Session Dispatch & Context Restore

Session continuity and agent dispatch. Agents can restore context from a past session for injection into a new conversation, or run a task on a registered agent (HTTP or local) through run_agent / dispatch_to_agent.

MCP Tools

restore_session

Load context from a past session for injection into a new conversation.

Args: session_id (str), last_n (int, default 20)
Returns: JSON with formatted context string + char_count

run_agent / dispatch_to_agent

Route a task to a specific agent, or to the best match by capability (see Agent Registry). An HTTP agent’s job is submitted to its own REST API and polled through that API. A local agent always runs to completion via run_claude — there is no background job id to poll.

Claude CLI Runner

agentibridge/claude_runner.py runs the claude CLI directly via subprocess for every caller that needs a one-shot Claude run: agent_search, run_agent / dispatch_to_agent against a local agent, and init-interactive-agent.

Permissions

Claude CLI invocations spawned this way run with --dangerously-skip-permissions. No approval prompts will block them.

Command Presets

run_claude takes a model directly (default from CLAUDE_DISPATCH_MODEL, sonnet); callers pick the model themselves rather than through a preset name.

Dispatch Bridge (Docker Mode)

When AgentiBridge runs inside Docker, the claude CLI binary isn’t available in the container. A lightweight HTTP bridge runs on the host and proxies every run_claude() call to the local CLI, whichever MCP tool triggered it.

┌─────────────────────┐        HTTP         ┌──────────────────────┐
│ Docker container     │◄──────────────────►│  Host bridge          │
│ AgentiBridge MCP     │   POST /dispatch    │  dispatch_bridge.py   │
│                      │   → 202 + job_id    │                      │
│ claude_runner.py     │                     │  Spawns claude CLI    │
│  _run_claude_http()  │   GET /job/{id}     │  Returns result       │
│  submit + poll       │   → job state       │                      │
└─────────────────────┘                     └──────────────────────┘

Submit + Poll Pattern

The bridge is fire-and-forget: POST /dispatch validates the request, spawns a background task, and returns HTTP 202 with a job_id immediately. The client polls GET /job/{id} with exponential backoff (2s → 10s cap) until the job finishes or the deadline is exceeded. This job id is internal to the bridge — it never reaches an MCP tool caller, who only sees the final result once run_claude() returns.

Bridge Endpoints

Method Path Auth Description
GET /health No Health check
POST /dispatch X-Dispatch-Secret Submit a dispatch request, returns 202
GET /job/{id} No Get job status and result
GET /jobs No List all bridge-level job summaries

Recursion Prevention

The bridge clears CLAUDE_DISPATCH_URL from its environment before calling run_claude(), preventing the CLI runner from routing back to the bridge (which would cause infinite recursion).

Components

File Role
agentibridge/dispatch.py Session context restore
agentibridge/claude_runner.py Claude CLI subprocess / HTTP bridge proxy (submit + poll)
agentibridge/dispatch_bridge.py Host-side HTTP bridge (Docker mode), fire-and-forget

Configuration

CLAUDE_BINARY=claude
CLAUDE_DISPATCH_MODEL=sonnet
CLAUDE_DISPATCH_TIMEOUT=300

# Docker mode — bridge on host proxies CLI calls
# CLAUDE_DISPATCH_URL removed (native dispatch)
DISPATCH_SECRET=<shared-secret>
DISPATCH_BRIDGE_HOST=0.0.0.0
DISPATCH_BRIDGE_PORT=8101