GitHub Releases

Changelog that follows the repo.

Release notes are pulled from GitHub Releases, then grouped by their Markdown sections so the public changelog stays aligned with what ships.

Subscribe or inspect the source.
Subscribe via RSSBrowse GitHub releases

MendCode v0.1.29

GitHub

Built from e96615d4fd75a6cf08f7aeff38b0a433a4c8aa41.

MendCode v0.1.28

GitHub

MendCode v0.1.28 adds durable first-class workflows and provider-aware image generation, strengthens detached background-task delivery, and improves workflow and session recovery across the TUI and server transports.

Added

  • Durable Workflow runs with validated declarative plans, policy gates, dependency-aware phases, background execution, scheduling, persistence, recovery, CLI/API routes, and a dedicated TUI monitor.
  • The image_gen tool with Codex OAuth, OpenRouter, and OpenAI-compatible adapters, persisted artifacts, optional captions, and safe image-edit inputs.
  • Workflow deletion plus richer progress, receipt, artifact, recovery, and terminal-outcome metadata in the server API and generated JavaScript SDK.
  • Focused workflow, image-generation, background-task, queue/compaction, server-route, scheduler, and TUI regression coverage.

Changed

  • First-class workflows now live in the dedicated Workflows view instead of being mixed into Agent View.
  • Workflow receipts, progress, queued work, terminal outcomes, and recovery state are presented consistently across TUI and server transports.
  • Image generation is configured independently from the active chat model and appears only when the provider contract and permissions support it.
  • Compaction, reconnect, long-session, error, and workflow-monitor presentation is clearer while the prompt remains focused during active work.

Fixed

  • Persist, replay, acknowledge, and instance-scope detached background-task completion notifications so idle owners resume reliably without duplicate or cross-project deliveries.
  • Keep the shared server alive while detached work is active and classify missing terminal outcomes as interrupted instead of leaving tasks indefinitely active.
  • Preserve queued prompts through compaction and reconnects, restore prompt focus after compaction, and clear stale working or disconnected indicators when execution settles.
  • Require workflow artifact consumers to depend on their producers, reject direct starts with unresolved gates, and improve recovery and cleanup of interrupted workflow runs.
  • Keep long-running non-shell tools alive across stream-idle periods and handle unknown image-generation results without crashing the session.

Tests and supply chain

  • Queue/compaction smoke confirmed against a local-only provider; no external provider was contacted.
  • Added targeted regressions for workflow plans, scheduler recovery, detached task wakeups, server lifetime, transcript state, prompt initialization, and workflow rendering.
  • The release passed supply-chain preflight, OSV scanning, frozen-lockfile installs, platform asset checks, SBOM generation, SHA-256 checksum generation, and provenance attestation.
  • The public installer was validated in an isolated non-interactive home and installed a binary reporting 0.1.28.

v0.1.27

GitHub

MendCode v0.1.27 improves TUI reliability, interrupt handling, prompt guidance, permission synchronization, and narrow-terminal rendering.

Highlights

  • Clear the frontend working indicator immediately when a user interrupts a running command, even if a stale unfinished assistant message remains cached.
  • Keep long Mermaid flowcharts within the terminal width and rotate layouts when the preferred orientation cannot fit.
  • Reconcile session permission-mode synchronization across transient local-server reconnects.

Added

  • Add compact, width-bounded Mermaid flowchart layouts for long vertical and horizontal diagrams, including automatic rotation when a horizontal diagram cannot fit.
  • Add explicit MendCode prompt guidance for task lifecycle discipline, cost-aware model routing, loop activation consent, background-subagent consent, custom tools/pages/widgets, and the supported TUI rendering surface.
  • Add regression coverage for interrupted prompt state, reconnecting permission synchronization, Agent View ordering, compact flowcharts, and the expanded prompt contract.

Changed

  • Expand the full prompt mode description with TODO, Loop Workflow, cost, TUI, and extension context while preserving the provider-focused harness contract.
  • Make loop activation and background delegation require explicit user intent when the request does not already provide it.
  • Bound normal and advanced Usage Insights fetch sizes separately and adapt heatmap cell widths and labels for narrow terminals.
  • Retry and reconcile session permission-mode synchronization across transient local-server reconnects instead of surfacing avoidable errors.

Fixed

  • Clear the TUI working indicator immediately after a user interrupts a running command, even when a stale unfinished assistant message remains in the frontend cache.
  • Order active Agent View rows by turn start time instead of later tool-update timestamps.
  • Keep long Mermaid flowcharts within the terminal width and avoid the unreadable fallback layout when the preferred orientation is too wide.

Tests & quality

  • Focused TUI, prompt-composition, permission-sync, and Markdown-rendering tests: 193 passed with 849 expectations.
  • bun typecheck passed.
  • Interactive queue/compaction smoke passed with 3 local mock requests and no external provider calls.
  • PR #228 passed repository guards, dependency review, deep security scanners, CodeQL, and JavaScript/TypeScript analysis.
  • Release builds passed artifact checks, SHA256 checksum generation, SBOM generation, and build-provenance attestation.

v0.1.26

GitHub

MendCode v0.1.26 restores reliable in-app update discovery and installation across TUI transport modes.

Fixed

  • Check for updates after the TUI is ready regardless of whether it uses the shared server, an external server, or a private worker.
  • Namespace the skipped-update preference so legacy upstream version state cannot suppress MendCode releases.
  • Run in-app installer upgrades without opening setup or modifying PATH, and surface the installer error when an upgrade fails.

v0.1.25

GitHub

MendCode v0.1.25 repairs cancellation and compacted-history state after interrupted or long-running TUI sessions.

Fixed

  • Finalize orphaned pending or running tool calls as cancelled when an unfinished assistant turn is explicitly stopped, preventing commands that no longer exist from remaining visibly active.
  • Preserve the latest completed compaction boundary while the bounded TUI message cache pages through long transcripts, so tool calls before compaction remain hidden unless explicitly revealed.
  • Add focused regression coverage for cancelling unfinished assistant turns with active tools.

v0.1.24

GitHub

MendCode v0.1.24 packages the current dev/main session and runtime recovery hardening from #218.

Added

  • Add focused regression coverage for shared-server lifecycle, reconnect recovery, session ordering, compaction boundaries, tool safety, and runtime cleanup.
  • Add queue/compaction smoke coverage for long-running sessions.

Changed

  • Harden shared-server lifecycle and client cleanup so reconnects do not leave stale runtimes or leases behind.
  • Preserve session, prompt, compaction, and tool state across reconnects, resumed work, and cancellation.
  • Improve diagnostics and safe handling for patch, edit, read, webfetch, and process-memory paths.

Fixed

  • Recover active sessions without duplicating or losing user/assistant messages.
  • Prevent stale queued work and session state from restarting after cancellation or transport recovery.
  • Keep cleanup and shutdown paths from leaving detached runtime instances behind.

Tests

  • Run focused regressions for TUI synchronization, session pagination, shared-server lifecycle, tool behavior, and release assets.

v0.1.23

GitHub

MendCode v0.1.23 hardens long-running sessions and Loop Workflows while making provider-aware compaction and TUI state easier to inspect.

Added

  • Add scheduler health, wake attempts, retry/backoff state, failure metadata, and safer CLI/API reporting to durable Loop Workflows.
  • Add regression coverage for session synchronization, virtual transcript windows, loop lifecycle, smart permissions, provider-aware compaction, reconnect recovery, and runtime/session cleanup.

Changed

  • Set GPT-5.6 ChatGPT OAuth models to a 256K effective input/context limit with a provider-specific 90% compaction threshold.
  • Improve loop runner recovery, scheduler error reporting, run leases, completion state, and persisted run metadata.
  • Improve shared-server lifecycle cleanup, transcript rendering, compaction summaries, Changes Review rendering, setup copy, and permission-review presentation.

Fixed

  • Prevent stale or duplicate session rows and user-message rendering during reconnects, transcript growth, compaction, and resumed work.
  • Preserve safe session and tool state when scheduled loop execution, cancellation, or a permission-review path fails.
  • Keep queued prompt/session state from restarting after cancellation or a recovered transport.
  • Make cross-platform release smoke tests validate the public Usage: help header.

Tests

  • Add focused regressions for session lifecycle, reconnect synchronization, compaction boundaries, tool safety, and release smoke behavior.

v0.1.22

GitHub

MendCode v0.1.22 makes startup updates visible for patch releases instead of silently attempting an automatic upgrade by default.

Added

  • Add regression coverage for the 0.1.20 to 0.1.21 update path.

Fixed

  • Show the startup Update Available modal for patch releases when autoupdate is omitted or set to notify.
  • Preserve silent patch upgrades for users who explicitly set autoupdate: true.

v0.1.21

GitHub

MendCode v0.1.21 is a reliability, interruption-recovery, and customization release. It adds project-owned prompt context and custom tool documentation, makes prompt/runtime state consistent across execution paths, protects active sessions during cancellation and reconnects, preserves useful media through compaction, and refreshes setup and customization guidance. This release contains the cumulative TUI, runtime, memory, Loop Workflow, setup, package, installer, documentation, security, and release-automation updates accumulated since v0.1.20.

Highlights

  • Project-local and package-shared prompt customization with bounded, root-safe loading and visible origin/status information.
  • Safer session interruption and recovery: explicit user cancellation is distinguished from connection loss, queued work cannot restart from stale notifications, and recovered terminal state reconciles with pending questions and permissions.
  • Shell execution now bounds cancellation cleanup, reports timeout or connection-loss results as unknown, preserves partial output, and gives safe retry guidance without encouraging blind repetition of destructive commands.
  • More stable TUI transcript scrolling, paging, compaction boundaries, assistant continuations, prompt ownership, and compacted tool-call presentation.
  • Vision-aware compaction media retention and Codex ChatGPT OAuth Responses Lite affinity rotation when effective prompt instructions change.

Added

  • A project-local custom Prompt Context mode backed by .mendcode/prompts/custom.md, including optional frontmatter names, bounded loading, UTF-8 normalization, project-root safety checks, and fallback diagnostics.
  • Package-shared custom prompt support through runtime packs without leaking a local project prompt into another project.
  • mendcode prompt inspection, build, and mode commands, with setup/status/TUI/runtime-plan visibility for prompt origin, name, path, byte count, and fallback state.
  • Public custom tool-call documentation covering project-local and package-shared tools, typed arguments, execution context, metadata, permission checks, abort handling, and package distribution.
  • Regression coverage for custom prompt parsing, package resolution, prompt snapshots, interruption recovery, reconnect synchronization, compaction media, OAuth affinity, transcript ordering, and shell retry metadata.

Changed

  • Prompt context is resolved from the active project root and carried as one snapshot through streaming and token estimation.
  • Custom prompt metadata stays out of provider instructions while a readable project-defined label is exposed in setup and the TUI.
  • Prompt-status scripts run through a Bash-compatible shell, and the public prompt route is exposed from the installed CLI.
  • Session scroll positions and anchors remain stable across transcript growth, paging, compaction, route remounts, and later assistant continuations.
  • Local shared-server fingerprints refresh from source metadata so TypeScript changes do not reuse stale runtime processes.

Fixed

  • Manual Esc cancellation can no longer be undone by late background-task owner-wake notifications or stale prompt deliveries.
  • Pending prompt deliveries are cancelled during interruption and concurrent abort requests are coalesced.
  • Hidden prompt input no longer remains mounted while a question or permission prompt owns session input.
  • Worker-backed terminal recovery now reconciles session state, pending questions, and permissions after system sleep or reconnect.
  • Queued session work no longer restarts after explicit cancellation.

Verification

  • Local TypeScript validation: bun typecheck from src/mendcode/packages/opencode.
  • Focused regression validation: 155 tests passed, 0 failed across the prompt, shell, interruption, session initialization, and session prompt suites.
  • CLI smoke validation: mend --version returned 0.1.21; mend --help completed successfully.
  • Release dry run 30509944842 passed Linux/Windows and both native Darwin builds, the complete 12-archive asset contract, checksums, release manifest, SBOM generation, and build provenance attestations.
  • Required repository checks passed, including CodeQL, dependency review, repository guards, and deep security scanners.

v0.1.20

GitHub

Maintenance release with the TUI, runtime, memory, Loop Workflows, setup, package, installer, documentation, and security updates accumulated since v0.1.19. See CHANGELOG.md for categorized details.

v0.1.19

GitHub

MendCode v0.1.19 Fixed Distinguish explicit user cancels from transport, WiFi, lid-sleep, or connection aborts so interrupted session state no longer falsely implies the user stopped the run. Preserve partial shell/task output with non-user-disconnect metadata instead of labeling it as User aborted the command

Fixed

  • Distinguish explicit user cancels from transport, WiFi, lid-sleep, or connection aborts so interrupted session state no longer falsely implies the user stopped the run.
  • Preserve partial shell/task output with non-user-disconnect metadata instead of labeling it as User aborted the command.
  • Prevent Loop Workflows from completing as 0/0, reject invalid fixed zero-turn budgets, and keep unbounded monitors open instead of storing maxTurns: 0.
  • Keep manual and signal-driven Loop Workflows from auto-running on every scheduler tick; only interval/adaptive loops become due from nextWakeup.
  • Let implementation loops run with normal execution when the objective explicitly asks to write, edit, fix, implement, code, or create files, while preserving explicit report-only and custom-gated modes.

Changed

  • Improve Loop Workflow tool/prompt guidance so assistants inspect existing loops before recreating them, never use maxTurns: 0, and report invalid completed 0/0 states instead of spawning duplicates.
  • Expand Loop Workflow tool metadata for TUI cards with permission mode, trigger, budget, model, objective, agent, and timestamps.
  • Add /loops dashboard warnings for legacy invalid zero-turn budgets.
  • Preserve unfinished assistant/tool phases through compaction and keep Herdr status accurate while local tools, child sessions, and loop workflows are still active.

Tests

  • Add regression coverage for non-user abort classification in shell, task, and message error handling.
  • Add Loop Workflow regressions for zero-turn budgets, due scheduling, permission inference, delete lifecycle, legacy/HttpApi route parity, and loop prompt guidance.
  • Add chat presentation regressions for fenced long-line wrapping, Unicode display width, text-fence wrapping, hex color handling, table preservation, and regex backtracking avoidance.

v0.1.18

GitHub

MendCode v0.1.18 focuses on goal-driven Loop Workflows, safer provider/runtime behavior, and cheaper GitHub Actions usage without removing the security gates. Built from efcf5f3bbef2d0c2611aac490cec4c76ceccc9bb.

Highlights

  • Loop Workflows now distinguish fixed iteration jobs, max-goal work, and unbounded monitors.
  • Goal loops can stop early when their checkpoint proves completion instead of spending every available turn.
  • Compaction keeps a bounded recent-tail snapshot so unfinished work survives summaries more reliably.
  • Provider auth can be disconnected from the Connect Provider dialog.
  • CI now uses fewer redundant Action jobs and shorter release artifact retention.

Added

  • Goal-driven Loop Workflow budget modes: fixed, max-goal, and unbounded-monitor.
  • Loop Workflow completion criteria, success checks, target turns, reserved verification turns, machine-readable checkpoints, and optional owner-session completion notifications.
  • Provider disconnect support from the Connect Provider dialog for locally saved provider auth.
  • TSM shortcut support for creating an explicit missing branch before opening it in a split MendCode pane.

Fixed

  • Fix Loop Workflow execution so implementation loops that explicitly allow edits are not downgraded to report-only by a conservative background service.
  • Keep loop session headers, run checkpoints, and the /loops dashboard aligned with the currently running iteration instead of lagging one completed run behind.
  • Keep Herdr loop-session panes in a working loop state between iterations and only report completion after the workflow reaches a terminal state.
  • Resume active provider compaction automatically and add trigger context so compaction summaries do not falsely close unfinished work.
  • Preserve a bounded snapshot of the retained recent tail during compaction so summaries keep the latest unfinished request, tool output, and resume context.

Changed

  • Treat maxTurns as an iteration budget for goal work instead of a requirement to spend every available loop turn.
  • Improve /loops timeline loading and scrolling so recent workflow events remain inspectable without overwhelming compact terminals.
  • Scope model variant selection to the newly selected model before deciding whether to open the variant picker.
  • Refresh opencode Go/OpenRouter reasoning variant mapping for GLM-5.2, MiniMax M3, and widely supported effort levels.
  • Clean up slash-command fallbacks and aliases so command names do not collide.

CI

  • Skip CodeQL for docs-only, changelog-only, and public asset-only PRs while keeping CodeQL on source changes, main pushes, scheduled runs, and manual dispatch.
  • Consolidate Security Guard scanners into one deep-scanner job while preserving gitleaks, OSV, zizmor, and Semgrep policy enforcement.
  • Reduce release workflow artifact retention from 30 days to 7 days to cut Actions storage pressure.

Tests

  • Added coverage for goal-budget loop completion/blocking, owner notifications, compaction tail snapshots, provider schema cleanup, reasoning variants, TSM/worktree shortcut edge cases, Herdr idle reporting, and structured-output schema transformation.

v0.1.17

GitHub

MendCode v0.1.17 Added Loop Workflows: durable workflow records, draft/activate/tick lifecycle, run journals, loop root sessions, Agent View grouping, terminal monitor, HTTP routes, SDK types, and a built-in loop assistant tool. Safe loop execution modes: dry-run ticks, --execute --report-only wakeups that deny mutation/shell/subagent tools, full execution for trusted contexts, and per-project OS service support for macOS LaunchAgent, Linux user systemd, and Windows Task Scheduler

Highlights

  • Loop Workflows: durable workflow records, draft/activate/tick lifecycle, run journals, loop root sessions, Agent View grouping, terminal monitor, HTTP routes, SDK types, and a built-in loop assistant tool.
  • Safe loop execution modes: dry-run ticks, --execute --report-only wakeups that deny mutation/shell/subagent tools, full execution for trusted contexts, and per-project OS service support for macOS LaunchAgent, Linux user systemd, and Windows Task Scheduler.
  • Built-in loop templates for PR watching, CI repair, research digests, and repo maintenance, plus CLI/status/monitor docs for loop lifecycle, daemon, service, logs, and troubleshooting.
  • Changes Review: a MendCode-native /changes TUI route for working-tree diffs with file navigation, diff-block navigation, line selection, reload, responsive layouts, inline comments, and stale-comment reconciliation.
  • Assistant-visible review context through bounded <mendcode_review_context> summaries and a review tool for current selection, file summaries, navigation, reload, comment creation, listing, and clearing.

Verify

  • Built from 8e354be833499ac8625da6fb7f23338ae06a0bd0.
  • Release workflow: https://github.com/MendCode/MendCode/actions/runs/28054534553

v0.1.16

GitHub

Patch release for Memory Center Dream scheduling. This keeps the v0.1.15 Memory Center work and fixes the accepted Dream proposal path so the Dream page actually becomes scheduled.

Highlights

  • Fix Memory Center Dream scheduling proposals so accepting a side-chat dream-dry-run proposal writes the Dream scheduler state instead of only saving an ordinary memory proposal.
  • Recover Dream scheduler state from already-applied Dream proposals, so users who accepted a Dream schedule proposal in 0.1.15 see the Dream page become scheduled after upgrading.
  • Add Dream schedule parsing for human time windows such as 6pm to 11pm local time, 18:00-23:00, and fixed times such as 21:00 Europe/London.
  • Add regression coverage for side-chat Dream proposal application, Dream schedule recovery, and human-readable Dream time parsing.

Added

  • Add Dream schedule parsing for human time windows such as 6pm to 11pm local time, 18:00-23:00, and fixed times such as 21:00 Europe/London.
  • Add regression coverage for side-chat Dream proposal application, Dream schedule recovery, and human-readable Dream time parsing.

Fixed

  • Fix Memory Center Dream scheduling proposals so accepting a side-chat dream-dry-run proposal writes the Dream scheduler state instead of only saving an ordinary memory proposal.
  • Recover Dream scheduler state from already-applied Dream proposals, so users who accepted a Dream schedule proposal in 0.1.15 see the Dream page become scheduled after upgrading.

v0.1.15

GitHub

Large Memory Center and terminal UX release. It adds the route-level memory workspace, Dream foundations, side-chat proposal flow, richer docs/screenshots, prompt/model metadata fixes, Markdown/Plan rendering upgrades, Herdr state integration, project MCP config, and release dependency hardening.

Highlights

  • Memory Center: Overview, Project memories, Global memories, Policy & categories, Dream, proposal review, and side chat.
  • Dream foundations: bounded evidence collection, logs, scheduler state, source manifests, missed-window handling, and safety reporting.
  • Terminal UX: Markdown/Plan rendering, Mermaid support, chat timeline polish, prompt chrome fixes, and Herdr agent-state reporting.
  • Workflow/runtime: project MCP config sync, package/setup coverage, worktree/mflow control-plane polish, and release dependency hardening.

Added

  • Add a route-level Memory Center in the TUI with Overview, Project memories, Global memories, Policy & categories, and Dream tabs.
  • Add saved/pending memory review surfaces with keyboard actions for applying, rejecting, editing, deleting, moving, and inspecting memories without leaving the terminal.
  • Add category-aware memory metadata, policy normalization, scope reasoning, evidence references, and reviewable proposal operations for verify, expire, recategorize, relink, demote-scope, and promote-scope flows.
  • Add a memory graph sidecar with legacy-entry bridging so memories can be grouped, indexed, and queried by category while preserving the existing flat memory files.
  • Add workspace-aware memory grouping so the Memory Center can show the current project plus other known project memory scopes.

Fixed

  • Make the prompt model, provider, and variant labels resolve from the selected prompt model metadata instead of stale local/session labels.
  • Preserve assistant provider/model/variant metadata when sync events arrive with partial message updates.
  • Keep prompt status script output tied to the current prompt-status identity so stale script output from an older model/session/root is not reused.
  • Replace the Plan Review renderer path with a stable styled Markdown component instead of relying on the experimental Markdown flag.

Changed

  • Label retrieved prompt memory by scope and category while preserving the existing <mendcode_memory> compatibility block.
  • Feed recent conversation context into post-turn memory extraction so proposals can distinguish durable facts from isolated one-line messages.
  • Track whether local model and variant state came from the user or session hydration, so explicit local changes win over older hydrated session state.
  • Persist submitted prompt agent/model selections onto the session before generation so resumed sessions and Agent View reflect the real active selection.
  • Improve Plan Review table rendering with grid mode, wrapped cells, and better handling for file/action/detail tables.

Docs

  • Add screenshot-backed public docs for the README hero, home layouts, Agent View, Plan Mode, and Memory Center.
  • Add docs/features.md as the product-facing feature map for README, website, screenshot, and demo work.
  • Add docs/memory-center.md and a Memory Graph/Dream spec to document the new memory workspace and safety model.
  • Refresh the README into a public product page with install paths, product surfaces, documentation map, development notes, security/community pointers, star history, and an expanded agent-context block.
  • Refresh the docs index, setup/configuration docs, customization docs, Plan Mode docs, Usage Insights docs, and package-sharing docs to point at implemented screenshots and provider-neutral examples.

Security and Release Integrity

  • Refresh release-gated dependencies and lockfile overrides for DOMPurify, Undici, form-data, esbuild, Slack Bolt, and OpenTelemetry so the OSV release scanner can pass without disabling supply-chain checks.

v0.1.14

GitHub

Package registry and setup release. It makes package installs target a selected package instead of an entire source, keeps source-level workflows explicit, and updates the release dependency gate.

Highlights

  • Make mendcode packages install <pack-id> install a selected package from the official or chosen registry catalog instead of applying a whole source by accident.
  • Keep source-level package application available as mendcode packages install-source <source-id> for existing local/team registry workflows.
  • Let the setup Package step skip packages, browse official packages, install a local package path, create/update a local package snapshot, or manage installed overlays.
  • Add registry coverage for multi-package catalogs, selected pack installation, source-level install compatibility, and runtime compatibility rejection.

Added

  • Add registry coverage for multi-package catalogs, selected pack installation, source-level install compatibility, and runtime compatibility rejection.

Fixed

  • Make mendcode packages install <pack-id> install a selected package from the official or chosen registry catalog instead of applying a whole source by accident.
  • Keep source-level package application available as mendcode packages install-source <source-id> for existing local/team registry workflows.

Changed

  • Let the setup Package step skip packages, browse official packages, install a local package path, create/update a local package snapshot, or manage installed overlays.

Security and Release Integrity

  • Update protobufjs to 7.6.3 so the release dependency gate can pass with the fixed version.