- **Telegram remote questions** — receive and respond to GSD questions via Telegram bot alongside existing Slack and Discord channels (#645)
- **`/gsd quick`** — execute a quick task with GSD guarantees (atomic commits, state tracking) without the full planning overhead (#437)
- **`/gsd mode`** — workflow mode system with solo and team presets that configure defaults for milestone IDs, git commit behavior, and documentation settings (#651)
- **`/gsd help`** — categorized command reference with descriptions for all GSD subcommands (#630)
- **`/gsd doctor`** — 7 runtime health checks with auto-fix for common state corruption issues (#646)
- **Agent instructions injection** — `agent-instructions.md` loaded into every agent session for persistent per-project behavioral guidance (#437)
- **Skill lifecycle management** — telemetry tracking, health dashboard, and heal-skill command for managing custom skills (#599)
- **SQLite context store** — surgical prompt injection from structured knowledge base for precise context engineering (#619)
- **Context-window budget engine** — proportional prompt sizing that allocates context budget across system prompt sections based on relevance (#660)
- **LSP activated by default** — Language Server Protocol now auto-activates with call hierarchy, formatting, signature help, and synchronized edits (#639)
- **Extension smoke tests** — CI catches import failures, circular deps, and module resolution issues across all bundled extensions
- **`gsd --debug` mode** — structured JSONL diagnostic logging for troubleshooting dispatch and state issues (#468)
- **Worktree post-create hook** — run custom setup scripts when GSD creates a new worktree (#597)
### Fixed
- **CPU spinning from regex backtracking** — replaced `[\s\S]*?` regex in preferences parser with indexOf-based scanning (#468)
- **Model config bleed between concurrent GSD instances** — isolated model configuration per session (#650)
- **Onboarding wizard repeats** — skip onboarding for extension-based providers that don't require auth.json credentials (#589)
- **Session tool rebuild on cwd change** — tools now rebuild correctly when working directory changes mid-session (#633)
- **Auto mode state derivation after discussion fallthrough** — re-derives state to prevent stale dispatches (#609)
- **Milestone branch preservation on auto stop** — prevents work loss when stopping auto mode (#601)
- **Infinite loop when milestone detection silently fails** — `findMilestoneIds` now logs errors and warns instead of looping (#456)
- **Google Search OAuth fallback** — uses Google Cloud Code Assist API when `GEMINI_API_KEY` is not set (#466)
### Changed
- **Preferences wizard** — replaced serial flow with categorized menu for faster configuration (#623)
- **Slack remote questions** — brought to feature parity with Discord integration (#628)
- **YAML support in hooks** — hooks now support YAML configuration alongside JSON (#637)
- **Workflow visualizer** — `/gsd visualize` opens a full-screen TUI overlay with four tabs: Progress (milestone/slice/task tree), Dependencies (ASCII dep graph), Metrics (cost/token bar charts), and Timeline (chronological execution history). Supports Tab/1-4 switching, per-tab scrolling, auto-refresh every 2s, and optional auto-trigger after milestone completion via `auto_visualize` preference (#626)
- **Mid-execution capture & triage** — `/gsd capture` lets you fire-and-forget thoughts during auto-mode. The system triages accumulated captures at natural seams between tasks, classifies impact into five types (quick-task, inject, defer, replan, note), and proposes action with user confirmation. Dashboard shows pending capture count badge. Capture context injected into replan and reassess prompts (#512)
- **Dynamic model routing** — complexity-based model routing classifies units into light/standard/heavy tiers and routes to cheaper models when appropriate, reducing token consumption 20-50% on capped plans. Includes budget-pressure-aware routing, cross-provider cost comparison, escalation on failure, adaptive learning from routing history (rolling 50-entry window with user feedback support), and task plan introspection (code block counting, complexity keyword detection) (#579)
- **Feature-branch lifecycle integration test** — proves milestone worktrees branch from and merge back to feature branches, never touching main (#624)
- **Discord integration parity with Slack** — plus new remote-questions documentation (#620)
### Fixed
- **Absolute paths in auto-mode prompts** — write-target variables now passed as absolute paths, eliminating LLM path confusion in worktree contexts that caused artifacts written to wrong location and loop detection (#627)
- **Worktree lifecycle on mid-session milestone transitions** (#616, #618)
- **Eager template cache warming** — prevents version-skew crash in long auto-mode sessions (#621)
- **Milestone queue reorder** — `/gsd queue` supports reordering milestone execution priority with dependency-aware validation, persistent ordering via `.gsd/QUEUE-ORDER.json` (#460)
- **`.gsd/KNOWLEDGE.md`** — persistent project-specific context file loaded into agent prompts. New `/gsd knowledge` command with `rule`, `pattern`, and `lesson` subcommands for adding entries (#585)
- **Dynamic model discovery** — runtime model enumeration from provider APIs (Ollama, OpenAI, Google, OpenRouter) with per-provider TTL caching and discovery adapters. New `ProviderManagerComponent` TUI for managing providers with auth status and model counts (#581)
- **Expanded preferences wizard** — all configurable fields now exposed in the setup wizard, model ID validation, and `updatePreferencesModels()` for safe read-modify-write of model config (#580)
- **Comprehensive documentation** — 12 new docs covering getting started, auto-mode, commands, configuration, token optimization, cost management, git strategy, team workflows, skills, migration, troubleshooting, and architecture (#605)
- **`resolveProjectRoot()`** — all GSD commands resolve the effective project root from worktree paths instead of using raw `process.cwd()`, preventing path confusion across worktree boundaries (#602)
- **1,813 lines of new tests** — 13 new test files covering discovery cache, model discovery, model registry, models-json-writer, auto-worktree, derive-state-deps, in-flight tool tracking, knowledge, memory leak guards, preferences wizard fields, queue order, queue reorder E2E, and stale worktree cwd
### Fixed
- **Heap OOM during long-running auto-mode sessions** — four sources of unbounded memory growth: activity log serialized all entries for SHA1 dedup (now streaming writes with lightweight fingerprint), uncleaned `activityLogState` Map between sessions, unbounded `completedUnits` array (now capped at 200), and `dirEntryCache`/`dirListCache` growing without bounds (now evicted at 200 entries) (#611)
- **Stale worktree cwd after milestone completion** — three-layer fix: `escapeStaleWorktree()` at auto-mode entry, unconditional cwd restore in `stopAuto()`, and cwd restore on partial merge failure (#608)
- **Worktree created from integration branch instead of main** — `createAutoWorktree` reads integration branch from META.json, merge targets integration branch not hardcoded main (#606)
- **Milestone merge skipped in branch isolation mode** — branch-mode fallback detects `milestone/*` branch and performs squash-merge (#603)
- **`parseContextDependsOn()` destroys unique milestone ID case** — was lowercasing IDs, breaking dependency resolution (#604)
- **Tool-aware idle detection** — prevents false interruption of long-running tasks in auto-mode (#596)
- **Remote questions onboarding crash** — extracted `saveRemoteQuestionsConfig` into compiled src/ helper to avoid cross-boundary .ts import (#592)
- **Token optimization profiles** — `budget`, `balanced`, and `quality` presets that coordinate model selection, phase skipping, and context compression to reduce token usage by 40-60% on budget mode
- **Complexity-based task routing** — automatically classifies tasks as simple/standard/heavy and routes to appropriate models, with persistent learning from routing history
- **`git.commit_docs` preference** — set to `false` to keep `.gsd/` planning artifacts local-only, useful for teams where only some members use GSD
### Changed
- Updated Ollama cloud provider model catalog
### Fixed
- Native binary hangs in GSD auto-mode paths (#453)
- Auto-mode can be stopped from a different terminal (#586)
- Auto-mode worktree path resolution — prompt templates now include working directory, preventing artifacts from being written to the wrong location and causing infinite re-dispatches
- Auto-mode resource sync detection — gracefully stops when resources change mid-session instead of crashing
- Auto-mode missing import for `resolveSkillDiscoveryMode` causing crash on startup
- Auto-mode recovery hardened — checkbox verification falls through correctly, corrupt roadmaps fail verification instead of silently passing, atomic writes for completed-units.json, and task completion verified via artifacts not just file existence
- Auto-mode progress widget now refreshes from disk every 5 seconds during unit execution instead of appearing frozen
- Undo command now invalidates all caches (not just state cache), preventing stale results after undoing completed tasks
### Changed
- CI pipeline supports prerelease publishing with `--tag next` for testing before stable release
### Added
- Unit tests for auto-dashboard, auto-recovery, and crash-recovery modules (46 new tests)
- **8 new commands**: budget enforcement, notifications, and quality-of-life improvements (#441)
- **Preferences schema validation**: detects unknown/typo'd preference keys and surfaces warnings instead of silently ignoring them (#542)
- **Pipeline-aware prompts**: each agent phase (research, plan, execute, complete) now knows its role in the pipeline, eliminating redundant code exploration between phases (#543)
- **Research depth calibration**: three-tier system (deep/targeted/light) so agents match effort to actual complexity (#543)
### Changed
- Auto-mode decomposed into focused modules for maintainability (#534)
- Dispatch logic extracted from if-else chain to dispatch table (#539)
- v1 migration code gated behind dynamic import — only loaded when needed (#541)
- Background shell module decomposed into focused modules
- Unified cache invalidation into single `invalidateAllCaches()` function (#545)
### Fixed
- Executor agents now receive explicit working directory, preventing writes to main repo instead of worktree (#543)
- Merge loop and .gsd/ conflict auto-resolution in worktree model, `git.isolation` preference restored (#536)
- Arrow keys no longer insert escape sequences as text during LLM streaming (#493)
- YAML preferences parser hardened for OpenRouter model IDs with special characters (#488)
-`@` file autocomplete debounced to prevent TUI freeze on large codebases (#448)
- Auto-mode stops cleanly when dispatch gap watchdog fails (#537)
- Synchronous I/O removed from hot paths (#540)
- Silent catch blocks now capture error references for crash diagnostics (#546)
-`ctx.log` error in GSD provider recovery path fixed
- TUI resource leaks patched in loader, cancellable-loader, input, and editor components (#482)
- Hardcoded ANSI escapes replaced with chalk for consistent terminal handling (#482)
- **Session cwd update** — `newSession()` now updates the LLM's perceived working directory to reflect `process.chdir()` into auto-worktrees. Previously the system prompt was frozen at the original project root, causing the LLM to `cd` back and write files to the wrong location. This was the root cause of complete-slice and plan-slice loops in worktree-based projects.
- **Copy planning artifacts into new auto-worktrees** — `createAutoWorktree` now copies `.gsd/milestones/`, `DECISIONS.md`, `REQUIREMENTS.md`, `PROJECT.md` from the source repo into the worktree. Prevents plan-slice loops in projects with pre-v2.14.0 `.gitignore`.
- **Dispatch reentrancy deadlock** — `_dispatching` flag was never reset after first dispatch, permanently blocking all subsequent unit dispatches. Wrapped in try/finally.
- **`.gitignore` self-heal** — existing projects with blanket `.gsd/` ignore now auto-remove it on next auto-mode start, replacing with explicit runtime-only patterns so planning artifacts are tracked in git.
- **Discuss depth verification** — render summary as chat text (markdown renders), use ask_user_questions for short confirmation only.
- **Quiet auto-mode warnings** — internal recovery machinery (dispatch gap watchdog, model fallback chain) downgraded to verbose-only. Users only see warnings when action is needed.
- **Dispatch recovery hardening** — artifact fallback when completion key missing, TUI freeze prevention, reentrancy guard, atomic writes, stale runtime record cleanup
- **Discussion manifest** — mechanical process verification for multi-milestone context discussions
- **Session-internal `/gsd config`** — configure GSD settings within a running session
- **Model selection UI** — select list instead of free-text input for model preferences
- **Startup performance** — faster GSD launch via optimized initialization
### Changed
- **Branchless worktree architecture** — eliminated slice branches entirely. All work commits sequentially on `milestone/<MID>` within auto-mode worktrees. No branch creation, switching, or merging within a worktree. ~2600 lines of merge/conflict/branch-switching code removed.
- **`.gitignore` overhaul** — planning artifacts (`.gsd/milestones/`) are tracked in git naturally. Only runtime files are gitignored. No more force-add hacks.
- **Multi-milestone enforcement** — `depends_on` frontmatter enforced in multi-milestone CONTEXT.md
### Fixed
- **Auto-mode loop detection failures** — artifacts on wrong branch or invisible after branch switch no longer possible (root cause eliminated by branchless architecture)
- **Nested worktree creation** — auto-mode no longer creates worktrees inside existing manual worktrees, preventing wrong-repo state reads and "All milestones complete" false positives
- **Dispatch recovery hardening** — artifact fallback when completion key missing, TUI freeze prevention on cascading skips, reentrancy guard, atomic writes, stale runtime record cleanup, git index.lock cleanup
- **Worktree isolation for auto-mode** — auto-mode creates isolated git worktrees per milestone, with `--no-ff` slice merges preserving commit history and squash merge to main on milestone completion
- **Self-healing git repair** — automatic recovery from detached HEAD, stale locks, and orphaned worktrees
- **Worktree-aware doctor** — git health diagnostics and worktree integrity checks
- **Isolation preferences** — choose between worktree and branch isolation modes
### Fixed
- **Dispatch loop: parse cache stale data** — `dispatchNextUnit()` cleared path cache but not parse cache, allowing stale roadmap checkbox state to persist through doctor→dispatch transitions (#462)
- **Dispatch loop: completion not persisted after agent session** — `handleAgentEnd()` now verifies artifacts and persists the completion key before re-entering the dispatch loop, preventing re-dispatch when `deriveState()` sees pre-merge branch state (#462)
- **Dispatch loop: recovery counter reset without persistence** — loop-recovery and self-repair paths now persist completion keys and include a hard lifetime dispatch cap of 6 (#462, #463)
- **Dispatch loop: non-execute-task units had no artifact verification** — `complete-slice`, `plan-slice`, and other unit types now verify artifacts on disk before bail-out (#465)
-`@` file autocomplete debounced to prevent TUI freeze on large codebases (#452)
- Guard against newer synced resources from future versions (#445)
- Prevent `web_search` tool injection for non-Anthropic providers serving Claude models (#446)
- **Parallel tool calling** — tools from a single assistant message execute concurrently by default, with sequential mode as opt-in (`toolExecution: "sequential"`) and `beforeToolCall`/`afterToolCall` hooks for interception
- **Ollama Cloud** as model and web tool provider
- **Extensible hook system** for auto-mode state machine — post-unit hooks fire after unit completion
- **Event queue settlement** for parallel tool execution — extension `tool_call`/`tool_result` handlers always see settled agent state
### Changed
- Inline static templates into prompt builders, eliminating ~44 READ tool calls per milestone
### Fixed
- Auto-mode dispatch loop when `cachedReaddir` returns stale data after unit writes files
- Parse and path caches cleared alongside state cache after unit completion
-`bg_shell` hangs indefinitely when `ready_port` server fails to start — now transitions to error state with stderr context
- Em dash and slash characters in milestone/slice titles corrupting GSD state management
- Guided-flow self-heals stale runtime records from crashed auto-mode sessions on wizard start
- **URGENT: auto-mode loops on research-slice and plan-slice** — `handleAgentEnd` called `invalidateStateCache()` but not `clearPathCache()` or `clearParseCache()`. The in-process directory listing cache in `paths.ts` retained the pre-subagent empty directory snapshot, so `resolveSliceFile()` returned `null` for artifacts the subagent had just written. This caused `dispatchNextUnit` to re-dispatch the same unit (`research-slice` or `plan-slice`) instead of advancing, incrementing the dispatch counter until the `MAX_UNIT_DISPATCHES=3` limit triggered a hard stop with "Loop detected" (#421)
- Fix `npx gsd-pi@latest` failing with `ERR_MODULE_NOT_FOUND: Cannot find package '@gsd/pi-coding-agent'`. The loader now creates workspace package symlinks at runtime before importing, so it works even when `npx` skips postinstall scripts (#380)
- Hoist workspace package dependencies (undici, anthropic SDK, openai, chalk, etc.) into root `dependencies` so they install for end users. v2.10.10 removed `bundleDependencies` but didn't promote the transitive deps (#376)
- Add `undici` as root dependency to resolve startup crash (#372)
- Check `GROQ_API_KEY` before entering voice mode to prevent crash (#367)
- Publish verification checks `dist/loader.js` is non-empty (`-s`) and uses `--ignore-scripts` on `npm pack --dry-run` to match actual publish behaviour (#298)
-`optionalDependencies` in published `gsd-pi@2.10.4` were still pinned to `2.10.2`, causing users to install the broken engine binaries that 2.10.4 was meant to fix (#276)
- Native Rust TTSR regex engine — pre-compiles all stream rule conditions into a single `RegexSet` for one-pass DFA matching instead of O(rules × conditions) JS regex iteration
- Native Rust diff engine — fuzzy text matching (`fuzzyFindText`, `normalizeForFuzzyMatch`) and unified diff generation (`generateDiff`) via the `similar` crate, replacing the `diff` npm package
- LSP tool — full Language Server Protocol integration with diagnostics, go-to-definition, references, hover, document/workspace symbols, rename, code actions, type definition, and implementation support
-`/thinking` slash command for toggling thinking level during sessions
- Interactive wizard mode for `/gsd prefs` with guided configuration
- Startup update check with 24-hour cache — notifies when a new version is available
### Fixed
- TypeScript type errors across gsd, browser-tools, search-the-web, and misc extension files
- Milestone ID generation uses max-based approach instead of length+1 (prevents ID collisions)
- Non-thinking models handled correctly in `/thinking` command
- Auto-mode pauses on provider errors to prevent reassess-roadmap loop
- TAB hint displayed for notes input in discuss-mode survey
- Slice branches merge to integration branch instead of main
- Prefs wizard audit findings addressed
- Deduplicated maxNum logic with test coverage
- Command injection eliminated in LSP config `which()` function
- Unhandled JSON.parse in LSP message reader wrapped with error handling
- Vendor Pi SDK source (tui, ai, agent-core, coding-agent) into workspace monorepo under `packages/`, replacing the compiled npm dependency and patch-package workflow. Pi internals are now directly modifiable as TypeScript source.
- Existing patches (setModel persist option, Windows VT input caching) applied as source edits.
- Build pipeline runs workspace packages in dependency order before GSD compilation.
- Removed `patch-package` from devDependencies and postinstall.
- Proactive secret management — planning phase forecasts required API keys into a manifest; auto-mode collects pending secrets before dispatching the first slice
-`--continue` / `-c` CLI flag to resume the most recent session
### Fixed
- Doctor post-hook no longer preempts `complete-slice` dispatch
-`main_branch` preference restored; `runPreMergeCheck` implemented for merge safety
- Recovery/retry prompt injection capped to prevent V8 OOM on large sessions
-`.gsd/` excluded from pre-switch auto-commits to prevent squash merge conflicts
- Postinstall no longer triggers hidden `sudo` prompt on Linux — Playwright's `--with-deps` flag is no longer run automatically, preventing `npm install -g` from appearing to hang (#67)
- Auto-commit dirty files before branch switch to prevent lost work during slice transitions
### Changed
- Updated README to reflect current commands, extensions, and step mode workflow
