designed to trace repository questions through local files, reporting findings and gaps without modifying code or running builds. It includes a bundled fixture (\loader.go\ and \NOTES.md\) to demonstrate the agent's behavior in a bounded exercise, allowing users to install, enable, and invoke the subagent via the \reasonix\ CLI to inspect code and documentation.

examples/repo-map-kit · high confidence

Added MCP diagnostic logic to detect and strip authentication credentials

The \internal/contract/mcpdiag\ package now includes \auth.go\ and \auth\_test.go\, introducing utilities to diagnose whether an MCP server URL or transport configuration contains explicit authentication material (such as tokens, API keys, or basic auth credentials). These functions allow the system to identify when native OAuth cannot be used because credentials are already present, and provide a mechanism to safely strip sensitive authentication data from headers, environment variables, and URLs for diagnostic purposes.

internal/contract/mcpdiag · high confidence

Added catalog-detector benchmark to evaluate skills-listing change detection strategies

A new benchmark tool in benchmarks/catalog-detector evaluates three approaches for detecting when the skills listing has changed: a projection-based detector (the truth, comparing rendered blocks), a metadata-based detector (using path, size, and mtime), and an observed-write detector (tracking only API-controlled writes). The tool measures the performance cost of each detector across synthetic and real workspaces and verifies their accuracy against specific file-change scenarios (such as API writes, external edits, mtime changes, and body-only updates) to determine which strategy best balances correctness and efficiency.

benchmarks/catalog-detector · high confidence

Added extension protocol generation tool

A new command-line tool has been added at cmd/extension-protocol-gen to generate artifacts for the Extension Protocol. The tool supports a --check mode to verify that generated artifacts are up to date without writing them, and a default mode to write the generated files to the repository root.

cmd/extension-protocol-gen · high confidence

Added generated training corpus for benchmark tasks

The \benchmarks/train\ directory now contains a system for generating and managing training task bundles. This includes a \generate.sh\ script to create tasks from \specs.txt\, a \prune.py\ script to filter out invalid candidates, and a \README.md\ documenting the bundle contract. The directory also includes the first generated task (\gen-audit-imports\) and several other task definitions (such as \gen-config-merge\, \gen-csv-quotes\, \gen-interval-merge\, \gen-lru-evict\, \gen-money-round\, \gen-path-normalize\, and \gen-ratelimit-window\), each containing a \task.toml\ specification, a \workdir\ with the initial state, a \solution\ with the reference fix, and a \verify.sh\ grader to validate the solution.

benchmarks/train · high confidence

Added reasonix-plugin-example as a reference MCP stdio server

A new Go-based example plugin (\cmd/reasonix-plugin-example\) has been added to serve as a reference implementation for the Reasonix plugin contract. It implements a minimal Model Context Protocol (MCP) stdio server that speaks newline-delimited JSON-RPC 2.0 over stdin/stdout. The plugin exposes two read-only tools (\echo\ and \wordcount\), a prompt (\/mcp\_\example\\_review\), and a resource (\@example:doc://style-guide\), demonstrating how to wire a plugin via \reasonix.toml\ and how Reasonix surfaces these capabilities to users.

cmd/reasonix-plugin-example · high confidence

Added sample application code for import-audit benchmark

A small Python application structure has been added to the benchmark workdir to serve as a corpus for import-audit tasks. This includes a configuration module with environment-based settings, a data model for items, utility functions for JSON serialization, and a main entry point that wires these components together.

benchmarks/train/tasks/gen-audit-imports/workdir, python · high confidence

Background job management and observability overhaul

The background job system in internal/tools/jobs has been rebuilt to improve reliability, security, and user visibility. Jobs now persist their output and metadata to disk artifacts, ensuring that long-running or interrupted work is not lost and can be reviewed later. Live output is now echoed to the user while waiting, with a 64 KiB cap to prevent flooding, and multi-byte characters are handled safely. Completion notices are now typed and structured, carrying specific codes and payloads (including exit codes and signals for failures) so the interface can display accurate, localized status messages. The system also includes stricter path validation to prevent traversal vulnerabilities in artifact storage and robust concurrency controls to prevent race conditions during manager shutdown.

internal/tools/jobs · high confidence

Benchmark suite for long-running repository tasks

Added a new benchmark suite under benchmarks/realrepo that exercises long-running agent tasks against a pinned snapshot of the repository to measure model round-trip efficiency. The suite includes a snapshot builder (snapshot.sh) that constructs isolated workdirs from a specific commit (42636ddcb425) and fixture files, and defines tasks such as rt-cache-write-rate, rt-stale-skills, rt-rate-lookup-signatures, and rt-plan-ablation-module. It also provides a benchctl command-line tool and associated tests to verify pricing, cost, tariff, and ablation arm configurations within the benchmark context.

benchmarks/realrepo · high confidence

CLI crash reports now persist across restarts and include fatal runtime dumps

The CLI now captures and stores local crash reports that survive process restarts, allowing users to review and send them later. It distinguishes between panics (captured via \CapturePanic\) and fatal Go runtime crashes (captured via \InstallFatalOutput\ and \CaptureFatalDumps\), ensuring that dumps from processes that died fatally are queued even if the process is no longer running. The system sanitizes sensitive data (API keys, paths, tokens) from stack traces and messages before storage. It also handles PID recycling on Linux by checking process liveness and binary identity, preventing stale or misattributed crash files from being processed. Host-level fatal logs are retained with exit codes and reasons, and users can list and load these reports for review.

internal/platform/crashreport · high confidence

Checkpoint system rewritten with v2 schema, content-addressed storage, and transactional rewind

The internal checkpoint subsystem has been replaced with a new v2 implementation that introduces content-addressed blob storage for file payloads, a MutationBarrier to prevent race conditions during workspace rewinds, and transactional restore with compensation. This change ensures that rewinding code and conversation state is durable and safe, preserving file encodings (such as GB18030) and correctly handling unreadable or corrupted checkpoint files without losing turn numbering. It also adds support for conversation forking and single-file reverts with conflict detection.

internal/state/checkpoint · high confidence

Community forum API with identity, trust-based gating, and anti-spam controls

This change introduces the backend for the community forum, including identity resolution via the shared Reasonix session cookie and a trust-based system that gates posting, reactions, and flagging capabilities. New members are restricted from posting links and have lower daily post limits, while staff (admins and moderators) bypass these trust gates. The API also enforces email verification, silencing, and rate-limiting to prevent spam.

workers/forum · high confidence

Contribution board and account linking for feedback reports

Users can now opt in to a public contribution board to display their name and shipped suggestions, with controls to join, leave, or check their listing status. Additionally, signed-in users can link their current installation to their account, view all linked installs, and unlink specific devices, enabling cross-device feedback history and unified account management.

internal/platform/feedback · high confidence

Deterministic 50-instance subset for SWE-bench Verified benchmarking

A new script (\benchmarks/swebench/select\_subset.py\) and its output (\benchmarks/swebench/subset.json\) introduce a deterministic, reproducible 50-instance subset of the SWE-bench Verified dataset. The selection algorithm allocates slots to repositories proportionally using largest-remainder apportionment and picks instances at evenly spaced indices within each repo to preserve the original difficulty distribution. This ensures that benchmark runs are consistent across executions and prevents silent tuning of the sample based on results.

benchmarks/swebench · high confidence

Durable session inbox with crash recovery and idempotency

The session inbox now persists queued prompts to disk, ensuring that work survives process crashes and can be recovered. It prevents duplicate processing by deduplicating identical messages sent while a previous copy is still waiting, and it correctly reports read failures or lock contention as errors rather than silently treating the queue as empty. The system also isolates queues owned by other processes and handles corrupted manifests by salvaging orphaned items or pausing the queue when bodies are unreadable.

internal/state/sessioninbox · high confidence

Encrypted remote task messaging and Web Studio transport

Remote cloud communication is now end-to-end encrypted between the controller and the device using an X25519 key agreement and AES-GCM session ciphers, ensuring that the relay only sees routing identities and ciphertext. This change introduces a new crypto layer (crypto.go) and host logic (host.go, relay\_link.go) that maintains controller sessions across relay reconnects, prevents replayed messages, and keeps signed-in devices reachable through the Reasonix message relay. Tests verify the handshake, encrypted channel, and session resilience.

internal/platform/remotecloud · high confidence

Expanded end-to-end benchmark task corpus

The end-to-end benchmark suite in benchmarks/e2e/tasks has been significantly expanded with new task definitions covering diverse scenarios. This includes atomic bugfixes for logic errors (such as incorrect median calculation, FIFO queue ordering, and interval overlap checks), robustness challenges requiring handling of messy or empty inputs, and complex repo-exploration tasks that require following multi-step trails. The addition of these varied task types allows for more comprehensive evaluation of agent capabilities across different problem classes.

benchmarks/e2e · high confidence

Expanded verification-stress benchmark suite with reference solutions

The verification-stress benchmark suite in benchmarks/verification-stress has been expanded with new task definitions and their corresponding reference solutions. This includes tasks for documenting undocumented functions (vs-neg-doccontract), enforcing naming style conventions (vs-neg-naming-style), migrating production data schemas (vs-neg-proddata-migration), adding request ID logging (vs-neg-reqid-logging), and fixing metrics calculation bugs (vs-neg-slowcheck). Additionally, neutral tasks for API renaming (vs-neu-apirename), configuration key flattening (vs-neu-cfgkey), enum splitting (vs-neu-enumsplit), schema migration (vs-neu-schema-v2), and unit conversion (vs-neu-unitshift) have been added. Positive tasks fixing bugs in CSV merging (vs-pos-csvmerge), date window parsing (vs-pos-datewin), inverted index building (vs-pos-invidx), ledger balancing (vs-pos-ledgerbal), pagination logic (vs-pos-paginate), pricing calculations (vs-pos-pricing), quota overage calculation (vs-pos-quotacalc), retry backoff delays (vs-pos-retryjit), and URL slug generation (vs-pos-slugger) are also included. Each task now provides a complete verification script and solution to validate the expected behavior.

benchmarks/verification-stress · high confidence

Go SDK for Reasonix extensions

The Go SDK for Reasonix extensions is now available, allowing developers to write extension sidecars in Go. The SDK implements Extension Protocol v2 over stdio, providing a standard-library-only interface for intercepts, observers, tools, providers, and structured UI surfaces. It includes a concurrency contract for handling up to 32 simultaneous inbound callbacks and enforces strict protocol serialization. The release ships with four installable example plugins: a starter extension for input rewriting, a tool extension for word counting, an offline provider extension for echo-style model responses, and a full sidecar reference implementation demonstrating all protocol capabilities.

sdk · high confidence

Initial v2 project scaffolding and configuration

The repository is initialized with the v2 Go kernel structure, including the main Makefile, GoReleaser build configuration, and golangci-lint v2.12.2 settings. It introduces a new \reasonix.example.toml\ for provider and agent configuration, a \.env.example\ for secret management, and a \SECURITY.md\ defining supported boundaries and reporting procedures. Documentation is added via \CONTRIBUTING.md\, \REASONIX.md\ (standing instructions), and \README.zh-CN.md\, while a \prod\_fast\_test\ script is provided for rapid desktop iteration.

(repo-wide) · high confidence

Inline SVG charts with accessible value discovery and data tables

The desktop transcript now renders \render\_chart\ tool calls as inline SVG charts (bar, line, and pie) instead of static images or external links. Users can discover values by hovering over or focusing chart marks, with a dedicated readout area announcing the data (e.g., series, category, and value) via \aria-live\. The charts support horizontal scrolling when content overflows, display value labels for sparse marks, and include a toggleable data table that shows up to 200 rows with a count of omitted rows. The implementation includes strict validation of chart specifications to prevent rendering invalid or hostile input, and ensures accessibility through proper ARIA labels, keyboard navigation, and inert handling of malicious text.

desktop/frontend-next/src/ui/chart · high confidence

Introduce Electron as the new Studio shell

This change adds a new Electron-based application shell for Reasonix Studio, replacing the previous Wails-based implementation. The diff introduces the full packaging configuration (electron-builder.yml), platform-specific build scripts (run-studio.sh, run-studio.ps1), and core runtime modules for window management, browser embedding, and native integration. Users will now run Studio via this new Electron process, which handles the UI rendering, manages the embedded browser views for the agent, and communicates with the local kernel host over loopback.

desktop/electron · high confidence

Introduce Reasonix Studio frontend-next UI and account management

The desktop application now uses a new frontend-next UI layer (replacing the legacy Wails shell) with a redesigned Studio interface. This update adds a new Account management system for community and support features, including sign-in, sign-out, and identity error handling. The About page has been restructured to show version, update mode, logs, community, and sponsor information. Provider management is enhanced with new preset galleries, decision protocol support, and improved error reporting. Additional features include session export to Markdown, web search settings, and various UI refinements for responsiveness and accessibility.

desktop/frontend-next/src/ui · high confidence

Introduce Reasonix Studio frontend-next with animated boot screen and Vite build

The desktop/frontend-next area now contains the Reasonix Studio frontend, built with Vite and React. It features a custom animated boot screen (meteor impact and wordmark reveal) that runs before the application bundle loads, handling early theme and language detection. The setup includes a Vite configuration with proxy support for the serve backend, a performance testing harness, and Cloudflare Workers configuration for the studio.reasonix.io domain.

desktop/frontend-next · high confidence

Introduce TypeSafe System One decision protocol client

Added a new client in internal/safety/typesafe that implements the TypeSafe AI System One decision protocol. This client handles API key validation, request construction, and response parsing for decision queries, including specific validation for question types (noul, choice, score) and criteria bounds. It also includes a Probe method for connectivity verification and enforces security by preventing cross-origin redirects when no custom HTTP client is provided.

internal/safety/typesafe · high confidence

Introduce WSL distribution discovery and probing

The platform now discovers installed Windows Subsystem for Linux (WSL) distributions and probes a selected one to verify it is a viable kernel host before any installation occurs. This change adds a new \internal/platform/wsl\ package that lists distributions via \wsl -l -v\ (handling localized output and UTF-16 encoding), validates the distribution name against the list, and runs fixed \--exec\ commands (\uname -m\, \id -un\) to determine architecture and user. It also ensures \wsl.exe\ is launched from the OS-reported System32 directory rather than relying on environment variables, and correctly distinguishes between a cancelled context and an empty machine when listing distributions.

internal/platform/wsl · high confidence

Introduce account client for identity, backups, and remote device management

The new \internal/platform/account\ package provides the client library for the \id.reasonix.io\ identity provider, enabling device-flow sign-in (printing a code for browser approval without embedded forms or local ports), session token management, and user profile resolution. It adds support for backing up signed-in account configuration to the provider, managing enrolled phone controllers (listing, approving, replacing, or revoking them), and registering remote devices with specific capabilities (tasks, logs, files, desktop, terminal) to facilitate remote access and file attachment grants.

internal/platform/account · high confidence

Introduce background work scheduler with foreground priority

Added a new scheduler in the backgroundwork package that manages background tasks by prioritizing interactive foreground reads over disposable projections. The scheduler allows foreground activity to preempt running background work and enforces a quiet period before background tasks resume, ensuring that user-visible operations are not blocked by background processing.

internal/base/backgroundwork · high confidence

Introduce chunk-based delta updates to reduce download size

The platform now updates installations by downloading only the content-defined chunks that are missing from the existing install, rather than fetching the entire installer package. This new delta mechanism splits files into chunks using a rolling hash, allowing updates to reuse unchanged bytes even when files are modified or reordered, significantly reducing network usage. The system validates every chunk against its SHA-256 hash to ensure integrity and assembles the new release into a staging directory before applying it. A built-in threshold prevents the chunked update from being used if the missing data exceeds 50% of the total release size, ensuring the full package is downloaded in cases where delta benefits are negligible.

internal/platform/delta · high confidence

Introduce closed v1 chart specification with strict validation and text summaries

The \internal/contract/chartspec\ package now defines a closed, versioned chart specification (v1) that models use to request charts and frontends to render. This spec enforces strict schema validation, rejecting unknown fields, duplicate keys, and unsafe data types (such as lossy integers or control characters) to prevent injection or rendering errors. It supports bar, line, and pie marks with defined limits on rows, columns, and series. For environments where rendering is not possible, the spec provides a deterministic text summary and a tabular representation of the data, ensuring consistent behavior across different client capabilities.

internal/contract/chartspec · high confidence

Introduce custom slash commands and slash\_command tool

Users can now define reusable prompt templates as custom slash commands by placing Markdown files in configured command directories. These commands support argument substitution (e.g., $1, $ARGUMENTS) and frontmatter metadata. The system exposes these commands via a new \slash\_command\ tool, allowing the model to list available commands and expand them into instructions. The implementation handles plugin-qualified names, respects override precedence (later directories win), supports symlinked command files, and correctly decodes files encoded in GB18030.

internal/ext/command · high confidence

Introduce dedicated accounts worker for email/password and device-authorization sign-in

A new \workers/accounts\ service has been added to handle core identity operations, including email/password registration, login, logout, and password resets, as well as a device-authorization flow for CLI and desktop clients. The worker uses Hono to expose routes for authentication (\/auth\), device sign-in (\/device\), user profiles (\/u\), and account management (\/me\). It implements secure session management with HTTP-only cookies and Bearer tokens, password hashing via PBKDF2-HMAC-SHA256, and email verification/reset flows using Resend. The device-authorization flow allows CLI/desktop apps to poll for approval via a user code entered in the browser. The service also includes rate limiting for auth endpoints and CORS configuration for cross-origin requests.

workers/accounts · high confidence

Introduce egress proxy to enforce domain allowlists in sandboxes

A new loopback HTTP proxy in the egress safety package now enforces egress restrictions for sandboxed commands. It validates outbound connections against a configurable allow/deny list of domains, blocking traffic to loopback, link-local, and cloud metadata addresses. When a command attempts to reach a host outside the allowed list, the proxy intercepts the request, optionally prompts the user for permission, and returns a specific refusal reason (such as 'blocked-by-allowlist' or 'blocked-by-address') that is recorded for attribution.

internal/safety/egress · high confidence

Introduce hostwrite package to centralize host directory and cache paths

Added the \internal/base/hostwrite\ package to define and manage the specific host directories (temporary folders and toolchain caches) that jailed commands are permitted to write to. This change introduces \Dirs()\ for jail grant lists and \Caches()\ for untrusted cache checks, ensuring that the process's home directory is read only once to derive both lists, thereby improving consistency and performance across platforms.

internal/base/hostwrite · high confidence

Introduce peermail for secure, audited inter-session messaging

A new peermail subsystem provides an on-disk spool for messages between sessions on the same machine, ensuring that messages are held for the user and delivered only after approval. The implementation enforces strict security by keeping the spool directory private to the user, rejecting symlinks, and validating message integrity via HMAC proofs. It includes an immutable audit trail that records message decisions (sent, held, approved, denied) using digests rather than full bodies to preserve confidentiality, and supports features like duplicate detection, hop limits, and body sanitization to prevent control character injection.

internal/state/peermail · high confidence

Introduce plan-mode policy, state machine, and refusal logic

This change adds the internal plan-mode subsystem (internal/contract/planmode) that enforces a strict planning workflow: side-effecting tool calls are blocked until the user approves a plan, while read-only tools remain available. It introduces a lifecycle state machine (Planning → AwaitingApproval → Executing) with epoch-based authority tracking to prevent stale actions, and a policy engine that classifies calls by effect and safety. Refusals now include bounded, escaped messages that name the command, cause, and exit path, and the policy explicitly blocks browser\_open during planning in favor of web\_fetch. Tests verify that the policy matches built-in tool declarations, that refusal messages are safe and distinguishable, and that the state machine correctly handles transitions and staleness.

internal/contract/planmode · high confidence

Introduce read-only 'observe' posture for unattended runs

A new internal 'observe' package defines a read-only operational posture for unattended runs. This mode restricts available tools to those explicitly declared as read-only with safe reach (local read or host control), ensuring no write operations are permitted. It includes a ledger system to park requests requiring human approval (such as approvals or questions) without executing them, enforcing a limit of 32 pending items and a 7-day expiration. The posture also sanitizes input to remove control and bidirectional characters, and provides metadata for frontends to display the enforcement level (OS sandbox vs. tool filter) and confinement status.

internal/contract/observe · high confidence

Introduce scheduled-run store with budget accounting and policy enforcement

The internal schedule state package now includes a durable store that manages scheduled runs, enforcing strict budget limits and policy constraints. Users benefit from automatic circuit breakers that pause schedules when lifetime token budgets are exhausted or consecutive failures occur, as well as global and per-schedule daily/weekly token and run-count windows. The system ensures atomicity for claims and budget charges, prevents duplicate execution via unique trigger IDs, and handles crash resilience by preserving previous manifests. Additionally, it supports one-shot schedules, human-confirmed creation, and configurable policy overrides that can only tighten (not loosen) limits. Also in this area: 6 commits (2 fixes) modifying internal; Host verification defaults to off and supports optional tool capabilities.

internal/state/schedule · medium confidence · unverified

The \internal/ext/configbackup\ package now provides a complete workflow for exporting and importing user configuration. Backups are encrypted with XChaCha20-Poly1305 and Argon2id, ensuring credentials are redacted by default and session tokens never leave the machine. Restoring requires a two-step process: a preview that compares items against the current state, followed by an explicit apply. The system enforces strict consent for high-risk changes, such as restoring hooks, MCP servers, or provider endpoints that differ from the local machine, and validates file paths to prevent directory traversal attacks. It also supports distinct identity snapshots for hooks and agent presets to ensure accurate diffs and safe round-trips.

internal/ext/configbackup · high confidence

Introduce session-private temporary directory management

Added a new \sessiontemp\ package that provides a manager for logical-session private temporary directories. This ensures that tool executions (such as sub-agent runs) share a private temporary directory distinct from parent agents and sibling runs, while still allowing each sandboxed invocation to enter its own namespace. The manager handles lifecycle events like hot rebuilds, session clears, and branch switches, ensuring that retired generations are cleaned up only after all active leases are released, and prevents fallback to the host public temporary directory if creation fails.

internal/state/sessiontemp · high confidence

Introduce structured surface specification contract for rich UI forms

The internal contract layer now defines a formal specification for rendering native UI surfaces, enabling non-blocking user interactions through structured JSON schemas. This change adds support for editable text fields with draft handling, visual diff previews for content changes, and rich interactive components such as map-based multichoice selectors, product comparison cards with pricing and facts, and data dashboards. It also establishes a strict validation framework for button actions, ensuring that user inputs are typed correctly and that surface definitions cannot execute arbitrary host operations, while providing mechanisms to redact sensitive data from the resulting specifications.

internal/contract/surfacespec · high confidence

Introduce the \`reasonix tui\` terminal interface

Adds a new in-process terminal UI (\reasonix tui\) built on Bubbletea, providing a full-screen chat experience with a composer, transcript, and footer. The interface supports slash commands (e.g., /status, /clear, /resume), local shell execution (!), and completion menus. It handles interactive approval workflows, including multi-choice questions and consent-based URL elicitation with security warnings for non-HTTPS or local hosts. Data visualization is supported via \render\_chart\ tool calls, which are displayed as aligned, foldable tables. The UI also features clipboard image pasting, vi-style navigation, and a first-run connection panel.

internal/frontend/tui · high confidence

Introduce trusted state store with tamper-evident object storage and record chains

The \internal/state/trustedstate\ package now provides a content-addressed object store and per-stream record chains protected by host write guards. Objects are immutable, named by SHA-256, and published via non-replacing hard links to prevent concurrent writers from racing. The store caches verified objects in memory while the file retains its size, modification time, and platform-specific identity (inode, device, and change time on Linux/BSD; size and mtime on other platforms), serving cached data only when these attributes match to detect tampering or race conditions. Streams maintain a chain of records ending in a head file, which is atomically replaced via rename to ensure crash safety. The store detects tampering by comparing the current head against previously observed heads and verifies record integrity by re-reading and hashing payload bytes, refusing rollbacks or forged histories.

internal/state/trustedstate · high confidence

Introduce typed port interfaces for the Reasonix Studio frontend

The \desktop/frontend-next/src/port\ directory now defines the TypeScript interfaces that structure the communication between the new Reasonix Studio frontend and the host kernel. These files establish the contract for account authentication (device grants), remote connection management (relay handshake failures, link resiliency, and binary framing), and local configuration (permission boundaries, sandbox settings, and backup/restore plans). This change provides the type-safe foundation for the new serve-backed UI, enabling features like remote device control, granular tool permissions, and encrypted configuration backups.

desktop/frontend-next/src/port · high confidence

Introduces Extension Protocol v2 UI Hub for structured surface management and secure event emission

The new \uihub\ package implements the host-side logic for the Extension Protocol v2, enabling extensions to publish structured UI surfaces (cards, forms, notifications, and views) and handle blocking prompts. It ensures that all user-visible text from extensions is automatically redacted of credentials before being emitted as frontend events, and it manages the lifecycle of these surfaces by isolating them per session and generation, dropping stale or crashed client updates to prevent state corruption.

internal/ext/extension/uihub · high confidence

Introduces chunked delta updates for Windows with platform-specific application swapping

The appupdate module now supports downloading and applying updates in chunks (deltas) on Windows, significantly reducing bandwidth usage for incremental releases. This feature includes logic to verify delta integrity, skip chunked updates if they would download more data than a full package, and handle swap failures by falling back to the full installer. The implementation is split by platform: Windows uses a dedicated installer execution path, while other platforms stage and swap the application bundle directly. The update process is user-controlled, requiring explicit confirmation before restarting the application, and includes robust error reporting and progress tracking for both delta and full downloads.

internal/platform/appupdate · high confidence

Introduces custom DNS resolver with DNS-over-HTTPS support

The \internal/base/netdns\ package now provides a custom DNS resolution layer that replaces the system default, addressing Android environments where standard resolution fails. It supports configuring specific name servers via the \REASONIX\_DNS\ environment variable or \\[network\]\ settings. A key addition is DNS-over-HTTPS (DoH) support, which encrypts queries to a specified endpoint, preventing tampering on untrusted networks. The implementation includes validation to ensure DoH endpoints use HTTPS and handles connection pooling and timeouts appropriately for mobile devices.

internal/base/netdns · high confidence

Introduces shared session management and usage catalog handling for headless runs

The \clisession\ package is introduced to provide shared logic for processes running agent sessions without a terminal UI, such as scheduled child processes. It centralizes session directory resolution, lease binding, and recovery branch cleanup, ensuring that headless runs can properly attach to session files and manage state. Additionally, it adds process-level usage catalog draining and closing, allowing scheduled children to safely flush and close shared statistics and history projections at process exit without interfering with other controllers.

internal/frontend/clisession · high confidence

MCP Registry client with pagination, caching, and exact-match resolution

The \internal/ext/mcpregistry\ package now provides a client for the official Model Context Protocol Registry that supports paginated search and precise installation name resolution. The \Resolve\ method follows pagination cursors to find exact server names and requires live registry metadata for installation (preventing stale cache usage), while \Search\ supports browsing with a local cache that expires independently per query. The implementation includes logic to handle transport types (stdio, http, sse) and generates safe local plugin names from registry identifiers.

internal/ext/mcpregistry · high confidence

Memory facts now support explicit activation and freshness controls

Users can now control how memory facts are presented to the model. Facts can be marked as 'pinned' to always appear in the stable session prefix or 'relevant' for retrieval-only recall, with legacy global user facts defaulting to pinned for compatibility. Additionally, facts now have a freshness classification (fresh, current, stale, expired) based on configurable volatility windows or type defaults, allowing the system to age out stale information and exclude expired facts from automatic recall while keeping them visible in management tools.

internal/state/memory · high confidence

MemoryBench adds CJK, conflict, contradiction, and distractor task suites

The MemoryBench benchmark suite in benchmarks/memorybench now includes four new task categories to evaluate memory handling: CJK (mb-cjk) tests naming conventions for database migration scripts; conflict (mb-conflict) checks whether the model correctly overrides a global default deploy region (us-east-1) with a project-specific one (eu-central-1); contradiction (mb-contradiction) verifies that the model follows a specific package manager instruction (pnpm) despite conflicting documentation mentioning npm; and distractor (mb-distractor) introduces 100 irrelevant historical notes to test the model's ability to ignore noise when answering a specific command-check question.

benchmarks/memorybench · high confidence

Mod host runs installed Claude Code mods as supervised sidecars

The new modhost package enables the application to run installed Claude Code mods as independent sidecar processes. It discovers mods by reading their hooks.json, resolves a compatible Node.js runtime (requiring Node 22+ or the host's embedded Electron), and launches the mod's hooks module via an embedded JavaScript host script. The host supervises these processes with strict timeouts, crash limits, and dynamic tool registration, while intercepting events like tool calls and prompts to enforce user-defined rules and security ceilings.

internal/ext/modhost · high confidence

New 'advise' tool for on-demand expert model consultation

The system now includes a new 'advise' tool that allows the working model to pause and consult a stronger, paid model for specific guidance. When invoked, the tool sends the current conversation transcript (including the user's request, previous tool calls, and results) to the stronger model, which returns a text-based plan, diagnosis, or review without executing any code. This enables the agent to seek expert input when stuck or before making risky changes, with usage billed separately to the advisor model.

internal/tools/advisor · high confidence

New CI signing and manifest generation tool for desktop releases

A new command-line tool (\desktop/cmd/sign\) has been added to the release pipeline to handle artifact signing and manifest generation. It uses minisign to create detached signatures for desktop binaries and generates a \latest.json\ manifest that maps platforms (including the new Windows ARM64 support) to their download URLs, sizes, and SHA256 digests. The tool also supports generating fresh signing keys and verifying signatures against the embedded public key, ensuring that the signing path and the runtime updater verify path share a consistent definition of the manifest and signature format.

desktop/cmd/sign · high confidence

New CLI account management and ACP agent server

The CLI now includes account management commands (\login\, \whoami\, \logout\) for networked features like the forum and crash reporting, using a device-flow sign-in process. It also introduces a new \--acp\ mode that runs Reasonix as an Agent Client Protocol (ACP) server, allowing editor integrations to drive sessions via stdio JSON-RPC with support for session-level permission postures and sandbox overrides.

internal/frontend/cli · high confidence

New Extension Protocol v2 sidecar client with content-ref externalization and strict capability enforcement

The sidecar component now implements the host half of Extension Protocol v2, introducing a new client that spawns and supervises extension sidecar processes with bounded timeouts for handshakes, shutdowns, and intercepts. It enforces strict capability declarations, rejecting sidecars that activate providers, UI actions, or dynamic tools without explicit manifest permission, and validates protocol version compatibility. Large payloads are automatically externalized into a per-connection content store using content refs when they exceed the threshold, with chunked reading and SHA-256 verification to ensure data integrity, while small payloads remain inline. The client also handles UI routing, provider stream notifications, and crash supervision, with error messages redacting sensitive credentials from peer responses.

internal/ext/extension/sidecar · high confidence

New Gemini provider using the Interactions API with video and reasoning support

This change introduces a new \gemini\ provider that communicates with Google's Interactions API (\POST {base}/interactions\) using a stateless, hand-written HTTP client. It supports streaming responses, including reasoning steps (thoughts) and function calls, which are stored and replayed verbatim to maintain conversation state. The provider also adds video understanding capabilities, allowing videos to be sent either inline as base64 data or via the Files API, with native sniffing for WebM and MOV formats. Configuration options include setting the API key source, custom headers, output token limits, and video protocol selection.

internal/model/gemini · high confidence

New LSP integration for code intelligence

Added a new Language Server Protocol (LSP) client in \internal/platform/lsp\ that provides read-only code intelligence tools (definition, references, hover, and diagnostics) for the agent. The implementation manages the lifecycle of external language servers (such as gopls, rust-analyzer, and pyright) on a per-language basis, handling JSON-RPC communication, document synchronization, and position encoding conversion. It includes a security fix that rejects oversized Content-Length headers to prevent unbounded memory allocation, and enforces workspace scoping to refuse network paths before they reach the server.

internal/platform/lsp · high confidence

New Reasonix Studio interface added

A new HTML structure for the Reasonix Studio has been introduced in the output directory. This interface includes a sidebar for navigation (workspace, inbox, library, projects, tasks), a main content area, and a status bar. It also features a modal dialog for interactions and accessibility elements like live regions for screen readers. The interface is designed for a 'personal workspace' and includes placeholders for various features like search, notifications, and settings.

output · high confidence

New Responses API provider with robust error handling and vendor-specific optimizations

Added a new provider implementation for the OpenAI Responses API wire protocol, enabling support for stateless (DeepSeek) and stateful (OpenAI) modes. The implementation includes automatic fallback to full history when a relay rejects the previous\_response\_id, flexible parsing of error codes that may be sent as strings or numbers, and specific optimizations for DeepSeek (shared context window replay) and OpenAI (input\_image detail metadata). It also enforces strict schema compliance for input items, ensures API keys are resolved dynamically per request, and adds a five-minute idle timeout for silent models.

internal/model/responses · high confidence

New Studio rail panels for agents, context, cost, and extensions

The Studio rail now includes dedicated panels for sub-agents, context usage, session cost, and extension surfaces. The Agents panel lists running and completed sub-agents with status indicators and links to full transcripts. The Context panel displays a gauge against the fold point (not just the window), showing usage breakdowns and pressure levels. The Cost panel shows session costs, wallet balances, and per-source breakdowns. The Extensions panel renders standing extension panels with progress bars, fields, and actions. These panels replace previous inline displays with a structured, scrollable rail view.

desktop/frontend-next/src/ui/panels · high confidence

New UI interaction census tool for static analysis

A new static analysis tool has been added to the \desktop/frontend-next/tools/ui-census\ directory to audit frontend UI interactions. The tool parses the codebase to map interaction roots (such as JSX handlers and DOM events) to their effects, classifying them as MUTATION, READ\_ONLY, or UNRESOLVED. It enforces structural invariants—such as ensuring every mutating capability reaches a known endpoint and that component identity maps are invertible—and provides a gate mechanism to fail CI if these invariants are broken beyond declared limits. The tool also tracks the provenance of state updates, promise continuations, and event parameters to identify hidden writes or unresolved dependencies.

desktop/frontend-next/tools/ui-census · high confidence

New account, crash-report, and forum workers with unified identity and enhanced diagnostics

This change introduces three new Cloudflare Workers: the \accounts\ worker at \id.reasonix.io\ provides the central identity service (email/password auth, sessions, and device-authorization for CLI/desktop); the \crash-report\ worker at \crash.reasonix.io\ now unifies dashboard access onto this shared identity, folds in the skill/MCP registry data plane, and adds structured crash diagnostics (Go panics, breadcrumbs, group summaries) alongside CLI telemetry and a daily stats purge; and the \forum\ worker at \forum.reasonix.io\ launches the community forum API with shared identity and built-in anti-spam trust levels. All three workers are backed by D1 databases with specific schema migrations and rate-limiting bindings.

workers · high confidence

New anonymous telemetry system with daily pings and content-free counters

The platform now includes a new telemetry client that collects and uploads anonymous, content-free usage counters. The system sends a daily 'ping' to verify installation status, with robust retry logic that persists and re-sends the ping each UTC day while the process is open. It also aggregates and uploads bounded counters (such as turn counts, token usage buckets, and error rates) without capturing any private user data or prompts. Telemetry is opt-in by default for interactive release builds, respects environment variables like DO\_NOT\_TRACK and CI flags, and stores pending data locally to survive restarts.

internal/platform/telemetry · high confidence

New automated UI audit and regression-sweep tooling

Added a suite of headless browser automation scripts (audit-sweep, contrast-sweep, fault-sweep, i18n-sweep) and a shared step-definition module to systematically verify the rendered interface. These tools walk through key UI states to detect specific regressions: contrast violations against actual painted backgrounds, Chinese text leaking into English interfaces, inaccessible or offscreen elements, and broken optimistic-write rollback behavior in settings panels. A fixture script is also included to automate the initial onboarding flow required for these sweeps to reach the main workbench.

desktop/frontend-next/tools · high confidence

New benchmark suite for measuring fan-out concurrency

A new benchmark suite named \fanout-width\ has been added to evaluate the performance benefit of running fan-out members concurrently versus serially. The suite includes a specific task, \fleet-parallel-three\, which requires the system to independently query three independent subsystems (alpha, beta, gamma) for their active runtime markers and then aggregate the results. This allows users to measure whether concurrent execution reduces wall-clock time without sacrificing accuracy, using the \e2ebench\ tool to compare wide (parallel) and serial execution modes.

benchmarks/fanout-width · high confidence

New benchmark to measure tool schema position impact on cache hits

Added a new benchmark tool (\benchmarks/tool-surface-cache\) that measures the cost of changing the provider-visible tool set between turns against a provider's prefix cache. The tool demonstrates that appending tools at the tail is cache-friendly (low miss rate), while inserting or editing tools earlier in the schema causes cache misses for subsequent tools, helping users understand how tool surface stability affects caching performance.

benchmarks/tool-surface-cache · high confidence

New completion report with model claim reconciliation and gap detection

The completion subsystem now generates a structured host-authored report that evaluates whether a task is done, partial, or incomplete based on evidence. It introduces a 'Claim' model allowing the AI to assert verifications, unverified items, and risks, which are reconciled against the evidence ledger to detect unbacked claims (e.g., assertions without matching runs, failed runs, or runs predating the latest change). The report also tracks changes, verifications (handling stale or inconclusive results), and specific gaps like unobserved mod effects or writes outside observed roots, providing a transparent audit trail of what is proven and what is not.

internal/runtime/completion · high confidence

New design system tokens and layout documentation

The frontend-next styles now ship a formalized design system. A new TOKENS.md document and the generated tokens.css define the interface's vocabulary, including registered CSS properties for smooth column tweening, a structured theme-pack mechanism that applies via a generated stylesheet, and a refined hue system with dedicated ink tokens to ensure text contrast. The app.css introduces a 'sky' background effect and a new elevation model for floating layers. Additionally, new stylesheets and tests cover specific UI components like the agent run header, the docked browser panel, and the changed-files pill, ensuring the layout behaves correctly across different viewports.

desktop/frontend-next/src/styles · high confidence

New desktop packaging infrastructure for multi-architecture builds

The desktop packaging system has been rewritten to support building and shipping native packages for Windows ARM64, macOS Universal (x64/arm64), and Linux. This change introduces a new build pipeline that compiles the Go-based host binary and the C++/Swift computer-use helper for specific architectures, ensuring correct binary compatibility. It adds logic to canonicalize artifact names to match the release vocabulary, strips unnecessary debug symbols to reduce bundle size, and generates platform-specific icons (including high-DPI tray icons and Windows Start menu tiles) from a single SVG source. Additionally, it enforces that Linux binaries are statically linked to avoid glibc dependency issues and refuses macOS blockmap files that are not needed for updates.

desktop/electron/packaging · high confidence

New diff preview package for file change visualization

A new \internal/base/diff\ package has been added to compute line-level diffs and render them as unified diffs with added/removed line counts. This enables the CLI to display a shell result showing the exact output of pending file operations (write, edit, multi-edit) before they are applied, allowing users to preview changes via approval cards or changed-files panels without touching the disk. The package handles file creation, modification, and deletion, detects binary files to avoid noisy byte-diffs, and includes a performance cap to skip expensive exact diffing for large rewrites, falling back to line tallies instead. It also provides a parser to validate whether text conforms to the standard unified diff format used by git.

internal/base/diff · high confidence

New encoding detection and conversion library for CJK and legacy files

The internal file utilities now include a new \encoding\ package that detects and converts file encodings (UTF-8, UTF-16, GB18030, GBK, and LossyUTF8) to ensure CJK Windows files remain editable without byte mangling. It introduces a detection cascade that prioritizes BOMs, then strict UTF-8, then legacy Chinese charsets, and finally lossy UTF-8. The library handles edge cases like cut fragments (partial characters at buffer boundaries) and distinguishes GBK from GB18030 by verifying if a write can restore the original bytes, preventing silent corruption of files containing specific single-byte characters like the euro sign.

internal/base/fileutil/encoding · high confidence

New eventwire package defines the frontend JSON contract

The internal/contract/eventwire package now provides the shared JSON schema for all frontend-facing events. This introduces structured wire types for approval requests (including recovery scenarios), tool dispatches and results (with argument normalization and file deliveries), extension UI surfaces (cards, forms, panels, and views), and usage/cost reporting (with dual-write support for legacy cost fields and new costQuote data). It also adds wire definitions for retry reasons, stream attempts, context maintenance, workspace changes, and completion receipts, ensuring that sensitive data like credentials are redacted in surface arguments before transmission.

internal/contract/eventwire · high confidence

New execfs layer for remote file operations without SFTP

A new \execfs\ package provides a remote file system layer for environments that cannot use SFTP, such as WSL distributions or SSH hosts with restricted subsystems. It implements the \sftpfs.Volume\ interface by executing small, fixed POSIX shell scripts over a transport, passing file paths as positional arguments to prevent shell injection. The implementation includes atomic file writes using unguessable temporary names, strict transfer size limits, and distinct error codes for missing files, directories, permission issues, and missing tools. Comprehensive tests verify correct handling of hostile filenames, path traversal protection, and atomicity guarantees.

internal/platform/remote/execfs · high confidence

New extension kernel with deterministic build pipeline and conflict handling

The extension subsystem now uses a new Builder pipeline to assemble a frozen RuntimeSnapshot from registered contributors. This pipeline enforces deterministic ordering regardless of contributor registration order, ensuring stable tool schemas, interceptor chains, and catalog listings. It introduces a ConflictPolicy mechanism: by default, same-tier duplicate IDs from different sources cause a hard failure (ConflictError), but the new ConflictCollect policy allows the build to succeed by recording the dispute in snapshot diagnostics while still resolving a deterministic winner. The system also includes a dependency graph resolver for component lifecycle management, performance baselines for CI, and adapters that map discovered skills, commands, hooks, MCP servers, and providers into the kernel's unified contribution model.

internal/ext/extension · high confidence

New extension protocol DTOs and validation layer

The extension protocol now defines a comprehensive set of data-transfer objects (DTOs) and strict validation logic for the host-sidecar communication. This includes structured types for lifecycle management (initialization, shutdown), provider streaming (credential-free messages, tool calls, usage), and UI surfaces (cards, forms, views, and status). The protocol also introduces session capabilities (reading usage/messages, submitting prompts) and content references for handling large payloads. All DTOs are validated against strict schemas to ensure wire compatibility and security, with specific bounds on view complexity and anchor safety.

internal/ext/extension/protocol · high confidence

New fileutil package for atomic writes, path containment, and network path scoping

The internal/base/fileutil package introduces a suite of utilities to improve file operation reliability and security. AtomicWriteFile and AtomicWriteFileStrict provide crash-consistent file publishing with retry logic for transient locks and cross-device fallbacks, while AtomicCreateFile and AtomicOverwriteFile handle non-overwriting and permission-preserving updates. Path safety is enforced through CanonicalWritePath (resolving aliases and rejecting ambiguous Windows paths), containment checks (Under, AtOrUnder), and NetworkScope, which restricts Windows network paths to the workspace root. Additional helpers include VCSStores for version control detection, MarkUntracked for gitignore management, and slash-normalized glob matching.

internal/base/fileutil · high confidence

New gitcommit platform for staging, committing, and discarding changes

The \internal/platform/gitcommit\ package introduces a new capability to manage Git operations directly from the application. Users can now stage and unstage specific files or all changes, create local commits with safety checks for sensitive content (like private keys or credentials), and discard individual file changes to revert them to their last committed state. The implementation ensures that only paths explicitly listed by Git are processed, preventing path injection, and it handles edge cases such as repositories with no prior commits and nested submodules.

internal/platform/gitcommit · high confidence

New gitstatus platform for branch switching, commit details, and history

The internal platform for Git status and operations has been expanded with a new \gitstatus\ package that provides robust, safe access to repository state. Users can now switch branches with pre-flight checks that prevent data loss by refusing to overwrite uncommitted changes or untracked files that would collide with the target branch, while also respecting linked worktrees. The platform exposes detailed commit information, including file paths, statuses (such as renames and copies), and line insertion/deletion counts, as well as a paginated history view across all branches. Additionally, the implementation hardens security by ensuring that repository-configured programs (like filters or diff drivers) are never executed during status reads or diffs, and it correctly handles edge cases like binary files and nested submodules.

internal/platform/gitstatus · high confidence

New history catalog for session search and indexing

The \internal/state/historycatalog\ package introduces a new SQLite-backed catalog that indexes session transcripts for full-text search. It handles incremental updates (appends and rewrites), manages source health by marking unreadable transcripts as corrupt and retrying them, and ensures search results only include the user's own words by stripping host-injected blocks. The catalog also supports cross-process safety during scans and includes a rebuild mechanism for the FTS projection.

internal/state/historycatalog · high confidence

New history search tool and indexed catalog system

Users can now search saved local session history using a new 'history' tool that supports lightweight BM25 retrieval and an 'around' operation to inspect messages near a hit. The system includes an indexed catalog that asynchronously indexes session files into a SQLite database for faster, scalable search, with support for project and global scopes, tool input/error indexing, and cleanup-pending session filtering. A new instance ID system ensures unique identification per data home, preventing session file conflicts across multiple Studio windows or installations.

internal/platform/instanceid, internal/tools/sessiontool · high confidence

New interactive card components for approvals, questions, and context compaction

The frontend now includes dedicated UI components for handling model requests and session management: ApprovalCard displays permission prompts with detailed scope explanations and decision options; AskCard presents multi-question interactions with stepper navigation, free-text input, and optional prompt refinement; CompactionCard shows context window compression status with specific trigger reasons and remaining content details. These components are accompanied by comprehensive test coverage for their interaction behaviors, rendering, and edge cases.

desktop/frontend-next/src/ui/cards · high confidence

New internal base utilities for file locking, network diagnostics, and search retrieval

This change introduces several new internal packages under \internal/base\. The \filelock\ package provides cross-process advisory file locking with bounded retries. The \netclient\ package adds a network diagnostic tool that traces connectivity failures through proxy, DNS, TCP, and TLS steps, along with a stream dialer for non-HTTP protocols. The \retrieval\ package implements a local BM25 search engine with CJK bigram tokenization and a shadow-only V2 ranking system with field weighting. Additionally, the \tempdir\ package improves test cleanup by waiting for Windows directory teardown, and \nilutil\ adds a helper to detect typed nil interfaces.

(repo-wide) · high confidence

New internal scheduler for managing scheduled runs

The internal/runtime/scheduler package introduces a new system for managing scheduled runs. It enforces strict security by only allowing schedules to operate within user-trusted project folders, rejecting any workspace not explicitly marked as trusted. The scheduler ensures that only one run executes at a time per schedule, handles missed slots by skipping them if they exceed a grace period, and respects a global pause switch. It also pins the AI model to the user's configuration to prevent repositories from altering execution parameters.

internal/runtime/scheduler · high confidence

New linting and pricing verification tools

Added \covergate\ to enforce minimum code coverage floors on declared sensitive paths, \pricecheck\ to verify vendor pricing tables against live web pages, and expanded \repolint\ with new rules for build artifacts and untracked file detection.

tools · high confidence

New release validation and automation scripts

The scripts directory now includes a suite of new tools to support the Stable and Preview release channels. This includes \compare-cli-release-tags.sh\ and \compare-desktop-release-versions.sh\ for strict version comparison, \decide-cli-pointer-update.sh\ and \decide-desktop-pointer-update.sh\ to manage public update pointers, and \decide-cli-release-publication.sh\ to validate release assets and checksums. Additional scripts like \brew-cask-guard.mjs\ protect Homebrew cask integrity, \backfill-issue-labels.mjs\ automates issue categorization using the DeepSeek API, and \credential-leak-check.mjs\ prevents test suites from printing real credentials.

scripts · high confidence

New remote attach layer for SSH and WSL connections

The \internal/platform/remote/attach\ package introduces a new connection management layer that allows the application to attach to remote workspaces via SSH or Windows Subsystem for Linux (WSL). This layer handles the lifecycle of remote kernels, including bootstrapping the \reasonix serve\ process, managing reference-counted connections and port forwards, and translating Windows paths for WSL distributions. It supports browsing remote directories without starting a kernel, probing host prerequisites, and securely staging launch scripts in private directories. The implementation includes a connection pool to share SSH sessions across multiple workspaces on the same host and specific handling for WSL session anchors to keep long-running processes alive.

internal/platform/remote/attach · high confidence

New remote host attachment and account authentication endpoints

The serve layer now exposes new HTTP endpoints to support remote machine attachment and user account management. The \remotehost\ package provides the implementation for dialing remote machines, browsing their file systems, and probing their status, exposing these capabilities via the serve layer. Additionally, new \/account\ routes (\/account\, \/account/login\, \/account/poll\, \/account/logout\) implement device-code based sign-in, which are strictly gated and only enabled when the local host explicitly grants permission via \AllowAccountAuth\. The \adjudications\ endpoint now exposes a snapshot of active and historical host adjudication barriers, and \agent-presets\ routes allow for the management of agent composition presets, including import, export, and selection, with strict edit grants required for mutations.

internal/frontend/serve · high confidence

New rules system with path-scoped application and strict authority boundaries

The system now supports a new \rules\ extension that loads markdown rule files from user and project directories (\.claude/rules\, \.reasonix/rules\). Rules can be unconditional or path-scoped using \paths:\ frontmatter globs, applying only when a session touches a matching file. The skill system enforces strict authority boundaries: external profiles cannot declare their own authority (e.g., \authority: review\), and the \delivery\ namespace is strictly scoped to \review-report\ values. Built-in skills like \review\ and \security-review\ are read-only and defer version control discovery to the workspace fact. Skill supplements (references, scripts) are confined to the skill directory and loaded only on invocation, not during listing.

internal/ext/skill · high confidence

New run\_script tool for batched tool execution via Starlark

A new \run\_script\ tool is introduced, allowing the model to execute short Starlark (Python-like) programs that invoke other tools in a single round trip. This enables batching dependent steps—such as reading multiple files or looping over results—so that only the script's printed summary is returned to the conversation. The implementation includes safety bounds (max 64 calls, 32 KB script size, 64 KB output) and provides \call\ (which stops on error) and \try\_call\ (which returns a success/failure struct) for flexible error handling within the script.

internal/tools/script · high confidence

New scratch package for safe, lock-owned temporary directories

The internal/base/scratch package provides a mechanism for managing temporary directories that are owned by a running process via an exclusive file lock. This ensures that abandoned directories (e.g., from killed kernels) are automatically reclaimed by a sweep process, while preventing the deletion of directories still in use by active owners. It includes a Create function to claim a new directory, a Remove function to safely release and delete it with retry logic for Windows file locks, and Sweep/SweepOnce functions to clean up unowned directories that have passed a creation grace period.

internal/base/scratch · high confidence

New selectable output styles for agent communication

Users can now choose from built-in communication styles (explanatory, learning, concise) or define custom ones via markdown files in project or home convention directories. These styles modify the agent's system prompt: they can append reasoning instructions while keeping coding capabilities (KeepCoding) or replace the prompt entirely for a pure persona. Custom styles override built-ins and support UTF-8 and GB18030 encodings.

internal/runtime/outputstyle · high confidence

New self-update mechanism for macOS and Linux

The update platform now supports self-updates on macOS and Linux. On macOS, updates are applied by swapping the entire .app bundle using a versioned directory layout and a handoff process that waits for the current application to exit. On Linux, updates are applied by extracting a tarball into a new version directory and activating it via a pointer swap, with support for .deb package installation on Debian-based systems. The update cache now stores artifacts with their version, platform, and SHA256 digest, and verifies them before installation. Rollbacks are supported by installing an older version from the cache.

internal/platform/update · high confidence

New shell command analysis and security enforcement in internal/base/shellparse

The internal/base/shellparse package now provides comprehensive static analysis for Bash and PowerShell commands to enforce security policies and improve error attribution. It introduces StaticFields and ParseStaticCommand to safely decompose Bash commands into executable arguments while rejecting dynamic expansions, redirections, and control operators. A new delete-analysis capability (AnalyzeDeleteCalls) identifies and flags unsafe file deletion commands, with specific logic to refuse recursive deletes targeting paths outside granted roots. The package also adds wrapper transparency for commands like sudo, env, and doas, allowing the system to see through these layers to the underlying command. Additionally, it implements precise exit-status and pipeline analysis (ExitZeroImplies, TerminalPipelines) to correctly attribute command outcomes in complex shell chains, and includes a PowerShell-specific analyzer for detecting delete operations in scripts.

internal/base/shellparse · high confidence

New structured event contract for agent turns

The internal event system has been replaced with a typed, structured event stream that decouples agent logic from frontend rendering. This change introduces a comprehensive set of event types (such as TurnStarted, Reasoning, Text, ToolDispatch, ToolResult, and Usage) and associated data structures (including Attachment, OutputBound, CacheDiagnostics, and CompletionSummaryInfo) to convey precise state changes, tool execution details, and cost diagnostics. It also adds a coalescing mechanism to merge rapid streaming deltas for performance, a cost-quote sink for currency-aware pricing, and an audit forwarder for host-side telemetry, enabling frontends to render rich, localized UIs without parsing fragile text formats.

internal/contract/event · high confidence

New studio-delta tool for building and simulating release deltas

A new command-line tool, studio-delta, has been added to the cmd/studio-delta directory. It enables the creation of delta indexes and chunk stores from Studio release archives (build mode) and allows for the simulation of update replays against existing installations (simulate mode). This tool measures the bandwidth and storage costs of delta updates by calculating missing chunks, reused data, and total download sizes before any client performs the actual update.

cmd/studio-delta · high confidence

New task control and persistence layer in the runtime

The \internal/runtime/taskmonitor\ package now provides the core infrastructure for managing task lifecycles. It introduces a \ControlService\ that handles atomic task operations (stop, cancel, requeue) with idempotency guarantees and version-based conflict resolution. A new \FileStore\ implementation persists task snapshots and event logs to the \.reasonix/tasks\ directory, featuring robust security measures such as path-traversal rejection and symlink chain validation to prevent state leakage. The system also includes cross-platform file locking (using \unix.Flock\ and \windows.LockFileEx\) to ensure data integrity during concurrent access, and ensures that all internal task state remains invisible to \git status\.

internal/runtime/taskmonitor · high confidence

New text utility library for safe display and preview handling

The internal/base/textutil package introduces a suite of functions to safely process text from untrusted sources (such as MCP servers or tool descriptions) before displaying it in the UI. It provides SanitizeDisplay and SanitizeLaunch to strip escape sequences, control characters, and bidirectional overrides that could repoint terminals or cause visual spoofing. It also includes BoundProse and BoundLiteral for generating safe, grapheme-aware previews with configurable limits, ensuring that long text is truncated without splitting complex characters (like emojis or CJK) and that hidden content removal is flagged to the user. Additionally, it offers HumanBytes for readable size formatting and helper functions for grapheme clustering.

internal/base/textutil · high confidence

New theme pack system with installation, removal, and plugin integration

Users can now install, remove, and manage custom theme packs via a new internal theme system. Packs are installed from ZIP archives or loose files, with strict validation of manifests and assets (background and preview images) to prevent path traversal and oversized downloads. Installed packs live in the user's library and can be removed, while shipped default packs remain immutable. The system also supports theme packs contributed by plugins, which are discovered and loaded in place without being copied to the user library, and are automatically hidden when the plugin is disabled or removed. Additionally, theme packs can declare background placement preferences (focus, safe area, opacity) and density/corner recipes, which are validated and applied if valid.

internal/ext/theme · high confidence

New trajectory recorder with audit trails and event merging

The system now records a complete, replayable JSONL trajectory for every run, capturing not just events but also structured audit data such as subagent handoffs, memory recall details, protocol recoveries, and capability versions. To improve readability and reduce file size, adjacent streaming text increments are coalesced into single records while preserving the original observation timestamps. Additionally, sensitive information like private URLs in elicitation requests is stripped from the persistent log to ensure security, while still being forwarded to the live frontend.

internal/state/trajectory · high confidence

New versioned install layout and repair subsystem

The platform now uses a versioned install layout (v1.20+) where updates are staged in unique directories and only swap the active version pointer in current.json after validation, ensuring atomic upgrades and safe rollback on failure. A new repair subsystem handles configuration and derived-state recovery by quarantining invalid files (such as broken TOML configs or dangling symlinks) and restoring from last-known-good snapshots, while also tracking update apply failures for diagnostics.

internal/platform/repair · high confidence

New wallet balance endpoint with multi-provider support and caching

The billing model now supports fetching wallet balances from multiple providers (DeepSeek, Moonshot, and OpenRouter) using a unified, format-declared approach. Users can now see their available credits across different currencies (USD, CNY) with proper symbol rendering and no implicit currency conversion. The system includes a caching layer to reduce network calls, ensuring status lines remain responsive. Security is improved by preventing API keys from being sent to unverified hosts and blocking cross-origin redirects.

internal/model/billing · high confidence

New write-claim scheduler for sub-agent concurrency and path safety

The runtime now includes a dedicated write-claim scheduler that manages sub-agent concurrency and prevents file-write conflicts. It enforces limits on total sub-agents and concurrent writers, serializing agents that claim overlapping workspace paths while allowing parallel execution for disjoint paths. The system supports 'optimistic' claims for agents that do not declare specific write paths, allowing them to run concurrently without blocking others, and provides detailed diagnostics on why a sub-agent is queued. Additionally, it introduces a write-grant mechanism that tracks declared and mid-run granted paths, ensuring the audit trail accurately reflects all allowed mutations while preventing unauthorized file access.

internal/runtime/writeclaim · high confidence

Opt-in export of GenAI traces to an OTLP collector

Users can now enable optional trace export for their runs to an OpenTelemetry-compatible collector via the \internal/platform/otlp\ package. When configured, the system structures run events into GenAI spans (agent invocation, chat attempts, and tool calls) and posts them over OTLP/HTTP JSON. The exporter is non-blocking and drops spans if the internal queue is full. By design, sensitive content such as prompts, tool arguments, results, file paths, and error text is never exported; only safe identifiers (model names, tool names) and telemetry metrics (token counts, timing, failure codes) are sent.

internal/platform/otlp · high confidence

Preview and apply skill updates from source folders

Users can now review proposed changes to an installed skill before applying them. The system generates a plan detailing added, changed, or removed files, ensuring that any local edits to the installed copy are flagged and require explicit confirmation to avoid accidental overwrites. The update process uses an atomic swap to maintain stability, and the system tracks the source to allow for future retries if an update fails.

internal/ext/skillupdate · high confidence

Privacy-preserving crash reporting and CLI version metadata

The Reasonix CLI now captures fatal crashes and background goroutine panics, saving them to disk so they are reported on the next start, with sensitive data (such as prompts) redacted from the reports. It also exposes build metadata (version, git commit, build time) via the \--version\ command and its verbose/JSON variants.

cmd/reasonix · high confidence

Push current branch to upstream via isolated throwaway repository

Reasonix now pushes the current branch to its configured upstream using a dedicated throwaway repository that borrows objects via alternates. This ensures that no repository-specific configuration (such as hooks, credential helpers, or URL rewrites) interferes with or redirects the push, while still respecting the user's own global or system-level URL rewrites. The workflow includes a preview step that displays the destination, commit count, and scheme (limited to https, ssh, or git) for confirmation, and enforces strict protocol policies and non-forced updates to prevent unintended changes.

internal/platform/gitpush · high confidence

Reliable storage root relocation with adoption and integrity checks

The storage subsystem now supports moving state roots to new locations with robust integrity guarantees. Moves are atomic: data is copied and verified before the configuration pointer is updated, and interrupted moves can be resumed via a journal. The system now detects when a target directory already contains the root's data (e.g., after a lost pointer) and offers an 'adopt' operation to point at existing files without copying. It also prevents merging with foreign files by refusing non-empty targets unless they are owned by the root or a pending move journal exists. Additionally, the storage survey now bounds directory walks with a time budget to prevent stalls, reporting partial results if a root is too large or slow to scan.

internal/state/storage · high confidence

Resumable npm CLI distribution with platform-specific packages and canary channel

The npm release process now publishes the Reasonix CLI as a main package with optional platform-specific sub-packages (darwin-arm64, darwin-x64, linux-arm64, linux-x64, win32-arm64, win32-x64) and supports an opt-in canary update channel. The build script embeds the version, git commit, and build time into the binary, and the publish logic ensures resumability by reusing already-published packages for the same candidate SHA, filling only missing ones, and tolerating registry visibility lag and forbidden cleanup errors without rolling back the canary tag.

npm · high confidence

Robust remote kernel lifecycle management with secure installation and attachment

The remote bootstrap system now reliably detects and attaches to existing \reasonix serve\ processes instead of replacing them, ensuring that multiple clients can share a single kernel instance without conflict. It introduces secure, cross-platform installation of the remote kernel binary, verifying SHA-256 digests for all downloads and properly escaping shell metacharacters in paths to prevent injection. The system also handles complex Windows scenarios, including launching detached kernels via script files, identifying hosts regardless of their OpenSSH login shell, and correctly resolving absolute workspace paths from the tree view.

internal/platform/remote/bootstrap · high confidence

SFTP file layer with text/binary detection and atomic operations

The remote platform now includes a new SFTP file system implementation that allows the application to interact with remote machines via SFTP. This layer introduces automatic detection of file content types (text vs. binary) to safely preview files without corrupting binary data, enforces read caps to manage memory usage during previews, and supports atomic file writes to prevent data corruption during uploads. It also provides robust directory listing and error handling for scenarios where remote servers may refuse the SFTP subsystem.

internal/platform/remote/sftpfs · high confidence

Sandbox now governs specific host services and protects Git metadata

The sandbox introduces a policy model for host services, allowing users to explicitly grant access to SSH agents, Docker, and Podman while masking their Unix sockets by default. It also enforces read-only protection on Git configuration files and hooks within write roots, preventing confined commands from altering repository metadata. Additionally, the sandbox can now bind specific host device nodes into the Linux environment and provides detailed diagnostics when the bubblewrap backend is unavailable due to missing installation or namespace restrictions.

internal/safety/sandbox · high confidence

Script added to slice Noto Studio font for faster first-screen rendering

A new Python script, slice-noto-studio.py, has been added to the frontend-next scripts directory to generate a smaller 'core' font subset from the full Noto Studio font. This script uses fontTools to extract common Simplified Chinese characters (via NotoSansSC-subset), Latin text, punctuation, arrows, and symbols, creating a lightweight WOFF2 file intended for the initial screen load. The full font remains available as a fallback for characters not included in this core set, allowing the application to reduce initial payload size while maintaining full typography support.

desktop/frontend-next/scripts · high confidence

Searchable conversation history with BM25 retrieval and robust index lifecycle management

The history subsystem now indexes saved session transcripts to enable full-text search across user and assistant messages, tool inputs, and tool errors using a lightweight BM25 algorithm. This change introduces a shared, process-wide indexed catalog that manages asynchronous opening, rebuilding, and closing of the search index, ensuring that lifecycle events like catalog closure properly fence in-flight operations and allow for safe restarts. The search functionality correctly handles edge cases, such as distinguishing between genuinely empty query results and failures caused by unreadable (corrupt) transcript files, and ensures that only the user's own words are indexed rather than host-injected blocks. Comprehensive tests verify the new search capabilities, the robustness of the catalog's lifecycle management, and the correct handling of unreadable transcripts and cleanup-pending sessions.

internal/state/history · high confidence

Secure, two-phase installation for skills, plugins, and MCP servers

The new installsource tool introduces a two-phase installation workflow for skills, plugins, and MCP servers, requiring explicit approval before any filesystem writes or network connections occur. It enforces strict security boundaries by bounding remote HTTP fetches to 2 MiB with a 30-second timeout, validating content digests to prevent supply-chain tampering, and blocking symlink targets that escape the project or home root. The system also redacts endpoint credentials from all preview and export surfaces, supports installing plugins from local ZIP archives, and handles directory root aliases while ensuring that only approved, pinned content is ever applied.

internal/ext/installsource · high confidence

Session store overhaul: auto-archiving, composition bindings, and robust metadata handling

The session store now supports automatic archiving of inactive conversations based on a configurable idle duration, while respecting user pins, active turns, and leases. Agent composition bindings are persisted within sessions, allowing immutable resources to be exported and imported via native archives without losing context. Metadata management has been hardened: sidecar reads are cached per transaction to improve performance, display fields are sanitized to remove internal wrappers, and cross-process updates are safely serialized. The store also handles legacy 1.x session imports, including specific ID formats and desktop prefixes, and ensures that composition history remains readable even if project launch overrides are rejected for security.

internal/state/sessionstore · high confidence

Shipped self-diagnostics skill for troubleshooting Reasonix capabilities

A new built-in skill named 'reasonix-guide' is now embedded in the binary, providing a comprehensive reference for troubleshooting and configuring Reasonix capabilities including Skills, Commands, Hooks, MCP servers, and plugin packages. The skill is designed to be inlined into the agent context and guides users through static and live diagnostic checks (e.g., \reasonix doctor capabilities\), explains configuration priority and discovery conventions, and details hook events, permissions, and MCP merge orders to help diagnose issues like missing skills, shadowed commands, or hook failures.

internal/ext/skill/builtincontent · high confidence

Static preview of mod hooks modules

The modpreview extension now provides a static analysis of a mod's hooks module to generate an install-time preview. By parsing the JavaScript/TypeScript syntax tree without executing the code, it lists registered events, API calls, and imports, while flagging constructs that prevent full visibility (such as eval, dynamic imports, or computed member access). This allows users to see what a mod intends to do before installation, with clear indicators for unsupported formats, oversized files, or unlistable code patterns.

internal/ext/modpreview · high confidence

Structured instruction resolution with host checks and import controls

The system now resolves standing instructions from convention files (REASONIX.md, AGENTS.md, CLAUDE.md) using a new internal resolver that enforces scope precedence (user, project, ancestor, local) and deduplicates shared imports to avoid redundant context. Users can define hard verification gates and sensitive path globs via a dedicated 'Reasonix host checks' section, which are extracted and enforced separately from ordinary guidance. The resolver also prevents instruction documents from importing files outside their workspace boundary or following symlinks to external locations, ensuring security and deterministic behavior.

internal/state/instruction · high confidence

Structured plan contract with approval gating and diff rendering

The \internal/runtime/plancontract\ package introduces a structured representation for proposed plans, replacing unstructured text with typed fields for objectives, steps, risks, and acceptance criteria. Plans are normalized to ensure consistent identity and structure, and a new \Diff\ mechanism compares revisions to detect expansions (such as new steps, increased risks, or broader file scopes) that require user re-approval, while allowing narrowings or reordering without interruption. The package also provides deterministic ordering of steps based on dependencies and phases, and renders both human-readable markdown and a serial task list for execution, ensuring the user's approved view matches the host's execution plan.

internal/runtime/plancontract · high confidence

Studio update manifest now supports delta updates and platform-specific installers

The new studio-manifest tool generates the latest.json update manifest, introducing support for chunked delta updates served from a mirror with signature verification, and correctly mapping platform-specific installers (Windows .exe, macOS .zip, Linux .deb) to their respective update channels while excluding portable archives from automatic updates.

desktop/cmd/studio-manifest · high confidence

Support pasting images from the system clipboard

The application now allows users to paste images directly from the system clipboard (e.g., via Ctrl+V), restoring functionality present in version 1.x. This change introduces a new \clipimage\ package that detects and reads image data from the clipboard using platform-specific tools: \osascript\ on macOS, PowerShell on Windows, and \wl-paste\ or \xclip\ on Linux. If the clipboard contains an image, it is made available for pasting; otherwise, the system falls back to handling the content as text.

internal/platform/clipimage · high confidence

Unified dashboard authentication and admin console

The crash-report worker now uses a shared identity service (id.reasonix.io) for authentication, replacing previous methods. This introduces a role-based access system (pending, viewer, admin) with an admin console for managing user roles and an audit log for tracking permission and data changes. Additionally, a community moderation interface has been added for reviewing and managing user-published skills, plugins, and MCP servers.

workers/crash-report · high confidence

User personas now compose prefix/suffix text into the system prompt

The runtime now supports user-authored persona settings that wrap the first-party system guidance with custom prefix and suffix text, rather than replacing it. Users can select from built-in presets (engineer, concise, mentor, critic) or define their own, with options to include or suppress runtime context. These changes are applied to new sessions and take effect after a rebuild of existing sessions, ensuring cache stability.

internal/runtime/persona · high confidence

The web search tool now supports an ordered chain of search backends, starting with a free, user-configured SearXNG instance and falling back to the DeepSeek Search API if the free backend fails. The DeepSeek API is a paid service that is only invoked after the user explicitly consents to the charge. The system provides detailed, structured error reporting for backend failures (e.g., transport issues, status codes, empty results) and includes security guards to prevent the SearXNG probe from connecting to link-local or cloud metadata addresses.

internal/tools/websearch · high confidence

Workspace repository discovery with caching

The gitrepos platform now discovers repositories within a workspace folder by scanning subdirectories (up to a depth of 2) to identify those that are top-level git work trees, rather than relying on names or request paths. The search is bounded to prevent excessive disk I/O (max 400 directories read, 24 repositories kept) and ignores symbolic links. A new caching layer ensures that if the workspace structure has not changed, the previous list of repositories is returned immediately without re-scanning.

internal/platform/gitrepos · high confidence

e2ebench adds anchor-resistance experiments and answer-isolation security controls

The benchmark harness now supports an 'anchor' experiment axis (blind, correct, or wrong) to measure how much a pre-supplied hypothesis influences agent performance, with reports explicitly naming skipped tasks and resistance against a blind baseline. To prevent answer leakage in memory-effectiveness suites, the harness enforces OS-level read sandboxes, stages isolated project configurations, and audits execution trajectories to detect and reject any attempt by the graded child to read forbidden answer files.

cmd/e2ebench · high confidence

Security

Hardened git command execution prevents repository-configured program execution

The internal git command wrapper now enforces strict security hardening to prevent repository configuration from executing arbitrary programs. It overrides dangerous settings (such as fsmonitor, hooks, and GPG programs) via baseline config, explicitly neutralizes repository-defined filter and merge drivers before content-affecting commands, and refuses command-line options that would load external code or configuration. Additionally, repository paths are resolved canonically once at open time to prevent nested .git files or symlinks from substituting the intended repository identity.

internal/platform/gitcmd · high confidence

New credential protection and redaction subsystem

The secrets package now enforces fail-closed protection for credential files and endpoint credentials. It introduces a built-in denylist that blocks read access to sensitive local files (such as SSH private keys, AWS credentials, and .netrc) and provides comprehensive redaction for URLs, environment variables, and command-line arguments to prevent secrets from leaking in diagnostics, exports, and logs. This change adds new security controls to the product's internal base layer. Also in this area: 6 commits (2 fixes) modifying internal.

internal/base/secrets · medium confidence · unverified

Prevent credential leakage on cross-origin redirects

The HTTP client now refuses to follow redirects that move to a different host, ensuring that credentials and authentication tokens are not sent to untrusted origins. This safety guard applies to requests carrying user-configured addresses, blocking cross-origin hops while still allowing safe same-origin redirects (including HTTP-to-HTTPS upgrades on default ports).

internal/safety/redirectguard · high confidence

Repository-declared MCP servers now require explicit user approval and re-verification on change

MCP servers declared in the workspace (e.g., via \.mcp.json\) are no longer started automatically. Instead, they are held in an 'awaiting decision' state until the user explicitly enables them. To prevent silent execution of modified code, the system now tracks a cryptographic digest of the server's declaration, including its command, arguments, environment variables, and any workspace files referenced by the command. If the repository changes any of these inputs, the server is disabled and requires the user to re-approve it. This approval is scoped to the specific project, so enabling a server in one project does not automatically enable a same-named server in another project within the same repository.

internal/contract/config · high confidence

Secure agent preset management with strict resource isolation and credential handling

The agent preset system now enforces strict security boundaries when managing agent configurations and their associated resources. User and project presets are stored with scoped authority, preventing project-level declarations from escaping their workspace or bypassing validation. Resource capture flattens symbolic links and junctions into static files, rejecting any links that point outside the source tree, target sensitive files (like .env or .git), or create cycles. Credentials in MCP plugin entries and environment variables are redacted in stored snapshots, retaining only references for live resolution, which prevents private tokens from being persisted. Additionally, the system validates retained resources against tampering by verifying file digests upon materialization and refuses to execute bindings that have been altered or lack proper local authorization.

internal/state/agentpreset · high confidence

Architecture

Provider contract layer refactored into modular, testable components

The internal provider contract has been restructured from a single monolithic file into a set of focused, modular packages. This change introduces dedicated modules for argument normalization (handling redundant envelopes), streaming chunk types (text, reasoning, tool calls, usage), client identity (Reasonix user-agent), configuration (including dynamic API key resolution), and error classification (context overflow detection). It also adds specific handling for OpenCode Go session routing, video and image data projection (stripping local metadata and handling unavailable assets), and tool call sanitization (pairing and deduplication). Comprehensive test coverage has been added for these new components to ensure correct serialization, projection behavior, and error handling.

internal/contract/provider · high confidence

Refactored terminal rendering into a dedicated package with improved diff and text handling

The terminal visual layer has been extracted into its own \termrender\ package, introducing robust support for Unicode grapheme clusters and ANSI escape sequences to ensure accurate text width calculations and layout. This refactor includes new capabilities for rendering unified diffs, including fenced code blocks (\ \\\diff \) and shell results, with optional support for external formatters like \delta\ via the \\[cli\].diff\_formatter\` configuration. Users will also benefit from improved clipboard handling that correctly falls back to OSC 52 in SSH sessions, and better visual consistency for boxed content and text wrapping.

internal/frontend/termrender · high confidence

Behavioural changes

1 commit (0 fixes) modifying (repo-wide)

A change to existing behaviour in (repo-wide) — 1 commit, 15 files: “bench(realrepo): long tasks over a pinned snapshot of this repository”.

(repo-wide) · medium confidence · unverified

1 commit (0 fixes) modifying internal/model/providerhealth

A change to existing behaviour in internal/model/providerhealth — 1 commit, 3 files: “feat(provider): add opt-in authorized connection health API”.

internal/model/providerhealth · medium confidence · unverified

13 commits (4 fixes) modifying internal/base/proc

A change to existing behaviour in internal/base/proc — 13 commits (4 fixs), 26 files: “perf(proc): resume a suspended launch through its own thread list”; “test(mcp): make the enable-binding tests hold on Linux”; “refactor: the kernel's packages live in the layer they belong to”; and 10 more.

internal/base/proc · medium confidence · unverified

14 commits (4 fixes) modifying internal/runtime/coordinator

A change to existing behaviour in internal/runtime/coordinator — 14 commits (4 fixs), 15 files: “feat: show how long each turn ran, measured by the kernel”; “fix(agent): enforce file access before previews”; “fix(agent): a plan check's own residue does not void it”; and 11 more.

internal/runtime/coordinator · medium confidence · unverified

14 commits (5 fixes) modifying internal/frontend/doctor

A change to existing behaviour in internal/frontend/doctor — 14 commits (5 fixs), 17 files: “refactor(sessionstore): a transcript's durable record leaves the agent”; “feat(shell): launcher arguments for a pinned shell (wsl.exe -d \<distro\> -- /bin/bash)”; “refactor: the kernel's packages live in the layer they belong to”; and 11 more.

internal/frontend/doctor · medium confidence · unverified

15 commits (3 fixes) modifying desktop/frontend-next/src/ui/document

A change to existing behaviour in desktop/frontend-next/src/ui/document — 15 commits (3 fixs), 31 files: “feat(desktop): the workbench's empty and failed states are cards, as the WB-Parts board draws them”; “feat(desktop): the compaction card, the execution group and delivered files lose their frames, as Chat-Cards-Light draws”; “fix(studio): name a property the sheets define”; and 12 more (refs \#15063).

desktop/frontend-next/src/ui/document · medium confidence · unverified

16 commits (8 fixes) modifying desktop/electron/assets

A change to existing behaviour in desktop/electron/assets — 16 commits (8 fixs), 16 files: “feat(studio): a flat, friendlier app icon, and the sidebar shows it”; “feat(serve): make the status icon a protocol rather than a binding”; “fix(studio): restore black-and-gold application and tray icons”; and 13 more.

desktop/electron/assets · medium confidence · unverified

165 commits (61 fixes) modifying desktop/frontend-next/src/design

A change to existing behaviour in desktop/frontend-next/src/design — 165 commits (61 fixs), 37 files: “feat(desktop): the plan is one card in the conversation, with the strip above the composer only while it is out of view”; “fix(studio): the pin button hides until its row is hovered, a pinned row keeps a marker, and the model menu's image badge no longer wraps”; “fix(studio): in a narrow conversation the composer's model pill shrinks the model name instead of letting the divider and reasoning level spill out of the pill and onto the send button”; and 162 more (refs \#15063, \#15040, \#15054, \#15050).

desktop/frontend-next/src/design · medium confidence · unverified

2 commits (0 fixes) modifying internal/ext/extension/providerconv, internal/ext/extension/providerext

A change to existing behaviour in internal/ext/extension/providerconv, internal/ext/extension/providerext — 2 commits, 8 files: “refactor: the kernel's packages live in the layer they belong to”; “refactor(agent): remove missing-tool-call-reasoning detection and recovery”.

internal/ext/extension/providerconv, internal/ext/extension/providerext · medium confidence · unverified

2 commits (0 fixes) modifying internal/runtime/taskcontract

A change to existing behaviour in internal/runtime/taskcontract — 2 commits, 4 files: “refactor(layers): contracts stop importing the evidence ledger”; “refactor(agent): stop sealing, shadowing and receipting each turn”.

internal/runtime/taskcontract · medium confidence · unverified

20 commits (10 fixes) modifying internal/state/stats

A change to existing behaviour in internal/state/stats — 20 commits (10 fixs), 24 files: “test(stats): every aggregate field declares its provenance”; “feat(usage): show how much is uploaded to model providers”; “fix(stats): the projection keeps a zero-amount quote”; and 17 more.

internal/state/stats · medium confidence · unverified

20 commits (8 fixes) modifying internal/runtime/usecap

A change to existing behaviour in internal/runtime/usecap — 20 commits (8 fixs), 39 files: “feat(plugin): pin MCP tool definitions at approval and hold changed tools”; “feat(composition): load preset resources on demand”; “fix(agent): give each child independent file observations”; and 17 more.

internal/runtime/usecap · medium confidence · unverified

26 commits (13 fixes) modifying internal/platform/worktree

A change to existing behaviour in internal/platform/worktree — 26 commits (13 fixs), 31 files: “feat(worktree): apply a candidate path by path against a workspace that moved on”; “fix(worktree): rebuild private index before capturing content”; “fix(gitcmd): resolve repository paths without rev-parse --path-format”; and 23 more.

internal/platform/worktree · medium confidence · unverified

3 commits (0 fixes) modifying internal/base/workspacefile

A change to existing behaviour in internal/base/workspacefile — 3 commits, 4 files: “test: every test directory is testenv.TempDir, and repolint refuses a bare one”; “Add structured file deliveries and workbench document previews”; “feat(workbench): preview Word and Excel files read-only”.

internal/base/workspacefile · medium confidence · unverified

3 commits (0 fixes) modifying internal/state/execjournal

A change to existing behaviour in internal/state/execjournal — 3 commits, 6 files: “refactor: the kernel's packages live in the layer they belong to”; “test: every test directory is testenv.TempDir, and repolint refuses a bare one”; “refactor(delegation): replace the task family with spawn\_agents”.

internal/state/execjournal · medium confidence · unverified

4 commits (2 fixes) modifying internal/base/netclient

A change to existing behaviour in internal/base/netclient — 4 commits (2 fixs), 9 files: “refactor: the kernel's packages live in the layer they belong to”; “fix(i18n): word the network test steps from codes so the English window stays English”; “test(netclient): pin the stream dialer's proxy boundary”; and 1 more.

internal/base/netclient · medium confidence · unverified

4 commits (2 fixes) modifying internal/model/providerbroker

A change to existing behaviour in internal/model/providerbroker — 4 commits (2 fixs), 7 files: “refactor(provider): remove the summary request flag and what only it served: the Request.Summary field and ErrSummaryToolsUnsupported, the adapters' no-tools branches, the broker's capability probe, and the test doubles and live test written for a summarizer call that no fold makes; the reported-usage effect test now detects a fold by its archive pointer and has a control arm”; “refactor: the kernel's packages live in the layer they belong to”; “fix(provider): stop attributing every refused body to dropped assistant reasoning”; and 1 more.

internal/model/providerbroker · medium confidence · unverified

5 commits (1 fix) modifying internal/runtime/recovery

A change to existing behaviour in internal/runtime/recovery — 5 commits (1 fix), 17 files: “refactor(sessionstore): a transcript's durable record leaves the agent”; “refactor: the kernel's packages live in the layer they belong to”; “fix(recovery): cut ArgsSummary on a rune boundary”; and 2 more.

internal/runtime/recovery · medium confidence · unverified

5 commits (2 fixes) modifying internal/base/filelock

A change to existing behaviour in internal/base/filelock — 5 commits (2 fixs), 7 files: “test(filelock): mirror relative-root rename admission”; “refactor: the kernel's packages live in the layer they belong to”; “refactor(delegation): replace the task family with spawn\_agents”; and 2 more.

internal/base/filelock · medium confidence · unverified

6 commits (2 fixes) modifying internal

A change to existing behaviour in internal — 6 commits (2 fixs), 415 files: “fix(ci): declare the picker's actions and move the peer tests that held a message to the switch-off case”; “feat(peers): with the dispatch switch on, a message from the user's other sessions reaches the model without a card”; “test: every test directory is testenv.TempDir, and repolint refuses a bare one”; and 3 more. Also in this area: New internal base utilities for file locking, network diagnostics, and search retrieval.

(repo-wide) · medium confidence · unverified

6 commits (2 fixes) modifying internal/platform/openwith

A change to existing behaviour in internal/platform/openwith — 6 commits (2 fixs), 4 files: “feat(openwith): say why an installed program is not offered”; “test(openwith): skip fakes of a POSIX disk on a Windows host”; “fix(openwith): resolve links before judging whether a program is inside a project”; and 3 more.

internal/platform/openwith · medium confidence · unverified

7 commits (0 fixes) modifying desktop/frontend-next/src/ui/surface/a2ui

A change to existing behaviour in desktop/frontend-next/src/ui/surface/a2ui — 7 commits, 10 files: “feat(studio): add nonblocking surface actions”; “feat(studio): add rich task workspaces”; “feat(studio): add editable drafts and text diff previews”; and 4 more.

desktop/frontend-next/src/ui/surface/a2ui · medium confidence · unverified

7 commits (1 fix) modifying desktop/frontend-next/src/ui/surface

A change to existing behaviour in desktop/frontend-next/src/ui/surface — 7 commits (1 fix), 18 files: “feat(studio): add nonblocking surface actions”; “feat(studio): add rich task workspaces”; “fix(studio): give the surface submit button AA contrast in the light theme”; and 4 more.

desktop/frontend-next/src/ui/surface · medium confidence · unverified

7 commits (3 fixes) modifying internal/state/mutation

A change to existing behaviour in internal/state/mutation — 7 commits (3 fixs), 33 files: “feat(mutation): admit explicitly opaque write effects”; “fix(mutation): reject ambiguous coordination record members”; “feat(proc): expose tracked root wait completion”; and 4 more.

internal/state/mutation · medium confidence · unverified

9 commits (4 fixes) modifying internal/frontend/capdiag

A change to existing behaviour in internal/frontend/capdiag — 9 commits (4 fixs), 13 files: “feat(rules): rules directories, delivered on the turn and scoped by paths”; “feat(shell): launcher arguments for a pinned shell (wsl.exe -d \<distro\> -- /bin/bash)”; “refactor: the kernel's packages live in the layer they belong to”; and 6 more.

internal/frontend/capdiag · medium confidence · unverified

98 commits (19 fixes) modifying desktop/frontend-next/src/i18n/msg

A change to existing behaviour in desktop/frontend-next/src/i18n/msg — 98 commits (19 fixs), 27 files: “fix(agent): a fold is described as the archive it is: the decline wording no longer says a summary is being paid for, /context and the TUI name an archived fold instead of calling it summarized or printing the raw mode, and the codes, modes and translations nothing produces any more are removed”; “feat(desktop): the workbench's empty and failed states are cards, as the WB-Parts board draws them”; “Show the cache hit rate after the cold start, and the output ratio”; and 95 more (refs \#15091, \#15096, \#15090, \#15040, \#15054, …).

desktop/frontend-next/src/i18n/msg · medium confidence · unverified

Centralized session artifact layout and remote path naming

The \internal/state/store\ package now serves as the single authority for on-disk persistence paths, consolidating session artifact sidecars (such as event logs, metadata, and checkpoints) that were previously derived independently across multiple packages. This change ensures that session listings correctly exclude internal sidecars (like \.events.jsonl\ and \.adjudication.jsonl\) to prevent phantom conversations from appearing in the user interface. Additionally, it introduces a new naming scheme for remote SSH module state, using a hash of the original workspace path to guarantee unique file slugs and prevent collisions between distinct workspaces.

internal/state/store · high confidence

Centralized settings contract with access classes and effect timings

The system now enforces a single, unified declaration for every user-configurable key, defining who may change it (Preference, Elevated, or Guarded), when the change takes effect (Immediate, NextTurn, OnOpen, or Restart), and whether the value is secret. Guarded keys are strictly refused by the model and require explicit user consent, while Elevated keys require a click even if proposed automatically. This centralization ensures that all frontend components and the model reference the same authoritative list of settings, preventing unauthorized model writes to sensitive areas like permissions, sandboxing, and provider access.

internal/contract/settings · high confidence

Computer use now relies on the CUA Driver with stricter action validation and effect grading

The internal computer platform has replaced its previous helper with the CUA Driver (pinned to version 0.34.1), introducing a new transport layer for native input and application control. This change enforces stricter validation on actions—such as bounding scroll amounts, limiting click counts, and requiring element references for specific operations like right-clicks—while introducing a graded effect system that distinguishes between confirmed, unverifiable, and suspected-noop outcomes based on driver evidence like value readbacks. The update also adds robust handling for modal interruptions, application launch attestation, and secure abort mechanisms via Unix sockets, ensuring that uncertain effects or refused operations are surfaced clearly for inspection rather than silently succeeding.

internal/platform/computer · high confidence

Durable session interruption tracking and turn admission control

The session control layer now persists a journal of 'barriers' (interruptions like user asks) to disk before they become answerable, ensuring that if the process crashes while waiting, the system can reconstruct the interrupted state on restart without losing the obligation or presenting it as answerable again. Additionally, a new admission guard strictly manages turn execution, classifying submissions as started, parked, or dropped (due to rotation, closing, or draining generation), and ensures that interrupted barriers are never included in the composed turn text sent to the model, preventing ghost obligations from persisting in conversation history.

internal/session · high confidence

Editor discovery now respects configured preference and known installer paths

The editor discovery logic now prioritizes a user-configured editor over automatic detection, ensuring that a specified editor is used if it exists on the system. If no editor is configured, the system searches for known editors (such as VS Code, Cursor, and VSCodium) by checking specific installer paths on Windows and macOS before falling back to PATH lookups. This change ensures that the executable installed by the editor's installer is preferred over potential PATH shims, providing a more reliable and predictable experience when opening workspaces in an external editor.

internal/platform/editor · high confidence

Enhanced file search with path-segment and subsequence matching

The file picker's search logic now supports matching against intermediate directory segments (e.g., typing "planind" finds files under a "planind" folder) and subsequence patterns (e.g., "proinf" matches "ProjectInfo"). These new match types are ranked below exact basename and directory name hits, and directories are prioritized in the results to allow direct navigation. The search also continues to exclude generated and vendor directories like \node\_modules\ and \build\ to keep results relevant.

internal/base/fileref · high confidence

Enhanced safety and reliability for the bash tool

The bash tool now includes robust pre-execution analysis to prevent destructive file operations, specifically refusing recursive deletes that target paths outside the workspace or use dynamic, non-literal arguments. It also enforces Windows command-line length limits to avoid OS-level launch failures and ensures that background jobs and foreground commands are properly cancelled and reaped, preventing orphaned processes. These changes improve security by containing file mutations to granted roots and increase reliability by handling edge cases like long commands and process lifecycle management.

internal/tools/builtin · high confidence

Extension dispatcher enforces strict intercept contracts and tighten-only permissions

The extension dispatcher now strictly validates extension rulings: answers that replace tool output are only permitted at the tool.before point, and any other usage is dropped with a warning. Additionally, extensions marked as 'tighten-only' are restricted to making permission decisions stricter (e.g., denying access) and are prevented from loosening host decisions or rewriting payloads, ensuring that untrusted or sandboxed extensions cannot override security policies.

internal/ext/extension/dispatch · high confidence

Goal evaluator now accepts verdicts from reasoning-only model responses

The goal completion evaluator now correctly interprets structured verdicts (complete, continue, blocked, uncertain) when the underlying model returns the answer exclusively within its reasoning stream rather than the main text output. This change introduces a shared bounded LLM call infrastructure that supports a 'reasoning answer' mode, allowing the evaluator to parse JSON verdicts from internal model thoughts when no final text content is produced, ensuring that models relying on reasoning-only outputs are no longer treated as failures or empty responses.

internal/model/boundedllm, internal/runtime/goaleval · high confidence

Graceful handling of unconfigured provider state

The application now supports a 'no provider' state when no provider is configured, allowing the runtime to start without failing. A new stub provider is registered for the unconfigured kind, which intercepts requests and returns a specific error indicating that no provider is available, rather than causing the kernel to refuse to start. This enables the user interface to display a connection guide or error message on the first turn instead of encountering a startup failure.

internal/model/noprovider · high confidence

Granular Bash and Browser permission controls with wrapper peeling and compound command decomposition

The permission system in internal/safety/permission now enforces stricter, more precise rules for Bash and browser tool calls. For Bash, commands are decomposed into segments so prefix rules (e.g., Bash(git push:\*) ) apply to each part independently, while dynamic or indirect executions (command substitution, eval, source, find -exec, inline interpreters) require explicit human approval or exact-match rules. Read-only classification is tightened: network-path operands, background execution flags, and file-output redirects are treated as non-read-only, and read-only git listings are explicitly recognized as reads. Wrapper peeling ensures that deny rules match through transparent wrappers (sudo, env, nohup, etc.), but allow rules do not widen to cover wrapped forms. For browser calls, credential entry and arbitrary script injection subjects require exact-match approvals and are not covered by site-level grants, preventing accidental authorization of sensitive actions.

internal/safety/permission · high confidence

Hardened shell command classification and recursive delete analysis

The safety engine now enforces stricter rules for classifying shell commands as read-only and analyzing recursive deletes. It prevents read-only escapes by detecting argument patterns that write or execute programs in tools like git, go, sort, uniq, and GitHub CLI (gh), and rejects gh calls where jq filters can reach the host environment. Bash redirect normalization is restricted to safe file descriptor duplications and null sinks (e.g., /dev/null), refusing to strip redirects that target real files. Recursive delete analysis (rm, find, cmd, xargs) now accurately identifies targets and recursion flags across wrappers like sudo and env, ensuring structural bypasses are closed.

internal/safety/shellsafe · high confidence

Hook system overhaul with improved execution, safety, and editing

The hook system has been significantly refactored to improve reliability, safety, and user experience. Command normalization now repairs malformed JSON-escaped hook commands (e.g., Node.js eval strings, PowerShell file paths) to prevent silent failures. Execution is more robust with platform-specific handling for Windows batch commands and better preservation of literal arguments and script content across different shells. Safety is enhanced by enforcing project hook approval, preventing project hooks from running without explicit trust, and ensuring that project settings are saved without overwriting unrelated configuration. A new dry-run feature allows users to test hook commands before applying them, providing immediate feedback on errors and blocking behavior. Additionally, the system now includes better diagnostics for inspecting hook configurations, detecting undefined environment variables, and suppressing repetitive warnings for broken hooks, while maintaining distinct identities for native hook configurations.

internal/ext/hook · high confidence

Introduce Standard and Minimal agent modes with strict project launch security

The agent composition system now provides two built-in modes: 'Standard' (日常搭档), which inherits all tools and configurations, and 'Minimal' (轻装搭档), which restricts capabilities to a terminal and optional image viewer. To enhance security, project-controlled agent compositions are now validated at execution time; they are strictly prohibited from overriding module arguments, URLs, environment variables, or headers, allowing only timeout adjustments. Additionally, the 'Minimal' mode supports optional tools, such as the image viewer, which gracefully degrades if unavailable.

internal/contract/agentcomposition · high confidence

Introduce per-turn language preference handling for reasoning and responses

The runtime now manages language preferences separately for visible reasoning and final answers on a per-turn basis. Users can explicitly set the language to Chinese (zh) or English (en), or leave it as auto. In auto mode, the system infers the reasoning language from the user's input script (e.g., Chinese characters trigger Chinese reasoning), while the response language follows a stable policy (auto commits to Chinese if the input is clearly Chinese, otherwise remains auto/English). The system injects transient context blocks (\<reasoning-language\> and \<response-language\>) into the turn without modifying the stable system prompt, ensuring that language preferences do not churn cached prefixes. This allows sub-agents to inherit preferences and ensures that code/technical terms remain in their original form regardless of the chosen language.

internal/runtime/langpref · high confidence

Introduces a structured, auditable pricing catalog with peak/off-peak billing and multi-currency support

The pricing subsystem now uses a formal catalog (catalog.go) to store official rate cards for DeepSeek, LongCat, MiMo, and Google Gemini, including dated rate steps and peak/off-peak windows (e.g., DeepSeek’s weekday peak hours and Chinese public holidays). The ledger (ledger.go) aggregates usage by pricing fingerprint and supports mixed-currency buckets, while coverage tracking (coverage.go) distinguishes complete, partial, and incomplete billing. Legacy telemetry is migrated (migrate.go) with clear completeness signals, and fixed-point money arithmetic (money.go) ensures precise cost calculations.

internal/contract/pricing · high confidence

Introduction of immutable Task Contract revisions with strict tightening policy

The runtime now enforces a canonical Task Contract model where accepted criteria (check commands and test requirements) are frozen and immutable. Users can only add new requirements (tightening); any attempt to drop or alter existing criteria is refused by default, ensuring that relaxing a contract remains an explicit user decision. The system supports versioned, digest-chained revisions stored in trusted state, with round-trip integrity checks to detect tampering.

internal/runtime/contract · high confidence

Linux desktop updates now use a hardened, Polkit-authorized .deb installer

Linux users can now receive desktop updates via a native .deb package installed through Polkit (pkexec). The new update helper validates the package identity and signature before installation, rejects symlinks and unauthorized inputs, and ensures the unprivileged desktop process never executes package management commands directly. This replaces the previous Windows-centric update path on Linux with a secure, audited installation flow that supports both Stable and Preview channels.

desktop/cmd/update-helper · high confidence

Linux packaging adds update authorization and cleanup scripts

The Linux package now includes a Polkit policy file to manage in-app update installation, requiring administrator authentication for the update helper, and an RPM post-uninstall script to properly clean up executable alternatives and AppArmor profiles when the application is removed.

desktop/packaging · high confidence

Localized compaction notices and configuration error messages

The interface now displays context-aware messages for context window compaction and configuration issues in the user's language. Compaction notices (e.g., summarization, degradation, native compression) now show specific details like token usage and compression mode, while failure reasons (e.g., summary timeout, hook refusal, context changes) are translated from typed codes. Additionally, when a settings save fails due to an unparsed configuration file, the error message now explicitly names the file path and line number, and guides the user on how to repair it, rather than just stating that the save failed.

desktop/frontend-next/src/i18n · high confidence

MCP plugin authentication errors now carry typed identity and cache entries persist after explicit connection

When an MCP server responds with a 401 Bearer challenge, the plugin now returns a typed \AuthRequiredError\ that clearly identifies the server, resource, and required scope, allowing the user to understand that a sign-in is needed rather than a generic retry. Additionally, when a server is explicitly connected via the \Add\ method, its handshake result is now persisted in the schema cache before the connection closes, ensuring that subsequent sessions can register the tools optimistically from the cache without waiting for a fresh network handshake.

internal/ext/plugin · high confidence

Market browsing now falls back to a cached copy when the registry is unreachable

The market client now maintains a local disk cache of the last successful anonymous browse response. If the registry becomes unreachable, returns a server error, or provides an unreadable response, the client serves this cached data instead of failing, marking the result with a note indicating it is stale and the reason it was served. This ensures that users can still browse the community market even during registry outages, while being clearly informed that the data is not current. The cache is strictly for anonymous browsing; authenticated actions like installing, publishing, or voting always require a live connection.

internal/ext/market · high confidence

New disk space warnings and improved notification reliability

The notification system now warns users when the workspace or state volume is nearly full, using a dedicated monitor that checks for low absolute or relative space and suppresses repeated alerts for six hours. On Windows, clicking a toast now correctly brings the Studio window to the foreground by handing the activation right to the shell before processing the click. Additionally, the system no longer silently discards notification send failures; errors are now reported to the host so they can be handled appropriately.

internal/platform/notify · high confidence

New structured frontmatter parser with namespace-aware lookups

The \internal/base/frontmatter\ package now introduces a canonical \Document\ type that preserves the original YAML structure, including nested namespaces and distinct empty shapes (mappings, sequences, bare keys). This replaces the previous flat-map approach for new consumers, allowing callers to use \Lookup\ and \Has\ to access fields by their full path without losing parent context. A \LegacyFlat\ compatibility view is retained for existing vocabularies, and a new \Decode\ function supports typed schema validation with optional strict field checking.

internal/base/frontmatter · high confidence

Projection database now caps WAL size and improves cross-platform reliability

The projection database (SQLite) now enforces a journal size limit on every connection to prevent Write-Ahead Log files from growing indefinitely. It also fixes cross-platform file URI generation to ensure Windows paths open correctly instead of silently falling back to memory. Additionally, the system now detects remote filesystems (NFS, SMB, etc.) and avoids using them for disk storage, while preserving database files when a newer schema version is detected.

internal/state/projectiondb · high confidence

Replace npm-based pre-push hook with Go-native validation

The pre-push git hook has been replaced with a Go-native implementation that runs 'go vet' and a repolint tool, removing the previous dependency on npm for local pre-push checks.

.githooks · high confidence

SSH agent integration and hardened authentication for remote connections

The remote platform now supports authenticating via the local SSH agent (using the standard SSH\_AUTH\_SOCK on Unix and the fixed Windows named pipe), allowing users to connect without storing private keys in the application. This change introduces OpenSSH-compatible behaviors such as the IdentitiesOnly flag to restrict agent keys to configured identities, and a fallback mechanism that attempts explicit identity files when the agent is unavailable or empty. Authentication logic has been hardened to prevent target credentials from leaking to ProxyJump hosts, and the SSH handshake is now bounded by a context deadline to prevent hangs on connections that do not support socket-level deadlines.

internal/platform/remote · high confidence

Scheduled runs now execute in supervised child processes with strict budget enforcement

The runtime now executes each claimed scheduled run in a separate child process rather than in-process. A Supervisor manages the child lifecycle, enforcing token ceilings and wall-clock limits by killing the process when budgets are exceeded. Communication between the supervisor and child occurs over standard streams using a JSON protocol, allowing the supervisor to track usage and enforce a hard kill if the child becomes unresponsive or crashes. This ensures that runs are isolated, metered accurately, and terminated reliably when they exceed their allocated resources. Also in this area: 6 commits (2 fixes) modifying internal.

internal/runtime/schedrun · medium confidence · unverified

Session-scoped isolation and stricter MCP security controls

The plugin specification layer now enforces tighter security and isolation for MCP servers. Process state is separated per session to prevent cross-session leakage, and absolute paths in local server commands and arguments are rebased to private session directories. Security is strengthened by refusing OAuth connections to servers that do not advertise PKCE support, unless the user explicitly opts out via their own personal configuration (project or package configs are ignored). Additionally, local stdio servers are now informed of writable folder boundaries to constrain file system access.

internal/ext/pluginspec · high confidence

Structured tool contract and approval policy definitions

The tool registry now exposes a stable, provider-visible contract snapshot (including schema, description, and read-only status) and enforces granular approval policies. Users will see clearer refusals when tools are disabled by configuration or when MCP tool definitions are held pending re-approval, and the system ensures that required tools cannot be removed even if explicitly requested.

internal/contract/tool · high confidence

Studio host runs as a loopback kernel with improved startup, language handling, and disk monitoring

The Studio host now operates as a dedicated kernel process behind a loopback HTTP socket, replacing the previous inline build model. This change introduces a background disk-space monitor that warns users when volumes are nearly full, ensures the kernel respects the configured desktop language (including explicit English for C/POSIX locales), and prevents the app from opening its own install folder as the first workspace. Startup performance is improved by sweeping session history in the background rather than during the initial build, and the host now logs elapsed milliseconds per startup phase for better observability.

cmd/reasonix-studio-host · high confidence

Unified rule parsing for allowed tools

A new shared parser in the rulelist package now handles splitting allowed-tools lists, accepting both commas and whitespace as separators while correctly preserving arguments inside parentheses. This ensures that skill front matter and the --allowed-tools flag interpret tool specifications consistently.

internal/base/rulelist · high confidence

Vision image attachments are now validated and resized before sending to models

A new \visionimage\ package enforces a maximum dimension of 1568 pixels for all images sent to AI models. When an attachment exceeds this limit, it is automatically downscaled while preserving aspect ratio; PNG and GIF files remain lossless, while JPEG and WebP files are re-encoded as JPEG at 85% quality. Images that cannot be decoded or exceed a 50-million-pixel decode budget are rejected with a specific error to prevent session failures caused by oversized formats.

internal/model/visionimage · high confidence

Windows system proxy bypass rules now correctly handled

The new sysproxy package on Windows now correctly interprets WinINET proxy bypass lists, supporting wildcards (e.g., \*.corp.local) and the implicit loopback bypass (e.g., 127.0.0.1, localhost) unless explicitly overridden with \<-loopback\>. This ensures that traffic to local or intranet addresses bypasses the configured proxy as expected by the OS settings.

internal/base/sysproxy · high confidence

Workspace identity and MCP launch authorization refactored

The system now derives a durable workspace identity that distinguishes between Git repositories and plain folders, ensuring that linked worktrees within the same repository share a single settings key while submodules and separate folders remain distinct. This identity is used by the MCP launch authorization manager to scope server launch grants and mutable launcher locks to the specific workspace, preventing cross-workspace interference. The change also introduces platform-specific process liveness checks for exclusive launch locks on Unix and Windows to prevent stale locks from blocking new launches.

internal/base/workspaceid, internal/ext/mcplaunch · high confidence

Fixes

1 commit (1 fix) fixing desktop/frontend-next/src/ui/kit

A fix in desktop/frontend-next/src/ui/kit — 1 commit (1 fix), 2 files: “fix(desktop): give every button the spec sizes”.

desktop/frontend-next/src/ui/kit · medium confidence · unverified

1 commit (1 fix) fixing internal/base/testscope

A fix in internal/base/testscope — 1 commit (1 fix), 2 files: “fix(testenv): isolate config locks from the real user home”.

internal/base/testscope · medium confidence · unverified

13 commits (9 fixes) fixing internal/platform/browser

A fix in internal/platform/browser — 13 commits (9 fixs), 33 files: “fix(browser): cut a long eval result on a rune boundary”; “fix(browser): a page the person opens answers when it is drawn, not when it loads”; “fix(browser): judge a click by where the pointer would really land”; and 9 more.

internal/platform/browser · medium confidence · unverified

13 commits (9 fixes) fixing internal/runtime/guardian

A fix in internal/runtime/guardian — 13 commits (9 fixs), 6 files: “fix(guardian): the reviewer sees the user's words and the agent's calls, nothing it was told”; “fix(agent): enforce file access before previews”; “refactor(provider): remove the summary request flag and what only it served: the Request.Summary field and ErrSummaryToolsUnsupported, the adapters' no-tools branches, the broker's capability probe, and the test doubles and live test written for a summarizer call that no fold makes; the reported-usage effect test now detects a fold by its archive pointer and has a control arm”; and 8 more.

internal/runtime/guardian · medium confidence · unverified

19 commits (11 fixes) fixing internal/model/anthropic

A fix in internal/model/anthropic — 19 commits (11 fixs), 28 files: “fix(anthropic): retain known usage when streaming is interrupted”; “fix(retry): say why a model request is retried, what it waits for, and expose idle\_timeout\_seconds”; “fix(providers): keep key-carrying provider requests on the configured origin”; and 16 more.

internal/model/anthropic · medium confidence · unverified

3 commits (2 fixes) fixing internal/platform/terminal

A fix in internal/platform/terminal — 3 commits (2 fixs), 13 files: “fix(ci): the Windows shell listing joins paths the way the OS it is computed for does, the save allocation check is skipped under the race detector that defeats regexp's pool, and the context-retrieval benchmark whose premise archived folds ended is removed with the surface helper only it used”; “fix(terminal): a Windows without a pseudoconsole is told so, not that the shell failed to start (\#15091)”; “feat(studio): the workbench gets a terminal of your own, as the Terminal board draws it” (refs \#15091).

internal/platform/terminal · medium confidence · unverified

5 commits (4 fixes) fixing internal/platform/environment

A fix in internal/platform/environment — 5 commits (4 fixs), 9 files: “refactor: the kernel's packages live in the layer they belong to”; “fix(git): give a session whose git did not answer one second attempt at its repository”; “fix(environment): attribute the reported shell to the bash tool”; and 2 more.

internal/platform/environment · medium confidence · unverified

6 commits (4 fixes) fixing internal/runtime/capability

A fix in internal/runtime/capability — 6 commits (4 fixs), 13 files: “refactor: the kernel's packages live in the layer they belong to”; “fix(capability): cut semantic-router candidate descriptions on a rune boundary”; “fix(capability): preserve MCP identities in catalog snapshots”; and 3 more.

internal/runtime/capability · medium confidence · unverified

7 commits (4 fixes) fixing internal/model/catalog

A fix in internal/model/catalog — 7 commits (4 fixs), 10 files: “fix(minimax): preserve discovered model ids”; “fix(providers): keep key-carrying provider requests on the configured origin”; “feat(providers): report why a whole-connection check failed as a typed finding”; and 4 more.

internal/model/catalog · medium confidence · unverified

9 commits (8 fixes) fixing internal/base/gitignore

A fix in internal/base/gitignore — 9 commits (8 fixs), 3 files: “fix(gitignore): read info/exclude from the common dir when .git is a file”; “refactor: the kernel's packages live in the layer they belong to”; “fix(gitignore): read core.excludesFile the way git does”; and 6 more.

internal/base/gitignore · medium confidence · unverified

Add tolerant JSON string decoder for inconsistent API error codes

Introduces the \flexstr\ package to handle JSON fields that may be sent as either strings or numbers by different API providers (e.g., OpenAI vs. DeepSeek). This allows the system to gracefully decode error codes regardless of whether they arrive as text (e.g., \"invalid\_request"\) or numeric values (e.g., \400\), preventing decode failures when server implementations disagree on the field type.

internal/base/flexstr · high confidence

Fix Windows uninstall registration for Reasonix

This change repairs how Reasonix registers itself in the Windows Add/Remove Programs list. It introduces a new \winuninstall\ package that manages the Windows registry entries for the application. The code ensures that the uninstall registration is correctly updated with the current version and install path, while also handling the migration from legacy Tauri-based installations by deleting outdated registry keys when appropriate. This fixes issues where the application might not appear correctly in the system's uninstall list or might retain stale uninstall information.

desktop/internal · high confidence

Fix workspace list state persistence and lock contention

The workspace list now correctly preserves loose entries when the state root moves by merging older copies into the current list exactly once, and updates that wait for a file lock will respect the caller's context deadline instead of hanging indefinitely. Also in this area: 6 commits (2 fixes) modifying internal.

internal/state/workspacelist · medium confidence · unverified

Fixes to legacy 1.x Desktop session migration routing and state adoption

The migration rescue process now correctly routes 1.x Desktop sessions into their original workspaces rather than dumping them into a single chosen project, and it preserves the archived status of conversations that were previously lost during the move. Additionally, the system now imports loose state entries (such as appearance settings and project lists) left behind by earlier storage relocations, ensuring a complete transition without overwriting existing data in the new state root.

internal/state/migration · high confidence

Improved reliability for OpenAI-compatible provider connections

This update enhances the stability and correctness of requests to OpenAI-compatible endpoints. It fixes an HTTP/2 protocol error with DeepSeek by holding request bodies until the server preface arrives, and ensures that stream cancellations reliably terminate connections even if the gateway goes silent. The model listing feature now supports custom authentication modes (such as x-api-key) and prevents credential leakage on cross-origin redirects. Additionally, reasoning effort overrides are now strictly validated against declared provider vocabularies, and empty or usage-only streams are correctly identified as upstream failures rather than valid model responses.

internal/model/openai · high confidence

Introduce fine-grained workspace write leases with contention reporting

The internal workspace lease mechanism has been replaced with a new system that allows multiple sessions to write to disjoint file paths in parallel, while serializing access to overlapping extents. Sessions now explicitly claim specific paths rather than holding the entire workspace, and when a session yields its claim (for example, while waiting for user input), it can report exactly which other sessions claimed the paths it gave up. The system also handles Windows path aliasing gaps, distinguishes between folder-scoped and repository-scoped claims, and provides detailed conflict and timeout errors that identify the holding session and the conflicting file paths.

internal/state/workspacelease · high confidence

Plugin package inventory and compatibility fixes

This update fixes several issues in the plugin package parser (internal/ext/pluginpkg). Agent directory scanning now correctly handles Unicode names, resolves symlink residency to prevent loading external files, and respects depth limits. Claude compatibility mapping now properly reports unsupported modules and filters out hooks that reference tools Reasonix does not support. Credential redaction is enforced in all export and inventory surfaces to prevent leaking secrets. Legacy v1 manifests are now rejected, with managed plugins auto-migrating to v2 and external ones quarantined. Tests cover these behaviors.

internal/ext/pluginpkg · high confidence

Scheduled runs now refuse execution when the target project or workspace is missing

The new \schedulerun\ package provides a standalone child process for executing scheduled runs, allowing the Studio host to trigger them without requiring the full command-line interface. As part of this change, scheduled runs now validate the existence of the target workspace and project trust status before execution; if the workspace is gone or the project is no longer trusted, the run is refused with a specific error cause (e.g., \CodeRootMissing\ or \CodeWorkspaceUntrust\) rather than proceeding or failing silently. This ensures that scheduled tasks only run against valid, confirmed, and accessible environments.

internal/frontend/schedulerun · high confidence

Secure MCP server setup with credential redaction and launcher-aware naming

The \internal/ext/mcpsetup\ package now provides a unified parser for adding MCP servers from JSON blocks, CLI commands, or URLs, ensuring that endpoint credentials are redacted in all display, preview, and export surfaces to prevent accidental leakage. It automatically derives clean server names from various launchers (such as npx, uvx, node, python, docker, and go) by stripping version pins, global options, and sensitive values, while correctly handling multi-line pasted commands and shell prompts.

internal/ext/mcpsetup · high confidence

Shell output now correctly decodes Windows code-page encodings

The shell runner now properly decodes non-UTF-8 output from Windows shells (such as PowerShell and Git Bash) that use the system code page (e.g., GB18030/GBK). Previously, these bytes were treated as invalid UTF-8 and replaced with replacement characters (U+FFFD), making error messages unreadable. The new implementation detects code-page bytes and decodes them into readable text, ensuring that users see the actual error details from failed commands rather than garbled placeholders.

internal/tools/shellrun · high confidence

Updated session1x test fixtures to include compaction summary and context sidecars

Added new test data files for the session1x boot assembly test, including a context JSON file that preserves the host-generated session-context sidecar and a compaction-summary message, alongside updated events and JSONL transcripts. This ensures the test suite validates that 1.x context sidecars are retained and their summaries are carried forward during session processing.

internal/assembly/boot/testdata/session1x · high confidence

Usage statistics now accurately distinguish billed costs, estimates, and incomplete quotes

The internal stats recording and query logic has been updated to provide more accurate usage reporting. Subscription-equivalent API estimates are now kept separate from billed costs, ensuring they do not inflate the total amount paid. Calls with no cost data or explicit pricing failures are correctly marked as incomplete rather than being treated as free or zero-cost. Additionally, the system now properly handles zero-amount quotes (where a call is priced at zero but still has a currency) and ensures that flush operations correctly report write failures to prevent data loss. Also in this area: 6 commits (2 fixes) modifying internal.

internal/state/usagecatalog · medium confidence · unverified

Test coverage

Add performance testing harness for Studio UI; Added comprehensive boot assembly tests for ablation, graph effects, and authority boundaries; Added conformance tests for extension lifecycle and protocol behavior; Added context-maintenance end-to-end benchmark; Added fixture suite for UI action and mutation analysis; Added live integration tests for the Electron desktop shell; Added relay harness for end-to-end testing of the desktop frontend; Added runtime-resume benchmark probe for graph state observation; Added test utilities for mocking LLM providers and handling summary contracts; Added tests for Surface validation and fallback logic; Added tests for ablation configuration parsing and arm naming; Added tests for sub-agent delegation lifecycle and capability routing; Automated enforcement of Studio release signing and CI security contracts; Introduce agent ablation testing and enforce immutable agent configuration; New benchmarks to measure the value of data-carrying dependency edges; New compaction benchmark to measure cost and fidelity of repeated folds; New evidence ledger tests for background leases, child merges, and completion adjudication; New i18n catalog parity tests and explicit C/POSIX locale handling; New project-check benchmark corpus for obligation derivation safety; New state management and test coverage for Reasonix Studio; New test coverage for frontend-next boot, actions, and error handling; New test environment utilities for isolation and stability.

Dependencies

Dependency updates across Go, Node.js, and benchmark fixtures

This release updates dependencies across the Go toolchain, Node.js packages, and benchmark fixtures. The Go module now targets Go 1.26.0 with toolchain 1.26.9, upgrading key libraries including golang.org/x/net to v0.60.0, golang.org/x/crypto to v0.57.0, and golang.org/x/sys to v0.48.0. The desktop Electron shell now uses Electron 44.5.1 and electron-builder ^26.15.3, while the frontend-next module upgrades to React 19.3.0, Vite 8.3.2, and TypeScript 7.0.2. The accounts worker pins esbuild to 0.28.1 and undici to 7.29.0 via overrides. Benchmark fixtures have been standardized to Go 1.21/1.22 and updated test dependencies.

(dependencies) · high confidence

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

How this codebase got here

Score

  • CAI 57 → 55 (-2.3)
  • Rubric changed (rubric-2026.10.5 → rubric-2026.11.3) — scores are not directly comparable.

Lenses

  • Code Health 74 → 70 (-4.3)
  • Architecture 96 → 92 (-3.5)
  • Maturity 65 → 64 (-1.3)
  • Readiness 64 → 64 (-0.3)
  • Security 74 → 80 (+5.6)
  • Accessibility 45 → 40 (-4.2)
  • Performance 74 → 77 (+3.5)

Resolved (56)

  • Change coupling: AddProvider.tsx ↔ EditConn.tsx (desktop/frontend-next/src/ui/AddProvider.tsx)
  • Change coupling: FeedbackForm.tsx ↔ FeedbackThread.tsx (desktop/frontend-next/src/ui/FeedbackForm.tsx)
  • Change coupling: en_kernel.ts ↔ refuse_common.go (desktop/frontend-next/src/i18n/en_kernel.ts)
  • Change coupling: en_kernel.ts ↔ routes.go (desktop/frontend-next/src/i18n/en_kernel.ts)
  • Change coupling: mock_hub.ts ↔ PhoneAccess.tsx (desktop/frontend-next/src/port/mock_hub.ts)
  • Change-coupling hub: actions.ts → preload.js, en_composition.ts, en_feedback.ts, host.ts, sse_boundary.ts, prefs.ts, Appearance.tsx, FeedbackMine.tsx, PhoneAccess.tsx, panetypes.ts, prefsnav.tsx (desktop/frontend-next/src/actions.ts)
  • Change-coupling hub: en_settings.ts → mock_boundary.ts, prefs.ts, prefsnav.tsx (desktop/frontend-next/src/i18n/en_settings.ts)
  • Change-coupling hub: kernel.ts → en_kernel.ts, boundary.ts, sse_boundary.ts, vite.config.ts, provider_edit.go, refuse_common.go, routes.go, themes.go (desktop/frontend-next/src/i18n/kernel.ts)
  • Critical CVE: [GHSA redacted] (android/gradle/verification-metadata.xml)
  • Critical CVE: [GHSA redacted] (android/gradle/verification-metadata.xml)
  • Critical CVE: [GHSA redacted] (android/gradle/verification-metadata.xml)
  • Documentation: written for insiders (examples/repo-map-kit/README.md)
  • Flaky test: workdir::example.com/summary.TestSummary
  • High CVE: [GHSA redacted] (android/gradle/verification-metadata.xml)
  • High CVE: [GHSA redacted] (android/gradle/verification-metadata.xml)
  • High CVE: [GHSA redacted] (workers/accounts/pnpm-lock.yaml)
  • High CVE: [GHSA redacted] (android/gradle/verification-metadata.xml)
  • High CVE: [GHSA redacted] (android/gradle/verification-metadata.xml)
  • High CVE: [GHSA redacted] (android/gradle/verification-metadata.xml)
  • High CVE: [GHSA redacted] (android/gradle/verification-metadata.xml)
  • …and 36 more

New (3845)

  • (anonymous) (cognitive 105) (desktop/frontend-next/tools/ui-census/verdicts.mjs)
  • (anonymous) (cognitive 18) (desktop/frontend-next/tools/ui-census/sites.mjs)
  • (anonymous) (cognitive 20) (desktop/frontend-next/tools/ui-census/schedulers.mjs)
  • (anonymous) (cognitive 22) (desktop/frontend-next/tools/ui-census/sinks.mjs)
  • (anonymous) (cognitive 25) (desktop/frontend-next/tools/ui-census/params.mjs)
  • (anonymous) (cognitive 27) (desktop/frontend-next/tools/ui-census/source.mjs)
  • (anonymous) (cognitive 44) (desktop/frontend-next/tools/ui-census/effects.mjs)
  • (anonymous) (cognitive 58) (desktop/frontend-next/tools/ui-census/state_dependency.mjs)
  • (anonymous) (cognitive 74) (desktop/frontend-next/tools/ui-census/state_lifecycle.mjs)
  • (anonymous) (cognitive 75) (desktop/frontend-next/tools/ui-census/state_identity.mjs)
  • (anonymous) (cognitive 92) (desktop/frontend-next/tools/ui-census/transport.mjs)
  • (anonymous) (cyclomatic 111) (desktop/frontend-next/tools/ui-census/state_identity.mjs)
  • (anonymous) (cyclomatic 17) (desktop/frontend-next/tools/ui-census/source.mjs)
  • (anonymous) (cyclomatic 19) (desktop/frontend-next/tools/ui-census/sinks.mjs)
  • (anonymous) (cyclomatic 20) (desktop/frontend-next/tools/ui-census/schedulers.mjs)
  • (anonymous) (cyclomatic 21) (desktop/frontend-next/tools/ui-census/sites.mjs)
  • (anonymous) (cyclomatic 24) (desktop/frontend-next/tools/ui-census/params.mjs)
  • (anonymous) (cyclomatic 32) (desktop/frontend-next/tools/ui-census/effects.mjs)
  • (anonymous) (cyclomatic 37) (desktop/frontend-next/tools/ui-census/state_dependency.mjs)
  • (anonymous) (cyclomatic 39) (desktop/frontend-next/tools/ui-census/state_lifecycle.mjs)
  • …and 3825 more

Changes since last survey

  • 300 commits — 181 feature/other, 119 fixes

By area

  • desktop/frontend-next — 180 commits
  • release-notes/studio — 51 commits
  • internal/assembly — 9 commits
  • internal/runtime — 8 commits
  • internal/frontend — 7 commits
  • internal/state — 7 commits
  • internal/session — 6 commits
  • (repo) — 4 commits
  • (root) — 4 commits
  • desktop/electron — 4 commits
  • internal/contract — 4 commits
  • .github/workflows — 3 commits
  • benchmarks/decision-model — 2 commits
  • internal/platform — 2 commits
  • .github/sponsor — 1 commit
  • android/app — 1 commit
  • benchmarks/context-retrieval — 1 commit
  • desktop/go.mod — 1 commit
  • internal/base — 1 commit
  • internal/ext — 1 commit

Notable commits

  • fix: build: Go 1.26.11, which fixes GO-2026-6859 in net/http
  • fix: fix(agent): a compact asked for by hand no longer answers "too little added" after a checkpoint; asking waives how much was added, and an unchanged context is still declined
  • fix: fix(agent): a fold is described as the archive it is: the decline wording no longer says a summary is being paid for, /context and the TUI name an archived fold instead of calling it summarized or printing the raw mode, and the codes, modes and translations nothing produces any more are removed
  • fix: fix(ci): declare the picker's actions and move the peer tests that held a message to the switch-off case
  • fix: fix(ci): the Windows shell listing joins paths the way the OS it is computed for does, the save allocation check is skipped under the race detector that defeats regexp's pool, and the context-retrieval benchmark whose premise archived folds ended is removed with the surface helper only it used
  • fix: fix(cli): a Cygwin or MSYS terminal is told why the UI cannot run there (#15091)
  • fix: fix(cli): every mcp subcommand prints usage for -h/--help and mcp import refuses arguments
  • fix: fix(control): a skipped question is marked as a skip before its answer reaches the turn: the turn that got the answer first could run on or end before the gate knew why, so it finished as cancelled, was handed a dismissal, or did not requeue the guidance queued into it, which showed as two intermittent failures under the race detector
  • fix: fix(control): in yolo mode forget no longer asks for approval; it archives a memory rather than destroying it, and remember still asks
  • fix: fix(control): in yolo mode remember no longer asks for approval either, so neither memory tool prompts under the mode that skips every other approval
  • fix: fix(desktop): Settings fills the window as the board draws it
  • fix: fix(desktop): keep the usage chart's hues in the design sheet
  • fix: fix(desktop): register the palette's scope chip action
  • fix: fix(desktop): the file-changes setting says the count sits at the top of the input box
  • fix: fix(desktop): the full-width option has its own words again, and a refused tool's failure line says the host's reason
  • fix: fix(desktop): the onboarding welcome step is centred with the wordmark, as the board draws it
  • fix: fix(electron): the instance probe and the kernel launch hide their console window on Windows
  • fix: fix(installer): a path-less computer-use driver no longer blocks an install or update
  • fix: fix(jobs): the view test waits for the live job's first line before it reads the running list: the job goroutine had not written it yet on a loaded runner, so the tail came back empty in one run in about two thousand
  • fix: fix(memory): the new save refusals carry wording in the desktop, and the remember freshness effect test states the mode it exercises
  • …and 280 more

Architecture

  • Containers 0 added · 0 removed · contexts 8 added · 1 removed · edges 13 added · 0 removed

Added bounded contexts (8)

-

  • benchmarks/realrepo/tasks/rt-cache-write-rate/solution
  • benchmarks/realrepo/tasks/rt-cache-write-rate/workdir
  • benchmarks/realrepo/tasks/rt-plan-ablation-module/solution
  • benchmarks/realrepo/tasks/rt-plan-ablation-module/workdir
  • benchmarks/realrepo/tasks/rt-rate-lookup-signatures/solution
  • benchmarks/realrepo/tasks/rt-rate-lookup-signatures/workdir
  • benchmarks/realrepo/tasks/rt-stale-skills/workdir

Removed bounded contexts (1)

  • app

Added dependency edges (13)

  • → benchmarks/realrepo/tasks/rt-cache-write-rate/solution
  • → benchmarks/realrepo/tasks/rt-cache-write-rate/workdir
  • → benchmarks/realrepo/tasks/rt-rate-lookup-signatures/solution
  • benchmarks/realrepo/tasks/rt-cache-write-rate/workdir → benchmarks/realrepo/tasks/rt-cache-write-rate/solution
  • benchmarks/realrepo/tasks/rt-plan-ablation-module/solution → benchmarks/realrepo/tasks/rt-cache-write-rate/workdir
  • benchmarks/realrepo/tasks/rt-plan-ablation-module/workdir → benchmarks/realrepo/tasks/rt-cache-write-rate/solution
  • benchmarks/realrepo/tasks/rt-plan-ablation-module/workdir → benchmarks/realrepo/tasks/rt-cache-write-rate/workdir
  • benchmarks/realrepo/tasks/rt-rate-lookup-signatures/solution → benchmarks/realrepo/tasks/rt-cache-write-rate/solution
  • benchmarks/realrepo/tasks/rt-rate-lookup-signatures/solution → benchmarks/realrepo/tasks/rt-cache-write-rate/workdir
  • benchmarks/realrepo/tasks/rt-rate-lookup-signatures/workdir → benchmarks/realrepo/tasks/rt-cache-write-rate/solution
  • benchmarks/realrepo/tasks/rt-rate-lookup-signatures/workdir → benchmarks/realrepo/tasks/rt-cache-write-rate/workdir
  • benchmarks/realrepo/tasks/rt-stale-skills/workdir → benchmarks/realrepo/tasks/rt-cache-write-rate/solution
  • benchmarks/realrepo/tasks/rt-stale-skills/workdir → benchmarks/realrepo/tasks/rt-cache-write-rate/workdir

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

Survey your own repository

esengine/DeepSeek-Reasonix was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.

About this page

  • The score is its most recent published measurement, taken on 13 November 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
  • Measured at commit cd20bb50be3b1eac5fad0648373a158c35767ae4 — the exact code this score is about.
  • Scored under rubric-2026.11.3 — the same rubric and the same method as every other entry in this index.
  • Measured by Canine Code Analyzer (cca.canine.dev) using cca measure-2026.11.61+preprod-f3449e6eb431.