Skip to content

v0.5.0 — Agents pane: match the best of Claude Code & Codex

Latest
Latest

Choose a tag to compare

View all tags
@packetloss404 packetloss404 released this 04 May 06:07
· 819 commits to main since this release
v0.5.0
b9085e9

Added — Agents pane "match the best of Claude Code & Codex" initiative

Driven by a 6-agent deep-dive on Claude Code, Codex, Cursor, Windsurf,
Aider, Cline, Continue.dev, Zed, Copilot Workspace, JetBrains Junie, and
Warp; followed by a 4-agent deep-dive on the OpenAI "Codex for (almost)
everything" April 16 release + GPT-5.5 + CLI 0.107→0.128 cuts.

Tier 1 — visible polish

  • Drag-drop and clipboard-paste images in the launcher (5 MB cap, removable thumbnail chips); image blocks land in the SDK content array on send
  • SessionHealthBar in chat header: model · context % gauge · cumulative tokens · session $ · git branch
  • Mid-turn steering: Tab queues a follow-up; Alt+. / Alt+, nudge the model toward thorough / fast within the same provider
  • Shift+Tab cycles a single mode chip (default | plan | manual | yolo)
  • New slash commands /usage, /history, /review, /goal, /template; saved prompt templates surface as native /<slug> commands
  • Header context badges: provider auth, linked Mission with click-to-jump, MCP N/M server toggle dropdown, memory-context tooltip previewing the actual injected patterns
  • One-time onboarding overlay on first Agents-view visit

Tier 2 — killer features

  • Persistent dockable PlanPanel parsing Anthropic SDK TodoWrite and the markdown task_list tool
  • PendingApprovalsRollup with "Apply / Reject / Cancel all" when 2+ pending writes or permissions stack up
  • /review spawns a Reviewer subagent fed a unified diff of the parent conversation's pending writes — returns 🛑 Blockers / ⚠️ Concerns / 💡 Nits
  • Durable agent profiles (Default / Scout / Reviewer built-ins, plus user-created); AgentProfilesCard editor in Settings → Agent Profiles
  • AGENTS.md / CLAUDE.md auto-injection from the project root
  • Memories panel inline editor (edit text + category, Ctrl+Enter saves)
  • RunningAgentsChip in toolbar with live count of streaming agents, click-to-jump and stop

Tier 3 — sidecar protocol v3 → v4 + frontend

  • Sidecar PROTOCOL_VERSION bumped 2 → 4
  • New events: plan_block (structured TodoWrite mirror), tool_output_extended (Bash exit code + stdout/stderr; Write/Edit modified paths), turn_summary (running tokens between turns)
  • New requests: set_permission_mode, set_model, retry, cancel_pending_tools (drains parked permission/edit prompts as denied without killing the loop)
  • StartSessionRequest gains attachments and resume; EditResponseRequest gains mergedContent (per-hunk acceptance honored sidecar AND every in-process provider)
  • permission_request gains batchId/batchSize for grouped rollups
  • done payload gains resumeToken; persisted on the conversation
  • Auto-failover heuristic on rate-limit (Opus → Sonnet → Haiku, o3 → gpt-5 → o4-mini, MiniMax → highspeed) with a one-retry-per-turn guard
  • Worktree-per-conversation toggle in launcher (.pkt-worktrees/<convId> on a fresh pkt/<convId> branch)

Codex Spring 2026 absorption (A1–A5 + B1–B9)

  • Codex todo_list items map to the existing plan_block event so PlanPanel works for Codex too
  • reasoning_tokens + cached_input_tokens from usage flow through turn_summary and roll into aggregateConversationCost (was: under-reporting GPT-5.5 spend)
  • Codex MultiAgentV2 sub-agent attribution: turn_summary.address (/root/agent_a etc.) routes child tokens into a per-address bucket on the conversation; CostDashboard rolls every bucket into the total
  • AGENTS.md cascading resolver in Rust core (core::agents_md) walking ~/.claude/AGENTS{.override,}.md → git-root → cwd, picking one of AGENTS.override.md / AGENTS.md / CLAUDE.md per directory, concat with <!-- source: <path> --> headers, capped at 32 KiB. Honors CLAUDE_HOME env override for CI parity with Codex's CODEX_HOME
  • ProjectRulesCard in Settings → Project Rules reads + writes both AGENTS.md and CLAUDE.md on save; surfaces a Unify affordance when the two files diverge; offers a starter template when neither exists
  • Hover-+ Codex-App-style diff comments: per-line + button in ToolDiffView opens an inline composer; queued comments fold into the next user turn as a File comments: preamble
  • Smart-approval prefix-rule proposal: PermissionPrompt gains a fourth row "Always allow rule <pattern>"; one click writes the derived pattern into conversation.allowedTools
  • Composer-mode segmented control (Local / Worktree / Cloud) replaces the binary worktree toggle; persisted via localStorage
  • Right-rail tabbed mode (AgentTabbedRail) with Plan / Diff / Inspector tabs in a single 340 px column; toggleable from chat header
  • Persistent goals bridged to Missions: new goalStore + /goal slash command + goal-bound footer in PlanPanel (Pause / Resume / Complete) + 🎯 N badge per Mission row
  • LiveSpendChip in toolbar combining today's persisted total (analyticsStore) + live in-memory session $ across every open API conversation
  • Old-model pinning per profile via pinnedModel field; resolves as profile.pinnedModel ?? selectedModel ?? getDefaultModel(agent) at launch
  • Plan-with-Claude → Execute-with-Codex one-click handoff: PlanPanel "Hand off to Codex →" button when parent is Claude AND Codex auth is ready; spawns a fresh Codex conversation seeded with buildHandoffPrompt(parent) (distilled spec + plan + discussion summary, capped at 12 KiB); parentConversationId field wires a "← back to plan" link in the child's chat header

Follow-ups (F1–F10)

  • Auto-resume hydrated conversations: extracted listener block into installApiAgentListeners helper; sendMessage routes the first post-restart send through resumeApiConversation with the stored resumeToken
  • In-process providers honor mergedContent for per-hunk diff acceptance (parity with sidecar Anthropic)
  • Anthropic sidecar emits tool_output_extended (Bash exit code + stdout/stderr; Write/Edit modifiedPaths) and turn_summary (running per-message tokens for live SessionHealthBar updates)

Fixed

  • macOS title bar shows native traffic-light controls — config switched to decorations: true + titleBarStyle: "Overlay" + hiddenTitle: true; lib.rs setup hook strips decorations at runtime on Windows + Linux so the custom chrome stays the only chrome there. TitleBar.tsx detects macOS via userAgent, hides the Win-style min/max/close cluster, reserves 78 px of left padding for the traffic-light area
  • Standalone target/<profile>/packetade.exe reported "Sidecar down" — two stacked bugs:
    • Capability gate: app.shell().sidecar("node") is rejected by Tauri's permission layer unless an explicit shell:allow-execute entry lists node with sidecar: true (added in 74e6ba9)
    • Per-triple Node binary missing: Tauri's shell plugin on Windows resolves sidecar("node") to <exe_dir>/node-<target-triple>.exe, not generic node.exe; build.rs now copies binaries/node-<triple>.<ext> into the cargo output directory at compile time (added in 8f49083)

Removed

  • .github/workflows/{build,ci,release}.yml — builds and releases run locally; no GitHub Actions CI in this repo

Sidecar protocol

  • PROTOCOL_VERSION = 4. v4 added cancel_pending_tools request. v3 added typed attachments on start_session / send_message, mergedContent on edit_response, batchId/batchSize on permission_request, resumeToken on done, plus plan_block / tool_output_extended / turn_summary events. Old sidecars reply "Unknown request type" to v3+ requests; supervisor warns on version mismatch (does not refuse)
Assets 5
Loading