agent-orchestrator/.claude/CLAUDE.md

118 lines
21 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

# BTerminal — Claude Behavioral Guide
## Workflow
- v1 is a single-file Python app (`bterminal.py`). Changes are localized.
- v2 docs are in `docs/`. Architecture decisions are in `docs/task_plan.md`.
- v2 Phases 1-7 + multi-machine (A-D) + profiles/skills complete. Extras: SSH, ctx, themes, detached mode, auto-updater, shiki, copy/paste, session resume, drag-resize, session groups, Deno sidecar, Claude profiles, skill discovery.
- v3 Mission Control (All Phases 1-10 Complete + S-1 Phase 1/1.5/2 Conflict Detection + Provider Adapter Pattern): project groups, workspace store, 12 Workspace components, session continuity, workspace teardown, file overlap conflict detection, inotify-based external write detection, multi-provider adapter pattern (3 phases: core abstraction, settings UI, sidecar routing), dead v2 component cleanup. 202 vitest + 42 cargo tests.
- v3 docs: `docs/v3-task_plan.md`, `docs/v3-findings.md`, `docs/v3-progress.md`.
- Consult Memora (tag: `bterminal`) before making architectural changes.
## Documentation References
- Architecture & decisions: [docs/task_plan.md](../docs/task_plan.md)
- Implementation phases: [docs/phases.md](../docs/phases.md)
- Research findings: [docs/findings.md](../docs/findings.md)
- Progress log: [docs/progress.md](../docs/progress.md)
- v3 architecture: [docs/v3-task_plan.md](../docs/v3-task_plan.md)
- v3 findings: [docs/v3-findings.md](../docs/v3-findings.md)
- v3 progress: [docs/v3-progress.md](../docs/v3-progress.md)
## Rules
- Do not modify v1 code (`bterminal.py`) unless explicitly asked — it is production-stable.
- v2/v3 work goes on the `v2-mission-control` branch, not master.
- v2 architecture decisions must reference `docs/task_plan.md` Decisions Log.
- v3 architecture decisions must reference `docs/v3-task_plan.md` Decisions Log.
- When adding new decisions, append to the Decisions Log table with date.
- Update `docs/progress.md` after each significant work session.
## Key Technical Constraints
- WebKit2GTK has no WebGL — xterm.js must use Canvas addon explicitly.
- Agent sessions use `@anthropic-ai/claude-agent-sdk` query() function (migrated from raw CLI spawning due to piped stdio hang bug). SDK handles subprocess management internally. All output goes through the adapter layer (`src/lib/adapters/claude-messages.ts` via `message-adapters.ts` registry) — SDK message format matches CLI stream-json. Multi-provider support: message-adapters.ts routes by ProviderId to provider-specific parsers (currently only Claude registered).
- Sidecar uses per-provider runner bundles (`sidecar/dist/{provider}-runner.mjs`). Currently only `claude-runner.mjs` exists. SidecarManager.resolve_sidecar_for_provider(provider) finds the right runner file. Deno preferred (faster startup), Node.js fallback. Communicates with Rust via stdio NDJSON. Claude CLI auto-detected at startup via `findClaudeCli()` — checks ~/.local/bin/claude, ~/.claude/local/claude, /usr/local/bin/claude, /usr/bin/claude, then `which claude`. Path passed to SDK via `pathToClaudeCodeExecutable` option. Agents error immediately if CLI not found. Provider env var stripping: strip_provider_env_var() strips CLAUDE*/CODEX*/OLLAMA* vars (whitelists CLAUDE_CODE_EXPERIMENTAL_*). Dual-layer: (1) Rust env_clear() + clean_env, (2) JS runner SDK `env` option. Session stop uses AbortController.abort(). `agent-runner-deno.ts` exists as standalone alternative runner but is NOT used by SidecarManager.
- AgentPane does NOT stop agents in onDestroy — onDestroy fires on layout remounts, not just explicit close. Stop-on-close is handled externally (was TilingGrid in v2, now workspace teardown in v3).
- Agent dispatcher (`src/lib/agent-dispatcher.ts`) is a singleton that routes sidecar events to the agent store. Provider-aware: sessionProviderMap routes messages through adaptMessage(provider, event) from message-adapters.ts. Handles subagent routing (project-scoped sessions skip layout pane, render in TeamAgentsPanel; detached mode creates layout pane). Session persistence via registerSessionProject(sessionId, projectId, provider) + persistSessionForProject() (saves state + messages to SQLite on complete).
- AgentQueryOptions supports `provider` field (defaults to 'claude', flows Rust -> sidecar), `provider_config` blob (Rust passes through as serde_json::Value), `permission_mode` (defaults to 'bypassPermissions'), `setting_sources` (defaults to ['user', 'project']), `system_prompt`, `model`, `claude_config_dir` (for multi-account), `additional_directories`.
- Claude profiles: claude_list_profiles() reads ~/.config/switcher/profiles/ with profile.toml metadata. Profile set per-project in Settings (project.profile field), passed through AgentSession -> AgentPane `profile` prop -> resolved to config_dir for SDK. Profile name shown as info-only in ProjectHeader.
- ProjectBox has project-level tab bar: Model | Docs | Context | Files | SSH | Memory. Three mount strategies: PERSISTED-EAGER (Model, Docs, Context — always mounted, display:flex/none), PERSISTED-LAZY (Files, SSH, Memory — mount on first activation via {#if everActivated} + display:flex/none). Tab type: `'model' | 'docs' | 'context' | 'files' | 'ssh' | 'memories'`. Model tab = AgentSession+TeamAgentsPanel. Docs tab = ProjectFiles (markdown viewer). Context tab = ContextTab.svelte (LLM context window visualization: stats bar, segmented token meter, file references, turn breakdown; reads from agent store via sessionId prop; replaced old ContextPane ctx database viewer). Files tab = FilesTab.svelte (VSCode-style directory tree + CodeMirror 6 editor with 15 language modes, dirty tracking, Ctrl+S save, save-on-blur setting, image display via convertFileSrc, 10MB gate; CodeEditor.svelte wrapper; PdfViewer.svelte for PDF files via pdfjs-dist with canvas multi-page rendering + zoom 0.5x3x; CsvTable.svelte for CSV with RFC 4180 parser, delimiter auto-detect, sortable columns). SSH tab = SshTab.svelte (CRUD for SSH connections, launch spawns terminal tab in Model tab). Memory tab = MemoriesTab.svelte (pluggable via MemoryAdapter interface in memory-adapter.ts). Rust backend: list_directory_children + read_file_content + write_file_content (FileContent tagged union: Text/Binary/TooLarge). Frontend bridge: files-bridge.ts.
- ProjectHeader shows CWD (ellipsized from START via `direction: rtl`) + profile name as info-only text on right side. AgentPane no longer has DIR/ACC toolbar — CWD and profile are props from parent.
- Skill discovery: claude_list_skills() reads ~/.claude/skills/ (dirs with SKILL.md or .md files). claude_read_skill() reads content. AgentPane `/` prefix triggers autocomplete menu. Skill content injected as prompt via expandSkillPrompt().
- claude-bridge.ts adapter wraps profile/skill Tauri commands (ClaudeProfile, ClaudeSkill interfaces). provider-bridge.ts wraps claude-bridge as generic provider bridge (delegates by ProviderId).
- Provider adapter pattern: ProviderId = 'claude' | 'codex' | 'ollama'. ProviderCapabilities flags gate UI (hasProfiles, hasSkills, hasModelSelection, hasSandbox, supportsSubagents, supportsCost, supportsResume). ProviderMeta registered via registerProvider() in App.svelte onMount. AgentPane receives provider + capabilities props. SettingsTab has Providers section with collapsible per-provider config panels. ProjectConfig.provider field for per-project selection. Settings persisted as `provider_settings` JSON blob.
- Sidecar build: `npm run build:sidecar` bundles SDK into claude-runner.mjs via esbuild (no --external, SDK included in bundle). Future providers will have separate runner bundles.
- Agent preview terminal: `AgentPreviewPane.svelte` is a read-only xterm.js terminal (disableStdin:true) that subscribes to an agent session's messages via `$derived(getAgentSession(sessionId))` and renders tool calls/results in real-time. Bash commands shown as cyan ` cmd`, file ops as yellow `[Read] path`, results as plain text (80-line truncation), errors in red. Spawned via 👁 button in TerminalTabs (appears when agentSessionId prop is set). TerminalTab type: `'agent-preview'` with `agentSessionId` field. Deduplicates — won't create two previews for the same session. ProjectBox passes mainSessionId to TerminalTabs.
- Maximum 4 active xterm.js instances to avoid WebKit2GTK memory issues. Agent preview uses disableStdin and no PTY so is lighter, but still counts.
- Store files using Svelte 5 runes (`$state`, `$derived`) MUST have `.svelte.ts` extension (not `.ts`). Import with `.svelte` suffix. Plain `.ts` compiles but fails at runtime with "rune_outside_svelte".
- Session persistence uses rusqlite (bundled) with WAL mode. Data dir: `dirs::data_dir()/bterminal/sessions.db`.
- Layout store persists to SQLite on every addPane/removePane/setPreset/setPaneGroup change (fire-and-forget). Restores on app startup via `restoreFromDb()`.
- Session groups: Pane.group? field in layout store, group_name column in sessions table, collapsible group headers in sidebar. Right-click pane to set group.
- File watcher uses notify crate v6, watches parent directory (NonRecursive), emits `file-changed` Tauri events.
- Settings use key-value `settings` table in SQLite (session.rs). Frontend: `settings-bridge.ts` adapter. v3 uses SettingsTab.svelte rendered in sidebar drawer panel (v2 SettingsDialog.svelte deleted in P10). SettingsTab has two sections: Global (single-column layout, split into Appearance [theme dropdown, UI font dropdown with sans-serif options + size stepper, Terminal font dropdown with monospace options + size stepper] and Defaults [shell, CWD] — all custom themed dropdowns, no native `<select>`, all persisted via settings-bridge with keys: theme, ui_font_family, ui_font_size, term_font_family, term_font_size, default_shell, default_cwd) and Group/Project CRUD.
- Notifications use ephemeral toast system: `notifications.svelte.ts` store (max 5, 4s auto-dismiss), `ToastContainer.svelte` display. Agent dispatcher emits toasts on agent complete/error/crash.
- StatusBar → Mission Control bar: running/idle/stalled agent counts (color-coded), total $/hr burn rate, "needs attention" dropdown priority queue (up to 5 cards sorted by urgency score, click-to-focus), total tokens + cost. Uses health.svelte.ts store (not workspace store for health signals).
- health.svelte.ts store: per-project health tracking via ProjectTracker map. ActivityState = inactive|running|idle|stalled (15-min stall threshold). Burn rate from 5-min EMA costSnapshots. Context pressure = tokens/model limit. File conflict count from conflicts.svelte.ts. Attention scoring: stalled=100, error=90, ctx>90%=80, file_conflict=70, ctx>75%=40. 5-second tick timer (auto-stop/start). API: trackProject(), recordActivity(), recordToolDone(), recordTokenSnapshot(), getProjectHealth(), getAttentionQueue(), getHealthAggregates().
- conflicts.svelte.ts store: per-project file overlap + external write detection. Records Write/Edit/Bash-write tool_call file paths per session. Detects when 2+ sessions in same worktree write same file. S-1 Phase 2: inotify-based external write detection via fs_watcher.rs — uses 2s timing heuristic (AGENT_WRITE_GRACE_MS) to distinguish agent writes from external. EXTERNAL_SESSION_ID='__external__' sentinel. Worktree-aware. Dismissible. recordExternalWrite() for inotify events. FileConflict.isExternal flag, ProjectConflicts.externalConflictCount. Session-scoped, no persistence.
- tool-files.ts utility: shared extractFilePaths(tc) → ToolFileRef[], extractWritePaths(tc) → string[], extractWorktreePath(tc) → string|null. Bash write detection via regex (>, >>, sed -i, tee, cp, mv). Used by ContextTab (all ops) and agent-dispatcher (writes + worktree tracking for conflict detection).
- ProjectHeader shows status dot (green pulse=running, gray=idle, orange pulse=stalled, dim=inactive) + external write badge (orange ⚡ clickable, shown when externalConflictCount > 0) + agent conflict badge (red ⚠ clickable with ✕) + context pressure badge (>90% red, >75% orange, >50% yellow) + burn rate badge ($/hr). Health prop from ProjectBox via getProjectHealth(). ProjectBox starts/stops fs watcher per project CWD via $effect.
- session_metrics SQLite table: per-project historical session data (project_id, session_id, timestamps, peak_tokens, turn_count, tool_call_count, cost_usd, model, status, error_message). 100-row retention per project. Tauri commands: session_metric_save, session_metrics_load. Persisted on agent completion via agent-dispatcher.
- Session anchors (S-2): Preserves important turns through compaction chains. Types: auto (on first compaction, 3 turns, observation-masked — reasoning preserved in full, only tool outputs compacted), pinned (user-created via pin button in AgentPane), promoted (user-promoted from pinned, re-injectable). Configurable budget via AnchorBudgetScale ('small'=2K|'medium'=6K|'large'=12K|'full'=20K) — per-project slider in SettingsTab, stored as ProjectConfig.anchorBudgetScale in groups.json. Re-injection: anchors.svelte.ts → AgentPane.startQuery() → system_prompt field → sidecar → SDK. ContextTab shows anchor section with budget meter (derived from scale) + promote/demote. SQLite: session_anchors table. Files: types/anchors.ts, adapters/anchors-bridge.ts, stores/anchors.svelte.ts, utils/anchor-serializer.ts.
- Agent tree (AgentTree.svelte) uses SVG with recursive layout. Tree data built by `agent-tree.ts` utility from agent messages.
- ctx integration opens `~/.claude-context/context.db` as SQLITE_OPEN_READ_ONLY — never writes. CtxDb uses Option<Connection> for graceful absence if DB doesn't exist.
- SSH sessions spawn TerminalPane with shell=/usr/bin/ssh and args array. No SSH library needed — PTY handles it natively.
- Theme system: 17 themes in 3 groups — 4 Catppuccin + 7 Editor (VSCode Dark+, Atom One Dark, Monokai, Dracula, Nord, Solarized Dark, GitHub Dark) + 6 Deep Dark (Tokyo Night, Gruvbox Dark, Ayu Dark, Poimandres, Vesper, Midnight). All map to same 26 --ctp-* CSS custom properties — zero component changes needed. ThemeId replaces CatppuccinFlavor. getCurrentTheme()/setTheme() are primary API (deprecated wrappers exist). THEME_LIST has ThemeMeta with group metadata for custom dropdown UI. Open terminals hot-swap via onThemeChange() callback registry in theme.svelte.ts. Typography uses --ui-font-family/--ui-font-size (UI elements, sans-serif fallback) and --term-font-family/--term-font-size (terminal, monospace fallback) CSS custom properties (defined in catppuccin.css). initTheme() restores all 4 font settings (ui_font_family, ui_font_size, term_font_family, term_font_size) from SQLite on startup.
- Detached pane mode: App.svelte checks URL param `?detached=1` and renders a single pane without sidebar/grid chrome. Used for pop-out windows.
- Shiki syntax highlighting uses lazy singleton pattern (avoid repeated WASM init). 13 languages preloaded. Used in MarkdownPane and AgentPane text messages.
- Cargo workspace at v2/ level: members = [src-tauri, bterminal-core, bterminal-relay]. Cargo.lock is at workspace root (v2/), not in src-tauri/.
- EventSink trait (bterminal-core/src/event.rs) abstracts event emission. PtyManager and SidecarManager are in bterminal-core, not src-tauri. src-tauri has thin re-exports.
- RemoteManager (src-tauri/src/remote.rs) manages WebSocket client connections to bterminal-relay instances. 12 Tauri commands prefixed with `remote_`.
- remote-bridge.ts adapter wraps remote machine management IPC. machines.svelte.ts store tracks remote machine state.
- Pane.remoteMachineId?: string routes operations through RemoteManager instead of local managers. Bridge adapters (pty-bridge, agent-bridge) check this field.
- bterminal-relay binary (v2/bterminal-relay/) is a standalone WebSocket server with token auth, rate limiting, and per-connection isolated managers. Commands return structured responses (pty_created, pong, error) with commandId for correlation via send_error() helper.
- RemoteManager reconnection: exponential backoff (1s-30s cap) on disconnect, attempt_tcp_probe() (TCP-only, no WS upgrade), emits remote-machine-reconnecting and remote-machine-reconnect-ready events. Frontend listeners in remote-bridge.ts; machines store auto-reconnects on ready.
- v3 workspace store (`workspace.svelte.ts`) replaces layout store for v3. Groups loaded from `~/.config/bterminal/groups.json` via `groups-bridge.ts`. State: groups, activeGroupId, activeTab, focusedProjectId. Derived: activeGroup, activeProjects.
- v3 groups backend (`groups.rs`): load_groups(), save_groups(), default_groups(). Tauri commands: groups_load, groups_save.
- Telemetry (`telemetry.rs`): tracing + optional OTLP export to Tempo. `BTERMINAL_OTLP_ENDPOINT` env var controls (absent = console-only). TelemetryGuard in AppState with Drop-based shutdown. Frontend events route through `frontend_log` Tauri command → Rust tracing (no browser OTEL SDK — WebKit2GTK incompatible). `telemetry-bridge.ts` provides `tel.info/warn/error()` convenience API. Docker stack at `docker/tempo/` (Grafana port 9715).
- v3 SQLite additions: agent_messages table (per-project message persistence), project_agent_state table (sdkSessionId, cost, status per project), sessions.project_id column.
- v3 App.svelte: VSCode-style sidebar layout. Horizontal: left icon rail (GlobalTabBar, 2.75rem, single Settings gear icon) + expandable drawer panel (Settings only, content-driven width, max 50%) + main workspace (ProjectGrid always visible) + StatusBar. Sidebar has Settings only — Sessions/Docs/Context are project-specific (in ProjectBox tabs). Keyboard: Ctrl+B (toggle sidebar), Ctrl+, (settings), Escape (close).
- v3 component tree: App -> GlobalTabBar (settings icon) + sidebar-panel? (SettingsTab) + workspace (ProjectGrid) + StatusBar. See `docs/v3-task_plan.md` for full tree.
- MarkdownPane reactively watches filePath changes via $effect (not onMount-only). Uses sans-serif font (Inter, system-ui), all --ctp-* theme vars. Styled blockquotes with translucent backgrounds, table row hover, link hover underlines. Inner `.markdown-pane-scroll` wrapper with `container-type: inline-size` for responsive padding via `--bterminal-pane-padding-inline`.
- AgentPane UI (redesigned 2026-03-09): sans-serif root font (`system-ui, -apple-system, sans-serif`), monospace only on code/tool names. Tool calls paired with results in collapsible `<details>` groups via `$derived.by` toolResultMap (cache-guarded by tool_result count). Hook messages collapsed into compact `<details>` with gear icon. Context window meter inline in status strip. Cost bar minimal (no background, subtle border-top). Session summary with translucent surface background. Two-phase scroll anchoring (`$effect.pre` + `$effect`). Tool-aware output truncation (Bash 500 lines, Read/Write 50, Glob/Grep 20, default 30). Colors softened via `color-mix()`. Inner `.agent-pane-scroll` wrapper with `container-type: inline-size` for responsive padding via shared `--bterminal-pane-padding-inline` variable.
- ProjectBox uses CSS `style:display` (flex/none) instead of `{#if}` for tab content panes — keeps AgentSession mounted across tab switches (prevents session ID reset and message loss). Terminal section also uses `style:display`. Grid rows: auto auto 1fr auto.
- Svelte 5 event syntax: use `onclick` not `on:click`. Svelte 5 requires lowercase event handler attributes (no colon syntax).
## Memora Tags
Project tag: `bterminal`
Common tag combinations: `bterminal,architecture`, `bterminal,research`, `bterminal,tech-stack`
## Operational Rules
All operational rules live in `.claude/rules/`. Every `.md` file in that directory is automatically loaded at session start by Claude Code with the same priority as this file.
### Rule Index
| # | File | Scope |
|---|------|-------|
| 01 | `security.md` | **PARAMOUNT** — secrets, input validation, least privilege |
| 02 | `error-handling.md` | **PARAMOUNT** — handle every error visibly |
| 03 | `environment-safety.md` | **PARAMOUNT** — verify target, data safety, K8s isolation, cleanup |
| 04 | `communication.md` | Stop on ambiguity, scope discipline |
| 05 | `git-practices.md` | Conventional commits, authorship |
| 06 | `testing.md` | TDD, unit tests, E2E tests |
| 07 | `documentation.md` | README, CLAUDE.md sync, docs/ |
| 08 | `branch-hygiene.md` | Branches, naming, clean state before refactors |
| 09 | `dependency-discipline.md` | No deps without consent |
| 10 | `code-consistency.md` | Match existing patterns |
| 11 | `api-contracts.md` | Contract-first, flag breaking changes (path-conditional) |
| 12 | `performance-awareness.md` | No N+1, no unbounded fetches (path-conditional) |
| 13 | `logging-observability.md` | Structured logging, OTEL (path-conditional) |
| 14 | `resilience-and-config.md` | Timeouts, circuit breakers, externalized config (path-conditional) |
| 15 | `memora.md` | Persistent memory across sessions |
| 16 | `sub-agents.md` | When to use sub-agents and team agents |
| 17 | `document-imports.md` | Resolve @ imports in CLAUDE.md before acting |
| 18 | `relative-units.md` | Use rem/em for layout, px only for icons/borders |
| 51 | `theme-integration.md` | All colors via --ctp-* CSS vars, never hardcode |
| 52 | `no-implicit-push.md` | Never push unless explicitly asked |