Two fixes for the model configuration bleeding between simultaneous
GSD instances that share the same global settings.json.
## Root Cause
1. `setDefaultModelAndProvider()` always persisted to `~/.gsd/agent/settings.json`
(global), so when either instance's interactive mode changed models (via
Ctrl+P or /model), it overwrote the other instance's saved default.
2. When auto-mode dispatched a new unit (after context wipe), if no
per-unit-type model preference was configured, the session picked up
the default from the now-contaminated global settings file.
## Fix 1: Project-scoped model persistence (settings-manager.ts)
`setDefaultModelAndProvider()`, `setDefaultModel()`, and `setDefaultProvider()`
now persist to project-level settings (`.pi/settings.json`) when a project
settings file exists, falling back to global only when no project context
is available. This prevents concurrent instances from overwriting each
other's model choice.
Added `hasProjectSettingsFile()` helper to detect project context.
## Fix 2: Auto-mode model capture (auto.ts)
Captures the session's model at auto-mode start (`autoModeStartModel`).
At each unit dispatch, if no model preference is configured for the unit
type, the captured model is re-applied with `persist: false`. This
ensures each auto-mode session maintains its own model regardless of
what other instances write to the shared settings file.
## Tests
3 new tests covering:
- Project settings file isolates model from global
- Two projects have independent model configs
- autoModeStartModel concept prevents model drift
All 448 existing tests pass.
Fixes#650
Tools (write, read, edit, bash) capture cwd at creation time via
createWriteTool(cwd), createReadTool(cwd), etc. When auto-mode
enters a worktree, process.cwd() changes but tools were not
recreated — they continued resolving relative paths against the
original project root.
This caused artifacts to be written to the main project's .gsd/
directory instead of the worktree's .gsd/ directory. The dispatcher
then couldn't find the artifact at the expected worktree path and
retried the unit indefinitely.
Fix: detect cwd change in newSession() and call _buildRuntime()
to recreate tools with the updated cwd. This is a targeted rebuild
that only fires when cwd actually changed (typically once per
auto-mode session when entering/exiting a worktree).
Fixes#633
LSP was never activated in interactive sessions because the default
active tools list hardcoded only read/bash/edit/write. This adds lsp
to that list and ships four new capabilities alongside edit sync and
stronger prompt guidance.
- Add "lsp" to default active tools in agent-session.ts
- New actions: incoming_calls, outgoing_calls, format, signature
- Wire edit/write tools to notify LSP clients on file changes
- Strengthen system prompt and GSD prompt with full LSP operation catalog
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: allow extensions to use 'yaml' and rework frontmatter parsing
* feat: add run-hook command for manual hook execution
* fix: sanitize slashes in unitType for runtime file paths
* fix: avoid native hangs in gsd auto paths
* fix: use .js extension in edit-diff.test.ts import for tsc compatibility
* fix: prevent OOM on large file diffs and implement context-line windowing
- Add size guard (MAX_DP_CELLS=4M) to buildLineDiff that falls back to a
linear-time prefix/suffix matching algorithm for large files, preventing
the O(n*m) DP table from causing OOM crashes
- Implement contextLines parameter in generateDiffString so only lines
within N lines of a change are rendered (with "..." separators), matching
unified diff behavior — the parameter was previously accepted but ignored
- Add tests for both context windowing and large-file fallback
---------
Co-authored-by: TÂCHES <afromanguy@me.com>
Add "success" to notify type union across ExtensionUIContext, interactive
mode, and RPC mode implementations. Fix null safety for readFileSync and
contextUsage.percent in auto.ts. Add discriminated union narrowing for
dispatch results. Add string type guards for select() return values in
commands.ts. Align ProviderErrorPauseUI notify signature. Simplify
AuthStorage return type.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
When auto-mode creates a worktree and chdir's into it, the Node process
cwd changes but AgentSession._cwd stays frozen at the original path.
Every newSession() builds a system prompt telling the LLM "Current
working directory: /original/path", so the LLM cd's back there and
writes files to the wrong location.
Update _cwd = process.cwd() at the start of newSession() so the system
prompt reflects the actual working directory after chdir.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* docs: add startup performance analysis and optimization plan
Profiled GSD CLI startup finding 2.2s for --version and ~3.8s for
interactive mode. Identified 5 root causes with measured timings and
created a phased optimization plan targeting <0.2s for --version
and ~0.8s for interactive startup.
* perf: speed up GSD startup with lazy loading and fast paths
- Fast-path --version/-v and --help/-h in loader.ts before importing
any heavy dependencies (2.2s → 0.15s, 14x faster)
- Lazy-load undici (~200ms) only when HTTP_PROXY env vars are set
- Skip initResources cpSync when managed-resources.json version
matches current GSD version (~128ms saved per launch)
- Lazy-load Mistral SDK (~369ms) on first API call instead of startup
- Lazy-load Google GenAI SDK (~186ms) on first API call instead of
startup
- Parallelize extension loading with Promise.all() instead of
sequential for-loop
---------
Co-authored-by: TÂCHES <afromanguy@me.com>
GitHub Copilot users with Claude models got 400 errors because the native
Anthropic web_search_20250305 tool was injected into requests to Copilot's
API proxy, which doesn't support it. The root cause was that model_select
never fires before the first API request on new sessions, so the fallback
heuristic (model name starts with "claude-") couldn't distinguish direct
Anthropic from proxied providers.
Fix: pass the resolved Model object through to the before_provider_request
event so extensions can check model.provider directly instead of relying on
model name heuristics.
Moves extension tool_call/tool_result interception from wrapToolsWithExtensions
(which fires inside the agent loop, bypassing event settlement) to
beforeToolCall/afterToolCall hooks that await _agentEventQueue. This ensures
extensions always see settled state — including the appended assistant message —
even when tools execute in parallel.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add Ollama Cloud (ollama.com) as a built-in provider with both model
hosting and web search/fetch capabilities.
Model provider:
- 13 curated models via OpenAI-compatible API (Llama 3.1, Qwen 3,
DeepSeek R1, Gemma 3, Mistral, Phi-4, GPT-OSS)
- Auth via OLLAMA_API_KEY environment variable
- Registered in onboarding, env hydration, and model resolver
Web tool provider:
- Search via POST ollama.com/api/web_search
- Page fetch via POST ollama.com/api/web_fetch (fallback after Jina)
- Added as third search provider option alongside Tavily and Brave
- /search-provider command updated with ollama option
Closes#430
When all credentials for a provider are exhausted, the system now
automatically falls back to the next available provider in a
user-configured fallback chain. Higher-priority providers are
restored automatically when their backoff expires.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Anthropic's 429 responses include retry-after and x-ratelimit-reset-*
headers that tell us exactly when to retry. Previously we ignored these
and used exponential backoff (2s, 4s, 8s), which is both wrong and
misleading in the UI countdown.
- Add retryAfterMs to AssistantMessage as the structured carrier
- Extract retry-after / x-ratelimit-reset-requests / x-ratelimit-reset-tokens
from Anthropic SDK APIError.headers in the provider catch block
- Session uses retryAfterMs when present (capped by maxDelayMs=60s),
falls back to exponential backoff for errors with no timing hint
The UI countdown now shows the actual Anthropic reset time. No UI changes needed.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat: replace pure-JS xxHash32 with native Rust implementation via napi
The hashline edit tool calls xxHash32 on every line of every file read/edit.
Moving this to a native Rust implementation (xxhash-rust crate) eliminates
JS overhead for this hot path. Hash output is identical -- verified by tests
comparing native vs JS reference across 11 input vectors including empty
strings, short/long inputs, unicode, and seeded variants.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: use typed native interface and remove version-drag comment in xxhash wrapper
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Replaces the multi-pass JS pipeline (TextDecoder → stripAnsi → sanitizeBinaryOutput)
in bash-executor.ts with a single native Rust call that handles UTF-8 decoding,
ANSI stripping, binary sanitization, and CR removal in one pass.
Key features:
- StreamState tracks incomplete UTF-8 and ANSI sequences across chunk boundaries
- Standalone stripAnsiNative() and sanitizeBinaryOutputNative() for use elsewhere
- Comprehensive test coverage for split multibyte, split ANSI, binary data
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Add the 1M context variant of Claude Opus 4.6 to the model registry
and fix model resolver to try exact match before glob detection, so
model IDs containing bracket characters (like [1m]) are not
misinterpreted as glob patterns.
* feat: add task isolation for subagent filesystem safety
Subagents can run in isolated git worktrees (or FUSE overlays on Linux)
so concurrent tasks don't stomp on each other's files. Changes are
captured as unified diffs and merged back via git apply.
- New isolation.ts module with worktree and FUSE overlay backends
- TaskIsolationSettings in settings-manager (mode + merge strategy)
- isolated parameter on the subagent tool schema
- Baseline capture/apply mirrors the parent repo's dirty state
- Process exit handler for best-effort cleanup of stale worktrees
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: correct delta capture to exclude parent baseline state
The worktree backend now commits a baseline snapshot after applying the
parent's dirty state, so captureDeltaPatch diffs only the subagent's
actual changes against the post-baseline HEAD (not the original HEAD).
The FUSE overlay backend tracks the parent's dirty file set at mount
time and filters the upper dir during delta capture to exclude inherited
dirty files.
Also removes dead code: findGitRoot (unused), readIsolationMergeStrategy
(exported but never called).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The async-jobs PR (#260) accidentally dropped `bashInterceptor` from the
Settings interface and the getBashInterceptorEnabled/getBashInterceptorRules
methods from SettingsManager, breaking the TypeScript build on main.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Fix cat rule to exclude heredoc syntax (cat <<EOF) via negative lookahead
- Fix write rule: exclude >> append and digit-prefixed fd redirects (2>)
using lookbehind (?<![|>\d])>(?!>)
- Add compileInterceptor() — pre-compiles rules once at construction time
instead of on every bash call; export CompiledInterceptor type
- Update createBashTool to use pre-compiled interceptor instance
- Add 33 unit tests covering all rules, edge cases, and pass-throughs
Regex-based pre-execution check in the bash tool blocks shell commands
(grep, cat, sed -i, etc.) when the dedicated replacement tool is available
in the session. Configurable via bashInterceptor settings.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Support multiple API keys per provider with automatic rotation:
- AuthStorageData accepts single credential or array per provider
- Round-robin selection across credentials (no sessionId)
- Session-sticky hashing when sessionId is provided
- Credential backoff on rate limits (30s), quota exhaustion (30min),
server errors (20s)
- markUsageLimitReached() backs off failing credential and returns
whether an alternate is available
- Login accumulation: duplicate provider logins append API keys
instead of replacing
- Agent retry handler tries credential fallback before counting
against retry budget (immediate retry, no delay)
- All getApiKey call sites thread sessionId for sticky selection
Backward compatible: single credentials work unchanged.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Move the edit tool's hot-path diffing operations from JS to native Rust:
- `normalizeForFuzzyMatch`: single-pass Unicode normalization (smart quotes,
dashes, special spaces, trailing whitespace)
- `fuzzyFindText`: exact-then-fuzzy substring search with UTF-16 index
conversion for JS compatibility
- `generateDiff`: unified diff generation using the `similar` crate
(Myers' algorithm with optimizations)
The Rust module at `native/crates/engine/src/diff.rs` exposes three napi
functions. The TypeScript wrapper at `packages/native/src/diff/` follows
the existing module pattern. `edit-diff.ts` now delegates to native
implementations while keeping line-ending handling and file I/O in JS.
18 tests covering normalization, fuzzy matching (including UTF-16 index
correctness with emoji/surrogate pairs), and diff generation.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The find tool's default path spawned `fd` and used the `glob` npm package
to discover nested .gitignore files. The native @gsd/native glob module
handles gitignore traversal natively via Rust's `ignore` crate, making
both dependencies unnecessary for this code path.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Check if the destination file exists before performing a move in
hashline-edit. If it does, return an error instead of silently
overwriting the file.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The delete operation in hashline-edit.ts wrapped both access() and
unlink() in a single try/catch. If access succeeded but unlink failed
(e.g., permissions), the error was silently swallowed and "Deleted" was
falsely reported. Now access and unlink have separate error handling:
access failures indicate the file doesn't exist, while unlink failures
propagate to the caller.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Implement hashline edit mode inspired by Oh My Pi's approach. Each line
in a file is identified by a content hash (xxHash32, 2-char nibble
alphabet), enabling the model to reference lines by stable LINE#ID tags
instead of reproducing full line text. This eliminates the most common
edit failure mode (slightly misquoted original text) and reduces output
tokens.
New files:
- hashline.ts: core hash computation, formatting, parsing, validation,
and edit application engine (pure JS xxHash32, no native deps)
- hashline-edit.ts: AgentTool wrapper for hash-anchored file edits
- hashline-read.ts: read tool variant that outputs LINE#ID:CONTENT format
- hashline.test.ts: 54 tests covering all core operations
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- config.ts: Replace execSync(`which ${command}`) with spawnSync("which", [command])
to prevent shell injection from malicious lsp.json config files
- client.ts: Wrap JSON.parse in parseMessage with try/catch and handle null messages
in the stream reader to prevent process crashes from malformed LSP output
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Tests initialize, hover, go-to-definition, references, document symbols,
diagnostics (type error detection), and clean shutdown against a real
typescript-language-server instance with a temp TypeScript project.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
All 10 LSP files ported and adapted. Wired into tools/index.ts.
Remaining work: fix TypeScript compilation errors (see below).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
In RPC mode, `ctx.ui.custom()` returns `undefined as never`, causing
`showInterviewRound` to return undefined and `Object.keys(result.answers)`
to throw TypeError.
When `showInterviewRound` returns undefined (RPC mode), fall back to
sequential `ctx.ui.select()` calls for each question, forwarding the
abort signal (#171) and supporting `allowMultiple` (#165).
- Add `allowMultiple` to `ExtensionUIDialogOptions`
- Widen `select()` return type to `string | string[] | undefined`
- Add `allowMultiple` to RPC select request and `values` array to response
- Update RPC `select()` to forward `allowMultiple` and parse array responses
- Guard existing `ctx.ui.select()` callers against the widened return type
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Three locations used lastIndexOf("/") or includes("/") for path
manipulation, which fails on Windows where paths use backslashes.
- auto.ts: writeBlockerPlaceholder directory extraction → dirname()
- interactive-mode.ts: parent directory traversal → path.dirname() loop
- path-utils.ts: non-null assertion on MSYS drive letter access
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
On Windows, LLMs convert absolute paths like F:\Projects\.gsd\... to
Unix-style /f/Projects/.gsd/... which Node's path.resolve interprets
as drive-root-relative, creating F:\f\Projects\.gsd\... instead.
Replace all *AbsPath template variables in prompt templates with
relative .gsd/... paths that resolve correctly on all platforms.
Add MSYS path normalization in resolveToCwd as defense-in-depth.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
LLM-generated commands with `> NUL` create undeletable files on Windows
because Git Bash treats NUL as a literal filename. Rewrite NUL redirects
to /dev/null at all three bash spawn sites.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
