spacebot-dev

name

spacebot-dev

description

Specialized Spacebot development skill with deep architectural knowledge. Use when working on any Spacebot feature, bug fix, refactor, or investigation. Covers the five-process agent model (Channel, Branch, Worker, Compactor, Cortex), memory system (SQLite + LanceDB + redb), task lifecycle, 60+ tool definitions, Rig framework integration, prompt templates, identity system, messaging adapters, config hot-reload, security/sandbox, and API server. Activate for any Rust code changes in src/, prompt template edits in prompts/, migration work, tool additions, or architectural questions.

Spacebot Development Guide

You are working on Spacebot, a multi-process agentic AI system built in Rust. This skill gives you the architectural knowledge to make correct decisions without reading every file.

Architecture Overview

Single binary crate (src/lib.rs + src/main.rs). No workspace. 40 public modules. Three database backends (relational SQL via enum DbPool { Sqlite, Postgres }, LanceDB for vectors, redb for key-value). Five process types. 60+ LLM tools. 9 messaging platform adapters (including Portal).

The core abstraction: every LLM process is a Rig Agent<SpacebotModel, SpacebotHook> with an isolated ToolServer, independent history, and typed ProcessEvent communication.

The Five Process Types

Each process type has a distinct role, tool set, and isolation boundary. Never mix their responsibilities.

Channel (src/agent/channel.rs)

• User-facing conversation. Delegates everything. Never blocks.
• Tools: reply, branch, spawn_worker, route, cancel, skip, react, cron, send_file, send_message, project_manage
• Conditional tools: attachment_recall (when save_attachments), send_agent_message (when links configured), set_outcome (cron executions)
• NO memory tools (delegated to branches)
• Max turns: 5. Model default: claude-sonnet
• Sub-modules: channel_attachments.rs, channel_dispatch.rs, channel_history.rs, channel_prompt.rs
• Key struct: Channel { id, agent_id, history, compactor, control_handle, ... }

Branch (src/agent/branch.rs)

• Fork of channel context for independent thinking. Short-lived (seconds).
• Tools: memory_save, memory_recall, memory_delete, channel_recall, spacebot_docs, email_search, worker_inspect, task_create, task_list, task_update, file_read
• Conditional tools: wiki tools (6, when wiki enabled), memory_persistence_complete (for persistence contract), spawn_worker (when channel state available)
• Max turns: 10. Model default: claude-sonnet
• Memory persistence contract: silent branches must save memories before completing
• Constants: MAX_OVERFLOW_RETRIES: 2, MAX_MEMORY_CONTRACT_RETRIES: 2

Worker (src/agent/worker.rs)

• Independent task executor. No channel context. Has execution tools.
• Tools: shell, file_read, file_write, file_edit, file_list, task_update, set_status, worker_inspect, optionally browser, web_search, mcp_*
• Max turns: 50 (segmented into 15-turn blocks via TURNS_PER_SEGMENT)
• States: Running, WaitingForInput, Done, Failed
• Two kinds: fire-and-forget and interactive (accepts follow-ups via route)
• Constants: MAX_OVERFLOW_RETRIES: 2, MAX_TRANSIENT_RETRIES: 5, MAX_SEGMENTS: 10

Compactor (src/agent/compactor.rs)

• Programmatic monitor, NOT an LLM process. Watches context size, triggers compaction workers.
• Thresholds: 80% background, 85% aggressive, 95% emergency truncation
• Called via check_and_compact() after each channel turn
• Actions: Background, Aggressive, EmergencyTruncate

Cortex (src/agent/cortex.rs)

• System-level observer. Singleton per agent. Sees across all channels.
• Tools (bulletin): memory_save only
• Tools (cortex chat): full superset — memory + task + shell + file + browser + factory + config tools
• Responsibilities: memory bulletin, worker/branch supervision, memory maintenance, ready-task loop
• Constants: BULLETIN_REFRESH_FAILURE_BACKOFF_BASE_SECS: 30, BULLETIN_REFRESH_CIRCUIT_OPEN_THRESHOLD: 3, MAINTENANCE_TASK_TIMEOUT_MIN_SECS: 300, MAINTENANCE_TASK_TIMEOUT_MAX_SECS: 3600

Process Interaction Flow

User message -> Channel LLM turn
  -> needs thinking -> spawns Branch (fork of history)
  -> needs execution -> spawns Worker (independent task)
  -> replies immediately via reply tool
Branch finishes -> result injected into channel history -> channel retriggered
Worker finishes -> status update injected -> channel retriggered

Retrigger events are debounced. Multiple branches completing within a short window batch into one turn. Limit: MAX_RETRIGGERS_PER_TURN: 3.

Event System

Two broadcast::channel<ProcessEvent> buses per agent:

• event_tx — control/lifecycle (capacity: 256)
• memory_event_tx — memory-save telemetry for cortex (capacity: 1024)

23 event variants in ProcessEvent enum (defined in src/lib.rs):

• Channel: BranchStarted, BranchResult, WorkerStarted, WorkerStatus, WorkerIdle, WorkerComplete, WorkerInitialResult, TextDelta
• Process: ToolStarted, ToolCompleted, StatusUpdate
• Memory: MemorySaved, CompactionTriggered
• Communication: AgentMessageSent, AgentMessageReceived
• Tasks: TaskUpdated
• Workers: WorkerPermission, WorkerQuestion, WorkerText
• OpenCode: OpenCodeSessionCreated, OpenCodePartUpdated
• System: CortexChatUpdate, SettingsUpdated

Type Aliases

pub type AgentId = Arc<str>;
pub type ChannelId = Arc<str>;
pub type WorkerId = uuid::Uuid;
pub type BranchId = uuid::Uuid;
pub enum ProcessType { Channel, Branch, Worker, Compactor, Cortex }
pub enum ProcessId { Channel(ChannelId), Worker(WorkerId), Branch(BranchId) }

Memory System (src/memory/)

Three Storage Backends

• SQLite (store.rs) — memories table (content, type, importance, timestamps) + associations table (graph edges with weights and relation types)
• LanceDB (lance.rs) — Vector embeddings (HNSW indexing) + full-text search (Tantivy). Joined on memory ID.
• Embedding (embedding.rs) — FastEmbed (local, no external API). Wrapped in Arc<Mutex<EmbeddingModel>> for thread safety.

Memory Types (src/memory/types.rs)

pub enum MemoryType {
    Fact,        // Something true (can be updated/contradicted)
    Preference,  // User likes/dislikes
    Decision,    // A choice that was made (carries why-context)
    Identity,    // Core user/agent info (never decayed, always surfaced)
    Event,       // Something that happened (temporal, naturally decays)
    Observation, // Inferred patterns
    Goal,        // Active goals (used in bulletin)
    Todo,        // Actionable task/reminder (cortex can promote to tasks)
}

Associations (src/memory/types.rs)

pub enum RelationType {
    RelatedTo,   // General semantic connection
    Updates,     // Newer version of same info
    Contradicts, // Conflicting information
    CausedBy,    // Causal relationship
    ResultOf,    // Result relationship
    PartOf,      // Hierarchical relationship
}

Auto-association: new memories search for similar ones. >0.9 similarity = Updates edge.

Search Modes (src/memory/search.rs)

pub enum SearchMode {
    Hybrid,    // Vector (HNSW) + FTS (Tantivy) + Graph traversal, merged via RRF (default)
    Recent,    // Most recent by created_at DESC, no query needed
    Important, // Highest importance DESC, no query needed
    Typed,     // Filter by MemoryType, requires memory_type param
}

pub enum SearchSort {
    Recent,       // Most recent first (created_at DESC, default)
    Importance,   // Highest importance first
    MostAccessed, // Most accessed first (access_count DESC)
}

RRF (Reciprocal Rank Fusion) works on ranks not raw scores. This handles different scales of vector and keyword results without normalization.

Memory Maintenance (cortex-managed)

• Decay: reduce importance of old, unaccessed memories (rate: 0.05/day)
• Prune: delete below importance floor (threshold: 0.1, min age: 30 days)
• Merge: combine near-duplicates (similarity threshold: 0.95)
• Identity memories exempt from decay/pruning

Memory Bulletin

Cortex periodically generates an LLM-curated summary injected into every channel system prompt. Eight retrieval sections (programmatic), then single LLM synthesis pass (~500 words). Cached in RuntimeConfig via ArcSwap. Refresh: warmup loop (900s) + fallback (3600s).

Working Memory (working.rs)

Ephemeral per-channel memory. Not persisted to SQLite/LanceDB.

Task System (src/tasks/)

Task Structure (src/tasks/store.rs)

pub struct Task {
    pub id: String,
    pub task_number: i64,
    pub title: String,
    pub description: Option<String>,
    pub status: TaskStatus,
    pub priority: TaskPriority,
    pub owner_agent_id: String,
    pub assigned_agent_id: String,
    pub subtasks: Vec<TaskSubtask>,
    pub metadata: Value,
    pub source_memory_id: Option<String>,
    pub worker_id: Option<String>,
    pub created_by: String,
    pub approved_at: Option<String>,
    pub approved_by: Option<String>,
    pub created_at: String,
    pub updated_at: String,
    pub completed_at: Option<String>,
}

pub enum TaskStatus { PendingApproval, Backlog, Ready, InProgress, Done }
pub enum TaskPriority { Critical, High, Medium, Low }

Lifecycle (from docs)

pending_approval -> ready (approval) or backlog (shelve)
backlog -> ready (manual promotion)
ready -> in_progress (cortex pickup)
in_progress -> done (success) or ready (failure, re-queued)

Cortex Ready-Task Loop

Every tick_interval_secs (30s): claim oldest ready task -> in_progress -> spawn worker -> bind worker_id -> execute -> done or back to ready.

Workers get restricted task_update (subtasks/metadata only, cannot change status/priority/title/description).

Tool System (src/tools/)

Global Constants

• MAX_TOOL_OUTPUT_BYTES: 50_000
• MAX_DIR_ENTRIES: 500

Five ToolServer Configurations

1. Channel — tools added/removed per turn via add_channel_tools()/remove_channel_tools()
2. Branch — isolated per branch, static tool set
3. Worker — isolated per worker, static + optional tools
4. Cortex — memory_save only
5. Cortex Chat — full superset including factory tools

Tool Implementation Pattern (Rig Framework)

impl Tool for SomeTool {
    const NAME: &'static str = "tool_name";
    type Args = SomeArgs;   // Deserialize + JsonSchema
    type Error = SomeError;  // Display
    type Output = SomeOutput; // Serialize
    async fn call(&self, args: Self::Args) -> Result<Self::Output, Self::Error>;
}

Doc comments on tool input structs serve dual purpose: Rust documentation AND LLM instructions.

Adding a New Tool

1. Create src/tools/<name>.rs with struct implementing rig::tool::Tool
2. Add pub mod <name>; to src/tools.rs
3. Add Jinja2 description template in prompts/en/tools/<name>_description.md.j2
4. Register in appropriate create_*_tool_server() factory function in src/tools.rs
5. If channel-dynamic, add to add_channel_tools()/remove_channel_tools()

Tool Categories

• Delegation: reply, branch, spawn_worker, route, cancel, skip, react
• Memory: memory_save, memory_recall, memory_delete, memory_persistence_complete
• Execution: shell, file_read, file_write, file_edit, file_list, project_manage
• Browser: browser_launch, browser_navigate, browser_snapshot, browser_click, browser_type, browser_press_key, browser_screenshot, browser_evaluate, browser_tab_open, browser_tab_list, browser_tab_close, browser_close
• Knowledge: channel_recall, attachment_recall, wiki_, skills_, spacebot_docs
• Tasks: task_create, task_list, task_update
• Network: web_search, email_search, send_message_to_another_channel, send_agent_message, send_file
• Secrets: secret_set
• Status: set_status, set_outcome, worker_inspect, config_inspect
• Scheduling: cron
• Factory: factory_create_agent, factory_update_identity, factory_update_config, factory_load_preset, factory_list_presets, factory_search_context

LLM Integration (src/llm/)

Key Types

pub struct LlmManager {
    config: ArcSwap<LlmConfig>,  // Hot-reloadable
    http_client: reqwest::Client,
    rate_limited: Arc<RwLock<HashMap<String, Instant>>>,
}

pub struct SpacebotModel {
    pub name: String, pub provider: String,
    pub context_window: usize, pub max_output_tokens: usize,
}

pub struct RoutingConfig {
    channel_model, branch_model, worker_model, cortex_model: String,
}

Model Name Format

<provider>/<model> — e.g., anthropic/claude-sonnet-4-20250514, openai/gpt-4o

Providers (14+)

Anthropic, OpenAI, OpenRouter, Gemini, Groq, DeepSeek, XAI, Fireworks, Together, Mistral, Nvidia, Minimax, Moonshot, Azure, GitHub Copilot, custom providers

Hot Reload

Uses arc_swap::ArcSwap<T> for lock-free reads. File watcher with 2s debounce re-parses config.toml and updates RuntimeConfig fields atomically.

Hot-reloadable: model routing, compaction thresholds, max_turns, context_window, browser config, warmup config, identity files, skills, bindings Needs restart: LLM API keys, messaging tokens, agent topology, database paths

Configuration (src/config/)

Config File

~/.spacebot/config.toml or $SPACEBOT_DIR/config.toml

Value Resolution

• secret:NAME — secret store lookup
• env:VAR_NAME — environment variable
• literal value
• Resolution chain: agent override > [defaults] > env fallback > hardcoded default

Key Config Sections

• [defaults] — max_concurrent_branches (5), max_turns (5), context_window (128000), history_backfill_count (50)
• [defaults.routing] — channel, branch, worker, compactor, cortex model assignments + rate_limit_cooldown_secs (60)
• [defaults.compaction] — background_threshold (0.80), aggressive_threshold (0.85), emergency_threshold (0.95)
• [defaults.cortex] — tick_interval_secs (30), bulletin_interval_secs (3600), bulletin_max_words (500), worker_timeout_secs (600), branch_timeout_secs (60), maintenance_* settings
• [defaults.warmup] — enabled (true), eager_embedding_load (true), refresh_secs (900)
• [defaults.browser] — enabled (true), headless (true), evaluate_enabled (false)
• [defaults.ingestion] — enabled (true), poll_interval_secs (30), chunk_size (4000)
• [defaults.opencode] — enabled (false), max_servers (5)
• [agents.sandbox] — mode ("enabled"), writable_paths, passthrough_env

Sub-modules

• types.rs — Config, RuntimeConfig, AgentConfig, DefaultsConfig structs
• load.rs — Parse config.toml + env vars + secrets
• toml_schema.rs — TOML structure definitions
• runtime.rs — In-memory state (hot-reloadable via ArcSwap)
• permissions.rs — Platform permission specs
• providers.rs — Provider default configs
• onboarding.rs — Interactive setup wizard
• watcher.rs — File watcher for hot reload

Database Layer (src/db.rs)

Triple-Database Bundle (post-PR-11.1, dual SQLite/Postgres backend)

pub struct Db {
    /// Backend-typed SQL pool. SQLite default; Postgres opt-in via
    /// `[database] url = "postgres://..."` in config.toml. Postgres
    /// hard-errors at connect until PR 11.3 lands `migrations/postgres/`.
    pub pool: Arc<DbPool>,
    pub lance: lancedb::Connection,   // Vector embeddings
    pub redb: Arc<redb::Database>,    // Key-value config
}

/// Backend-typed sqlx pool. Each variant holds the native typed pool so
/// chrono types, query_as!, and FromRow derives work per variant.
pub enum DbPool {
    Sqlite(SqlitePool),
    Postgres(PgPool),
}

impl Db {
    /// Transitional accessor returning the underlying SqlitePool for the
    /// SQLite variant; Err for Postgres. PR 11.2 widened the instance-tier
    /// stores to `Arc<DbPool>`; the remaining caller is `AgentDeps.sqlite_pool`
    /// on the per-agent path. PR 11.3 retires this accessor.
    pub fn sqlite_pool(&self) -> Result<&SqlitePool> {
        self.pool.as_sqlite()
    }
}

DialectAdapter (src/dialect.rs) is the companion trait for cross-backend DDL string differences (now_expr(), autoincrement_pk(), json_type()). SqliteDialect and PostgresDialect are zero-state unit structs. Per-query placeholder syntax (? vs $1) and JSON operators live in per-variant store dispatch (instance-tier stores shipped in PR 11.2; per-agent stores in PR 11.3), not in this trait.

SQLite Tables (per-agent spacebot.db)

memories, associations, conversation_messages, channels, cron_jobs, cron_executions, worker_runs, branch_runs, cortex_events, cortex_chat_messages, tasks, ingestion_progress, ingestion_files, agent_profile, portal_conversations, channel_settings, token_usage, working_memory

Global SQLite (global.db)

projects, project_repos, project_worktrees, notifications, wiki_pages, wiki_page_versions, wiki_pages_fts

redb Databases

• config.redb — UI prefs, feature flags
• settings.redb — runtime settings
• secrets.redb — encrypted secret store (AES-256-GCM + Argon2id)

Migration Rules

• NEVER edit existing files in migrations/
• Always create a new timestamped migration
• Migration files are immutable
• Format: YYYYMMDDHHMMSS_description.sql
• Global migrations go in migrations/global/

Prompt System (src/prompts/)

Template Engine

MiniJinja (Rust Jinja2). Templates embedded in binary via include_str!.

Template Files (prompts/en/)

• channel.md.j2, branch.md.j2, worker.md.j2, cortex.md.j2, compactor.md.j2
• memory_persistence.md.j2, ingestion.md.j2, cortex_bulletin.md.j2
• tools/*.md.j2 — 48+ tool description templates
• fragments/ — reusable prompt fragments (worker_capabilities, conversation_context, skills_*, retrigger, truncation, etc.)

Channel Prompt Assembly Order

1. Identity context (SOUL.md + IDENTITY.md + ROLE.md)
2. Memory bulletin (cortex-generated)
3. Base channel instructions
4. Skills prompt (if any active)
5. Worker capabilities
6. Conversation context (platform/channel metadata)
7. Status text

PromptEngine

pub struct PromptEngine { ... }
// Key methods:
engine.render_channel_prompt(identity_context, memory_bulletin, skills_prompt, worker_capabilities, conversation_context, status_text)
engine.render_static("worker")
engine.render_worker_capabilities(browser_enabled, web_search_enabled)

Identity System (src/identity/)

Per-Agent Files (in ~/.spacebot/agents/{id}/)

• SOUL.md — personality, values, communication style, boundaries
• IDENTITY.md — agent name, nature, purpose
• ROLE.md — responsibilities, scope, escalation rules

Live outside workspace so worker file tools cannot access them. Hot-reloaded on change.

Presets (presets/)

11 agent personas: community-manager, content-writer, customer-support, engineering-assistant, executive-assistant, integration-engineer, main-agent, project-manager, research-analyst, sales-bdr, sre. Each has meta.toml, IDENTITY.md, ROLE.md, SOUL.md.

Messaging (src/messaging/)

9 platform adapters: Discord, Slack, Telegram, Twitch, Signal, Email (IMAP/SMTP), Webhook, Mattermost, Portal (web chat).

All implement Messaging trait -> MessagingManager -> unified InboundMessage stream -> Channel.

Named adapter instances support multiple connections per platform (e.g., discord:ops).

Security (src/secrets/, src/sandbox/)

Sandbox

• bubblewrap (Linux) — mount namespaces, PID isolation, die-with-parent
• sandbox-exec (macOS) — SBPL deny-default
• Filesystem: read-only allowlist + writable workspace/tmp
• Agent data directory masked at kernel level

Secret Store

Instance-level redb at secrets.redb. Two categories:

• System — never exposed (LLM keys, messaging tokens)
• Tool — injected as env vars to workers (GH_TOKEN, etc.) Encryption: AES-256-GCM + Argon2id. Master key in OS credential store.

Output Protection

1. Stream scrubbing — exact-match redaction of stored secrets (handles chunk boundaries)
2. Leak detection — regex patterns at channel egress (reply tool)
3. Exec env var blocklist — blocks LD_PRELOAD, DYLD_INSERT_LIBRARIES, NODE_OPTIONS, etc.

Permissions (per-agent)

Five dimensions: file_read/file_write ("deny"|"workspace"|"allow"|globs), shell, exec ("deny"|"allowlist"|"allow"), browser (bool + url_allowlist), network_outbound (bool)

API Server (src/api/)

Axum-based REST API on port 19898. Embedded frontend via rust-embed. OpenAPI docs via utoipa.

39 route modules (excludes server.rs + state.rs scaffolding): activity, admin_access_review, admin_claim, admin_orphans, admin_teams, agents, attachments, audit, auth_config, bindings, channels, config, cortex, cron, desktop, factory, ingest, links, mcp, me, memories, messaging, models, notifications, opencode_proxy, portal, projects, providers, resources, secrets, settings, skills, ssh, system, tasks, tools, usage, wiki, workers

Special endpoints: events_sse() (real-time SSE), health(), backup_export()/backup_restore()

Hooks System (src/hooks/)

• SpacebotHook — emits ProcessEvents for every tool call/completion, tracks memory persistence contracts, enforces cancellation, tool nudging, leak detection
• CortexHook — lighter cortex-specific variant
• LoopGuard — detects and prevents infinite loops

Hook actions: Continue, Terminate, Skip

Key Constants Reference

Constant	Value	Location
EVENT_SUMMARY_MAX_CHARS	160	lib.rs
CONTROL_EVENT_BUS_CAPACITY	256	lib.rs
MEMORY_EVENT_BUS_CAPACITY	1024	lib.rs
MAX_TOOL_OUTPUT_BYTES	50,000	tools.rs
MAX_DIR_ENTRIES	500	tools.rs
MAX_RETRIGGERS_PER_TURN	3	channel_prompt.rs
TURNS_PER_SEGMENT	15	worker.rs
MAX_OVERFLOW_RETRIES	2	worker.rs, branch.rs
MAX_TRANSIENT_RETRIES	5	worker.rs
MAX_SEGMENTS	10	worker.rs
DETACHED_WORKER_LIFECYCLE_TERMINAL	3	process_control.rs
Background compaction	0.80	config defaults
Aggressive compaction	0.85	config defaults
Emergency truncation	0.95	config defaults
Context window default	128,000	config defaults
Max concurrent branches	5	config defaults
Bulletin refresh (warmup)	900s	config defaults
Bulletin refresh (fallback)	3600s	config defaults
Worker timeout	600s	cortex config
Branch timeout	60s	cortex config
Circuit breaker threshold	3	cortex config
Memory decay rate	0.05/day	cortex config
Memory prune threshold	0.1	cortex config
Memory min age for pruning	30 days	cortex config
Memory merge similarity	0.95	cortex config
Link conversation safety cap	20 turns	links

On-Disk Structure

~/.spacebot/
  config.toml
  embedding_cache/
  skills/
  data/
    global.db            # instance-wide SQLite
    secrets.redb         # encrypted secret store
  agents/{id}/
    SOUL.md, IDENTITY.md, ROLE.md
    workspace/
      skills/, ingest/
    data/
      spacebot.db        # per-agent SQLite
      lancedb/           # vector search
      config.redb        # key-value settings
      settings.redb      # runtime settings
      screenshots/, logs/
    archives/            # compaction transcripts

Multi-Agent Communication

Agents communicate via directed links (hierarchical/peer, one_way/two_way).

Link channels format: link:{self}:{peer}, safety cap of 20 turns.

Tools: send_agent_message, conclude_link.

Config: [[links]] with from/to agent IDs, type, direction. [[humans]] with id, display_name, role, bio. [[groups]] for visual containers.

Cron System (src/cron/)

User-defined scheduled tasks. Fresh channel per execution.

Config: id, prompt, cron_expr (wall-clock), interval_secs, delivery_target, active hours, enabled, run_once, timeout_secs.

Circuit breaker: 3 consecutive failures -> auto-disable.

CLI Commands

spacebot start [-f foreground] [-d debug]
spacebot stop
spacebot restart
spacebot status
spacebot secrets set <NAME>
spacebot skill add|install|list|info|remove

Environment: SPACEBOT_DIR, SPACEBOT_DEPLOYMENT (docker/hosted/native), SPACEBOT_MAX_AGENTS, SPACEBOT_CRON_TIMEZONE, SPACEBOT_CHANNEL_MODEL, SPACEBOT_WORKER_MODEL

Metrics (feature-gated: metrics)

Prometheus-compatible at /metrics on port 9090. All prefixed spacebot_.

Key: llm_requests_total, llm_request_duration_seconds, llm_tokens_total, llm_estimated_cost_dollars, tool_calls_total, tool_call_duration_seconds, active_workers, active_branches, memory_entry_count, process_errors_total, dispatch_while_cold_count

Anti-Patterns (from AGENTS.md)

These are validated constraints. Violating them produces architecturally broken code.

• Don't block the channel. The channel never waits on branches, workers, or compaction. If the channel awaits a branch result before responding, the design is wrong.
• Don't dump raw search results into channel context. Memory recall goes through a branch, which curates. The channel gets clean conclusions.
• Don't give workers channel context. Workers get a fresh prompt and a task. If it needs conversation context, that's a branch.
• Don't make the compactor an LLM process. The compactor is programmatic. The LLM work happens in the compaction worker it spawns.
• Don't use #[async_trait]. Use native RPITIT. Only add a Dyn companion trait when you actually need dyn Trait.
• Don't create many small files. Implement in existing files unless it's a new logical component.
• Don't add features without updating existing docs. Update relevant docs in the same commit. Don't create new doc files for this.

Patterns to Implement (from AGENTS.md)

These are validated patterns. Implement them when building the relevant module.

• Tool nudging / outcome gate: Workers cannot exit with text-only response until they signal a terminal outcome via set_status(kind: "outcome"). Hook fires Terminate and retries with nudge prompt (up to 2 retries). See docs/design-docs/tool-nudging.md.
• Fire-and-forget DB writes: tokio::spawn for history saves, memory writes, log persistence. User gets response immediately.
• Hybrid search with RRF: score = sum(1/(60 + rank)). RRF works on ranks, not raw scores.
• Leak detection: Regex patterns for API keys, tokens, PEM keys. Scan in SpacebotHook.on_tool_result() and before outbound HTTP.
• Workspace path guard: File tools reject writes to identity/memory paths with error directing LLM to the correct tool.
• Circuit breaker: Auto-disable recurring tasks after 3 consecutive failures. Apply to cron, maintenance, cortex routines.
• Error-as-result for tools: Tool errors are structured results, not panics. The LLM sees errors and can recover.
• Worker state machine: Validate transitions with can_transition_to() using matches!. Illegal transitions are runtime errors.

Build Order (6 phases from AGENTS.md)

When implementing from scratch, follow this order:

1. Foundation — error.rs, config.rs, db/, llm/, main.rs
2. Memory — memory/types.rs, store.rs, lance.rs, embedding.rs, search.rs, maintenance.rs
3. Agent Core — hooks/spacebot.rs, agent/status.rs, tools/, agent/worker.rs, branch.rs, channel.rs, compactor.rs
4. System — identity/, conversation/, prompts/, agent/cortex.rs, hooks/cortex.rs, cron/
5. Messaging — messaging/traits.rs, manager.rs, webhook.rs, telegram.rs, discord.rs
6. Hardening — secrets/, settings/, leak detection, workspace path guards, circuit breakers

Frontend Architecture (from docs/design-docs/spaceui-migration.md)

Component Library

Local UI primitives replaced by SpaceUI packages:

• @spacedrive/primitives — base components
• @spacedrive/ai — AI-specific components
• @spacedrive/forms — form components
• @spacedrive/explorer — file explorer components

Tailwind v4 via @tailwindcss/vite with SpaceUI design token system.

Layout

Persistent 220px sidebar with accordion agent sub-nav. Global workers popover in footer. Dashboard landing page with notifications, token usage, and activity cards.

Key UI Patterns

• Settings — decomposed from 2900-line monolith into 12 section components
• AgentConfig — decomposed from 1452 lines into ConfigSidebar + section editors
• TopologyGraph — decomposed from 2074 lines into OrgGraph + ProfileNode + GroupNode + edge/config panels
• Tasks — Linear-style task list (replaced kanban board), detail views, SSE-driven updates
• Portal — modular PortalPanel/Timeline/Composer/Header, file attachments, multipart upload

Backend Features Supporting UI

• ConversationSettings struct with memory mode, delegation, worker context, model selection
• ResponseMode enum: Active, Observe, MentionOnly (replaces old listen_only_mode)
• Per-channel persistence via ChannelSettingsStore with resolution chain: per-channel DB > binding defaults > agent defaults
• Settings hot-reload via ProcessEvent::SettingsUpdated
• "Direct mode" gives channels full worker-level tools
• UsageAccumulator for per-process token tracking

Metrics Instrumentation (from METRICS.md)

All telemetry behind #[cfg(feature = "metrics")]. Zero runtime cost without the feature. Docker always includes metrics.

Key Metrics (all prefixed spacebot_)

• LLM: llm_requests_total, llm_request_duration_seconds, llm_tokens_total{direction}, llm_estimated_cost_dollars
• Tools: tool_calls_total{tool_name}, tool_call_duration_seconds
• MCP: mcp_connections, mcp_tools_registered, mcp_tool_calls_total, mcp_tool_call_duration_seconds
• Messaging: messages_received_total, messages_sent_total, message_handling_duration_seconds, channel_errors_total
• Memory: memory_reads_total, memory_writes_total, memory_updates_total{operation}, memory_entry_count, memory_operation_duration_seconds, memory_search_results, memory_embedding_duration_seconds
• Workers: active_workers, active_branches, branches_spawned_total, worker_duration_seconds, worker_cost_dollars, context_overflow_total
• Cost: llm_estimated_cost_dollars
• API: http_requests_total, http_request_duration_seconds
• Cron: cron_executions_total
• Warmup: dispatch_while_cold_count, warmup_recovery_latency_ms

Total cardinality: ~800-9200 series (safe for any Prometheus deployment).

Uses private prometheus::Registry (not global) to avoid conflicts. Metrics server on 0.0.0.0:9090, separate from API server on 127.0.0.1:19898.

Adding New Metrics

1. Add metric definition in src/telemetry/registry.rs
2. Gate instrumentation with #[cfg(feature = "metrics")] at statement level
3. Access registry via crate::telemetry::Metrics::global()
4. Update METRICS.md with new metric documentation

Development Workflow

Before any push

just gate-pr    # Full validation pipeline

Quick checks

cargo fmt --all                    # Format
cargo clippy --all-targets         # Lint
cargo test --lib                   # Unit tests
cargo test --tests --no-run        # Compile integration tests
just check-typegen                 # Verify TypeScript schema sync

Cadence guidance: See .claude/rules/rust-iteration-loop.md for which of these to run per change class. The inner loop should use the narrowest check that catches the bug class. Reserve just gate-pr for pre-push.

Coding discipline: See .claude/rules/coding-discipline.md. Surface assumptions before implementing, keep changes surgical, and default to TDD with the named escape hatches for docs, config, and async-state paths.

Frontend

Always use bun, never npm/pnpm/yarn:

cd interface && bun install && bun run dev    # Dev server at :19840
cd docs && bun install && bun run dev         # Docs site

Common Patterns When Implementing Features

Adding a new process capability:

1. Modify the appropriate src/agent/*.rs process file
2. If adding a tool, follow the tool addition steps above
3. Update the prompt template in prompts/en/ if the LLM needs to know about it
4. Add ProcessEvent variant if new lifecycle events needed

Modifying the memory system:

1. Changes to src/memory/store.rs for SQLite operations
2. Changes to src/memory/lance.rs for vector/embedding operations
3. New migration in migrations/ if schema changes needed
4. Update search in src/memory/search.rs if query patterns change

Adding a messaging adapter:

1. Create src/messaging/<platform>.rs implementing Messaging trait
2. Add to MessagingManager in src/messaging/manager.rs
3. Add config section in src/config/toml_schema.rs

Config changes:

1. Add TOML field in src/config/toml_schema.rs
2. Add to Config struct in src/config/types.rs
3. If hot-reloadable, add ArcSwap field to RuntimeConfig in src/config/runtime.rs
4. Parse in src/config/load.rs

API endpoint:

1. Create route handler in src/api/<resource>.rs
2. Register in router in src/api/server.rs
3. Add OpenAPI annotations via utoipa macros
4. Run just typegen to regenerate TypeScript types

Related Reference Skills

• rust-async-patterns (from the systems-programming plugin). Broader Rust async tutorial covering JoinSet, tokio_util::sync::CancellationToken, Semaphore-based resource pools, and futures::stream patterns Spacebot does not document inline.

  The skill's Pattern 5 demonstrates #[async_trait]. Spacebot uses native RPITIT instead, which avoids the Box<dyn Future> allocation that #[async_trait] introduces. When you reach for Pattern 5, substitute the RPITIT pattern documented in .claude/rules/rust-patterns.md Trait Design and RUST_STYLE_GUIDE.md:459.