- Separate system prompt for local vs cloud models via isLocal flag
(isLocalBackendURL, buildSystemPrompt(..., isLocal), propagate through
createSession/createSessionFromRecord/compactSession/spawnSubAgent)
- Increase MAX_EMPTY_RESPONSE_RETRIES from 1 to 3 for cloud model resilience
- Upgrade mouse input: full SGR-1006 parsing with col/row/pressed/modifiers,
logicalButton() helper, and copyToClipboard() via OSC 52
- Add temporary mouse debug logging in App.tsx
- Increase DEFAULT_MAX_ITERATIONS from 100 to 300
- Update README mouse/scrollback docs, tweak diff-remove color
Replaces the flat todo_write approach for non-trivial multi-step work with a
structured, incrementally-updated task store — dependencies, ownership, status,
metadata — ported from origin/v0.6.0 and adapted to the local Session/ctx.
- src/tools/task.ts: TaskStore (in-memory, per-session) + four read-only tools:
task_create (returns id), task_list (summaries), task_get (full details),
task_update (status pending|in_progress|completed|deleted, rename, claim via
owner, addBlocks/addBlockedBy dependency edges, merge-patch metadata — null
deletes a key). Self-refs/unknown ids ignored; direct 2-cycles skipped;
delete prunes dangling refs. Schemas declared before use (fixes a TDZ in the
v0.6.0 version). todo_write kept alongside for simpler cases.
- types.ts: ctx.taskStore. session.ts: TaskStore on every Session. loop.ts:
expose session.taskStore in the tool ctx + on sub-agent sessions.
- tools/index.ts: register the four task tools.
- task.test.ts: 9 cases (create/list/get/update status, dependency edges both
ways, self-ref/unknown/2-cycle guards, delete+prune, metadata merge-patch,
absent-store error).
- README: document the structured tasks.
Verified: typecheck clean, build 275.10 KB, 301 tests pass (+9).
The markdown export used to include only plain user/assistant text — every tool
call and tool result was dropped, so a shared transcript lost half the work.
- exportSession.ts: the markdown export now includes tool calls (fenced JSON),
tool results (fenced), fallback tool_result blocks, and multimodal user
content (text + [image attached] placeholders). Adds sessionToJson() — the
full record (messages verbatim + meta + exportedAt) for cross-machine
replay/sharing. exportSession() takes a format; defaultExportFilename() picks
.md/.json.
- App.tsx: /export json [file] selects the JSON dump; /export [file] stays the
markdown transcript. The export prompt remembers the chosen format.
- exportSession.test.ts: 8 cases (markdown includes tool calls/results, tool-
only assistant turn, JSON shape, filename extensions, file writes, auto-name).
- /help + README: document /export json.
Verified: typecheck clean, build 268.34 KB, 292 tests pass (+8).
Mouse-wheel tracking (xterm ?1000h) was always on so the wheel could scroll
the transcript, but it also captures mouse events, which prevents the
terminal's native text selection/drag-to-copy — a frequent ask.
Now mouse tracking defaults to OFF, so selecting/dragging output to copy
works out of the box. PageUp/PageDown still scroll (they always did). Toggle
the wheel back on with '/mouse on' (re-captures mouse events, so selection
won't work while on); '/mouse off' restores selection. '/mouse' shows the
current state. Documented in the /help list and README.
Verified: typecheck clean, build 257.55 KB, 263 tests pass.
- Add todo_write tool with live checklist rendering in the UI
- Add permission modes (default/plan/auto-edit/auto-accept) cycled via Shift+Tab or /perm
- Load CLAUDE.md/AGENTS.md as project instructions into the system prompt
- Forward sub-agent tool calls/results to parent UI panel
- Auto-compact context mid-turn and keep model moving with a synthetic user turn
- Prune older image_url parts to cap vision token cost
- Idle-abort guard for stalled streaming backends
- Improve background bash jobs, kill whole process trees on timeout/exit
- Expand hooks: command/http hooks, blocking, JSON outputSchema
- MCP: JSON schema → Zod conversion, parallel connection, duplicate server detection
- Add tests for loop, confirmFn, projectInstructions, readFile, bash, todoWrite, contextWindowCache
A day of dogfooding locode on itself surfaced several real bugs, most centered on
the backend request lifecycle silently hanging with no error surfaced to the user:
- bash/hooks: execa's shell:true defaulted to cmd.exe on Windows, which lacks
Unix tools (cat, sed, grep, ...) and fails a whole pipe with an unhelpful exit
255 if any stage isn't found. Added src/utils/shell.ts to resolve real Git Bash
when available.
- agent/loop: the backend client's own `timeout` only bounds time-to-first-byte —
once a streaming response starts, a backend that goes silent mid-stream (or one
that sends periodic content-less "heartbeat" chunks to keep the connection alive
through a proxy) hung the request forever with zero output. Added an idle-abort
guard that pokes only on chunks carrying real progress (usage/text/tool-calls/
finish_reason), covering the main stream, the non-streaming malformed-JSON
retry, and compactSession.
- agent/loop: a single tool-call-heavy turn (e.g. reading dozens of files) had no
auto-compaction safety net at all — shouldAutoCompact was only ever checked
between turns, never during one, so a long turn could blow past the context
window with no mid-turn correction. Now checked every iteration; re-anchors
mutationCommitLength afterward since compaction replaces session.messages
wholesale.
- agent/loop: hitting the iteration cap surfaced as an opaque "Request failed"
even when every prior tool call (including file edits) had actually succeeded.
Added MaxIterationsError, rendered as a calm "paused, send another message to
continue" notice instead, and raised the default cap 25 -> 50.
- agent/loop: removed a dead-code branch (fallback-malformed retry handling
placed inside a native-mode-only code path, so session.mode === "fallback"
could never be true there) that TypeScript correctly flagged as a type error.
- ui/App: text_done fires with an empty fullText for pure tool-call turns (no
preceding prose); the assistant-message push had no guard, rendering a blank
line every time. Now skipped when fullText is empty.
- ui/App+index: alternate-screen-buffer mode and StatusBar-above-ChatInput
layout, fixing the reported "prompt starts below the dashboard, jumps up on
space" glitch. Added input history (up/down recall).
- utils/writeFileAtomic: random temp-file suffix instead of a fixed `<file>.tmp`,
so concurrent writes to the same directory can't collide.
Also fixes two hardcoded version strings (cli.ts --version, mcp/client.ts's
self-reported MCP identity) that were left stale at 0.2.0 through the 0.3.0
bump, and a stale maxIterations default noted in the README.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Bump to 0.3.0.
- bash_kill tool lets the model terminate a still-running backgrounded
job (Ctrl+B); all running jobs are killed on process exit so they
don't outlive locode as orphans.
- Finished background jobs are now evicted from the registry after a
10-minute TTL instead of accumulating for the life of the session.
- The chat completion request timeout is now configurable
(requestTimeoutMs / LOCODE_REQUEST_TIMEOUT_MS, default 180000) so
backends that queue requests behind a concurrency limit (e.g.
Ollama's OLLAMA_NUM_PARALLEL) under multi-session load aren't forced
into a hard failure at a fixed 3 minutes.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Builds out locode's Claude Code plugin parity: MCP servers, slash commands,
sub-agents, hooks (12 lifecycle events), and skills, all loadable from a
local path or git URL. Also adds image support (read_file, /import), @-mention
file autocomplete, a /dashboard stats view, /export with an editable filename
prompt, an expanded git tool (reset/stash/merge/rebase/delete_branch), and a
redesigned status bar.
On top of that, a review pass found and fixed 10 correctness/stability bugs:
argument-injection in git reset/merge/rebase (a ref like "--hard" was parsed
as a flag), tool-call image results interleaving with native tool_call_id
messages and breaking OpenAI-compatible message ordering, backgrounded bash
jobs still being silently killed by their original timeout with the kill
masked as a clean exit, a SubagentStart hook's block being ignored, the
sub-agent timeout clock starting before the hook it should exclude, a
template-expansion bug that could re-substitute $1..$9 placeholders,
unbounded background-job output buffers, a missing directory-target fallback
in /export, and image size caps checked after reading the whole file instead
of before.
Session saves and exports now go through a shared atomic
write-then-rename helper so a crash can't leave a truncated file.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The 8-iteration-per-turn cap was hardcoded from before this session's feature work and
too tight for genuine multi-file tasks, causing "Max tool-call iterations reached" on
legitimately multi-step work. Raised the default to 25 and made it configurable via
`locode config set maxIterations <n>` / $LOCODE_MAX_ITERATIONS, mirroring the existing
contextWindow config pattern.
Also bumps the version to 0.2.0 (package.json, --version, and the MCP client's
self-reported identity) to reflect the substantial feature additions since 0.1.0.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Substantially expands locode's tool loop toward feature parity with Claude Code:
- web_search / web_fetch: DuckDuckGo-backed search and URL fetching (HTML stripped to text)
- git_status / git_commit: structured git tools with real diff/status/commit-list previews
before any mutating operation is confirmed
- agent tool: sub-agents run an isolated, headless tool loop sharing the parent's permissions;
bounded by an explicit depth guard (no nested sub-agents) and a 120s hard timeout, independent
of the toolset already excluding `agent` for sub-agent sessions
- MCP client support (stdio + streamable HTTP transports) via the official SDK: locode connects
to servers configured in `.mcp.json` or via `locode mcp add`, and exposes their tools alongside
the built-in ones, namespaced as `mcp__<server>__<tool>`
- Session persistence: conversations auto-save after every turn; `--continue`/`--resume`,
`locode sessions list/rm`, and an interactive resume picker
- Context compaction: tracks real token usage via `stream_options.include_usage` (falling back to
a char-based estimate), auto-detects the model's context window (Ollama/LM Studio native APIs,
cached, else a configurable default), and auto-summarizes the conversation at 85% usage or on
demand via /compact
- Status bar and welcome banner: added a context-usage indicator and fixed several dim-by-default
colors that were hard to read
Session/loop architecture changed to support all of this: Session now carries its own toolset
(tools + registry + OpenAI schemas) built from local + MCP tools, so runTurn no longer depends on
a fixed global tool list — the same mechanism that lets sub-agents inherit MCP tools while
excluding `agent` itself.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
kim
