Agent Gitu v0.4.14

What's Changed in Agent Gitu v0.4.14

• Mobile Web & Responsive UI: Fixed Home/Cowork top-nav switcher geometry so it maintains a clean horizontal pill layout on mobile viewports.
• Company Workspace Modes: Added toggle to cleanly hide company icons and features when company mode is turned off.
• Cowork VCS & Git Support: Enabled git push and version control commands directly in Cowork mode without requiring delegation to the main agent.
• Coding Engine & Resiliency: Enhanced task loop resiliency, mobile UX, verification rigor, and safe Git operations.
• Pi Runtime Integration: Switched to Pi AgentSession loop with OAuth subscription sign-in/sign-out support.

Windows x64 Downloads

• Setup Installer: Agent-Gitu-Setup-0.4.14.exe
• Portable Executable: Agent-Gitu-Portable-0.4.14.exe

Verification Checksums (SHA-256)

2704349fce671bbe90210dee2317f2480ee9eca0b281c278eaba37ebf7d9eaae *Agent-Gitu-Setup-0.4.14.exe
27b48f73ab58160936621d0a91b4c5c516ab1f3933c3e1cee0cf48aee7a0029e *Agent-Gitu-Portable-0.4.14.exe

Agent Gitu v0.3.1

Agent Gitu v0.3.1 — quality fixes on top of the v0.3.0 architecture work.

Highlights

• 24 bug fixes across LLM, server, memory and tools; wired the memory learning loop
• v0.3.0: removed Atomics.wait, async memory lock, modular gitu.ts decomposition
• Fixed DeepSeek thinking-mode tool calls, vision gating and native streaming
• Restored the thinking/reasoning indicator on resumed runs
• Refined desktop UI: dark title bar integration, maximized launch, sidebar and grouped tool activity
• Improved connections and agent steering UX

Downloads

• Agent-Gitu-Setup-0.3.1.exe — Windows installer (NSIS, x64)
• Agent-Gitu-Portable-0.3.1.exe — Portable executable (x64)

Agent Gitu v0.2.1

Agent Gitu v0.2.1

Reliability hardening across the whole agent: specialists that checkpoint their work instead of losing it, dynamic auto-retry for flaky networks and provider outages, hybrid semantic retrieval, a vision-judged completion gate, and a three-phase UI/UX overhaul.

Highlights

Specialist reliability

• Mid-run checkpoints — every productive turn is committed on the specialist's branch
• Pause-not-discard — failed/partial attempts preserve their worktree + branch (durable index survives app restarts)
• Resume where it stopped — delegate {"resume":{"jobId":"sub-…"}} wakes a specialist exactly where it left off; the main agent knows which agent paused and how to wake it
• Generous budgets: base 30 turns, ceiling 150, progress-aware extensions

Dynamic LLM auto-retry

• Transport layer retries network errors (ECONNRESET, ETIMEDOUT, …) with jittered exponential backoff and Retry-After support
• Outer resilient wrapper (4 attempts, up to 45 s waits) around main-agent and specialist calls — visible recover … retry 2/4 in 4.2s events instead of silent deaths
• Selective by design: transient failures retry; auth/bad-request/app errors fail fast

Output quality

• Vision-judged completion gate — the model reviews the final screenshot before claiming UI work done
• IDF-weighted, identifier-aware retrieval + criteria-aware context packs; optional hybrid semantic search (embeddings)
• Built-in specialist roster: explore / backend / frontend / tester / reviewer / docs
• Truncated-JSON & empty-completion rescue retries; failure digests for long logs; 600 s command ceiling; rendered-DOM web_fetch

50+ bug fixes

Server security (double-resume race, DNS-rebinding, path scoping), git fingerprint/discard/parsing, LSP byte-framing, MCP spawn injection & bricked-client, apply_edit $ corruption, checkpoint resilience, CLI flags/hangs, and more.

UI/UX overhaul

Waiting-for-you state machine, SSE reconnect, ghost-bubble send retry, inline failure cards, keyless onboarding CTA, double-submit locks, tap-to-expand specialist cards, smart autoscroll, narrow-window panel overlay, global Esc handling, persistent error toasts, dark-theme dropdown fixes, accessibility polish.

---

Installers (Windows x64):

• Agent-Gitu-Setup-0.2.1.exe — standard installer (choose install dir)
• Agent-Gitu-Portable-0.2.1.exe — single-file portable, no install

Unsigned binaries: SmartScreen may show a warning — choose "More info → Run anyway".

v0.2.0 — LSP code intelligence

v0.2.0 — LSP code intelligence

Hermes now actively uses a Language Server Protocol layer during coding, not just exposes tools.

LSP intelligence layer

• 5 agent tools: lsp_diagnostics, lsp_definition, lsp_references, lsp_hover, lsp_symbols
• LspManager lifecycle: lazy server spawn (one per language, kept alive for the session), crash detection → single restart → clean fallback (never an infinite retry loop), graceful shutdown with full process-tree cleanup on Windows
• Configuration: .hermes/lsp.json overrides/extends the built-in registry (typescript, python, go, rust, c/cpp, java, csharp, css)
• Pull + push diagnostics with 1-based [ERROR] file:line:col — code message rendering
• Optional by design: when a server is unavailable or unsupported, Hermes falls back to search_files/read_file — LSP failure never fails a task

Hermes uses LSP during coding

• Automatic post-edit diagnostics gate: after every successful write_file/apply_edit, diagnostics for the touched files are surfaced to the model before the evidence gate
• Task-type → investigation strategy: the goal is classified (bug-fix / refactor / test-failure / explore / feature) and a proven LSP-first investigation sequence is injected at intake
• Prompt guidance: when to prefer LSP over blind text search, and the rule that "No diagnostics" is fast feedback — never a substitute for the real test/typecheck/build commands

Reliability

• 284 tests (up from 41 at the last release), including deterministic fake-LSP-server integration tests for lifecycle, kill-recovery, crash-once-restart-once, and fallback paths
• Windows spawn hardening: real executables spawn directly (no shell), only .cmd/.bat shims use a shell; taskkill /T /F tree cleanup

Installers

• Agent-Gitu-Setup-0.2.0.exe — NSIS installer (choose directory)
• Agent-Gitu-Portable-0.2.0.exe — portable build

Agent Gitu v0.1.9

Specialist Agent Delegation & Contract Clarity

v0.1.8: Orchestrator-Specified Criteria, Specialist Evidence Gates & Lying-Specialist Regression Tests

What's New in v0.1.8

1. Orchestrator-Specified Acceptance Criteria

• Added structured CriterionSpec contracts specifying required verification commands and evidence types (test_success, build_success, lint_success, etc.).
• Implemented exact normalized command matching (commandsMatch) in EvidenceEngine to prevent argument drift and command injection.

2. Execution-Authority & Hard Evidence Gate Enforcement

• Specialist Evidence Recording: Specialists running in isolated worktrees automatically record command verification into TaskLedger.
• Specialist claim_criterion: Added linking of command evidence against required acceptance criteria during delegation.
• Hard Gate on Answer: Specialist completion (answer / complete) is verified against EvidenceEngine.gate(). Unverified work is rejected and the worktree is discarded with no commit and no merge.
• Deterministic Prevention: Specialists running irrelevant commands (such as node --version) cannot satisfy verification criteria or merge changes into main.

Verification

• 178/178 tests passing across all 20 test files.
• TypeScript clean (tsc -p tsconfig.json).

v0.1.7

v0.1.7 — Worktree isolation hardening

Concurrency audit of specialist sub-agents running in isolated git worktrees — three real bugs found and fixed:

Fixes

• Merge race: a conflicting merge is now aborted inside the same serialized merge slot. Previously the abort ran outside the chain, so the next specialist's merge could start while conflict markers were still in the index (Merging is not possible because you have unmerged files).
• Conflict detection: conflict paths are read from the index before the abort wipes them — real conflicts are now reported as Merge conflict ... Conflicting paths: ... instead of a vague refusal.
• Poisoned merge chain: one failed merge can no longer silently skip every later specialist's merge.
• Fresh-machine identity: when no user.name/user.email is configured anywhere, both the worktree commit and the main-tree merge commit now fall back to a neutral author (Agent Gitu <agent@agentgitu.dev>). Previously only the worktree commit had a fallback — the merge still died with Committer identity unknown. Validated end-to-end with git identity fully unset.

Validation

• 156 tests pass (3 new: same-file collision, merge-after-failure, no-identity machine), typecheck clean.
• E2E on release build: two specialists colliding on one file → exactly one clean merge + one detected conflict; a disjoint-file specialist finishing after the failure merges cleanly; no leftover worktrees or hermes-sub-* branches; .hermes state never merged back.

Installers

• Windows x64 NSIS installer + portable build.

v0.1.6 — Audit Fixes: No Duplicate LLM Calls, Serialized Browser Actions, Worktree-Isolated Specialists

v0.1.6 — Audit Fixes: No Duplicate LLM Calls, Serialized Browser Actions, Worktree-Isolated Specialists

🔒 No more duplicate LLM requests

• A healthy stream that returns reasoning-only or empty content no longer triggers a second, duplicate (paid) completion request
• The non-stream fallback now only fires when the stream genuinely broke (connection died mid-stream) or the endpoint doesn't support streaming

🖥️ Serialized browser actions

• Parallel tool batches run independent tools concurrently, but browser actions are executed one at a time — a click/type/screenshot sequence can no longer race against itself
• Post-action verification: click, fill, select, and press now settle the page (wait for load or a quiet window) before reporting state, so the agent sees the post-action page instead of a mid-navigation snapshot

🏷️ Honest effort levels

• DashScope keeps genuinely distinct effort budgets (low 1024 → max 38912 thinking tokens)
• Generic OpenAI-compatible endpoints can't express "max" — the UI now labels it "max (= high)" instead of silently collapsing it
• New effortStyleFor / effortWireValue helpers make provider capability explicit in code

🌳 Worktree-isolated specialists

• Each specialist subagent now runs in its own git worktree (branch hermes-sub-*) — concurrent specialists can never clobber each other's files mid-task
• On success, changes are committed and merged back cleanly (private .hermes state is excluded)
• Merge conflicts abort the merge and mark the result failed instead of silently overwriting work
• Failed agents have their worktree discarded; non-git projects fall back to direct mode

✅ 153 tests passing (typecheck clean)

📦 Downloads

• Agent-Gitu-Setup-0.1.6.exe — NSIS installer (recommended)
• Agent-Gitu-Portable-0.1.6.exe — portable, no install required

v0.1.5 — Robust Function-Call Parsing, Queued Specialists, Usage Dashboard

v0.1.5 — Robust Function-Call Parsing, Queued Specialists, Usage Dashboard

🛠️ Three-Format Action Parsing

• Model replies now parse reliably from JSON, XML function calls (with streaming hold-back so partial tags never break the parser), and dots-style JSON where the tool name is the action type
• Up to 6 delegate/parallel tasks per step, with a background delegate mode that keeps the main loop moving

👥 Queued Specialist Sub-Agents

• Named worker agents run with a 5-worker concurrency pool; excess work stays visible as queued with live status updates
• Each agent keeps its own provider, model, and reasoning-effort setting

💰 Usage & Cost Dashboard

• Per-session panel with message count, input/cached/output tokens, and total USD cost
• Cost keeps accumulating across chat continuations of the same session

💻 Desktop & UI

• In-app browser back/forward now waits for the page to finish loading before screenshots
• Tool cards get per-kind icons with success/error states and animations; project sessions can be collapsed in the sidebar
• Edit-and-resend follow-ups can supersede the old message; optional auto-approve toggle

✅ 146 tests passing (typecheck clean)

📦 Downloads

• Agent-Gitu-Setup-0.1.5.exe — NSIS installer (recommended)
• Agent-Gitu-Portable-0.1.5.exe — portable, no install required

v0.1.4 — Persistent Code Index, LLM Retry, Workflow Switching

v0.1.4 — Persistent Code Index, LLM Retry, Workflow Switching

✨ New: Persistent Code Index

• SQLite-backed inverted index builds once, updates on file changes
• Background fs.watch keeps the index fresh between tasks
• Content-aware scoring finds relevant files by content, not just path

🔁 LLM Retry with Backoff

• Retries transient 429/5xx errors with exponential backoff + jitter
• Replaces raw JSON errors with actionable messages (rate limit, no credits, model denied, upstream 503)

🔄 Workflow Mode Switching

• Plan/Build/Chat dropdown now takes effect on continuations
• Switching a chat session to build (or vice versa) updates the durable session mode

🏷️ Free Model Labeling

• Free models show (free) in the model picker and sidebar

📦 Downloads

• Agent-Gitu-Setup-0.1.4.exe — NSIS installer (recommended)
• Agent-Gitu-Portable-0.1.4.exe — portable, no install required

🤖 Generated with Codebuff
Co-Authored-By: Codebuff noreply@codebuff.com