FN-7023: forward MCP runtime config to AI lanes
Forward enabled MCP server configuration through AI session creation and validation paths. - Resolve and materialize effective MCP servers with secret values at runtime for triage, execution, review, merge, evaluation, reflection, mission validation, and dashboard chat/planning lanes. - Add runtime support guards, Claude CLI MCP config generation, and a dashboard validation route for MCP server reachability probes. - Cover MCP forwarding, provider support behavior, settings resolution, validation, and pi extension config generation with targeted tests and docs. Files changed: .changeset/fn-7023-mcp-runtime-forwarding.md | 7 + docs/agents.md | 2 + docs/secrets.md | 1 + docs/settings-reference.md | 6 +- .../src/__tests__/mcp-lane-forwarding.test.ts | 79 +++++++++++ .../src/__tests__/mcp-validate-route.test.ts | 111 ++++++++++++++++ packages/dashboard/src/chat.ts | 3 + packages/dashboard/src/planning.ts | 5 + packages/dashboard/src/routes.ts | 90 ++++++++++++- .../src/__tests__/mcp-lane-forwarding.test.ts | 95 ++++++++++++++ .../engine/src/__tests__/mcp-resolution.test.ts | 104 +++++++++++++++ .../src/__tests__/mcp-runtime-support.test.ts | 44 +++++++ .../src/__tests__/mcp-validation-service.test.ts | 69 ++++++++++ packages/engine/src/__tests__/pi.test.ts | 65 +++++++++ packages/engine/src/agent-reflection.ts | 3 + packages/engine/src/agent-runtime.ts | 36 ++++- packages/engine/src/agent-session-helpers.ts | 2 + packages/engine/src/evaluator.ts | 3 + packages/engine/src/executor.ts | 14 ++ packages/engine/src/index.ts | 3 + packages/engine/src/mcp-resolution.ts | 81 ++++++++++++ packages/engine/src/mcp-runtime-support.ts | 41 ++++++ packages/engine/src/mcp-validation-service.ts | 146 +++++++++++++++++++++ packages/engine/src/merger-ai.ts | 5 + packages/engine/src/merger.ts | 14 ++ packages/engine/src/mission-execution-loop.ts | 3 + packages/engine/src/pi.ts | 44 +++++-- packages/engine/src/reviewer.ts | 3 + packages/engine/src/runtime-resolution.ts | 8 +- packages/engine/src/step-session-executor.ts | 6 +- packages/engine/src/triage.ts | 5 + packages/pi-claude-cli/index.ts | 20 ++- packages/pi-claude-cli/src/mcp-config.ts | 65 ++++++++- 33 files changed, 1150 insertions(+), 33 deletions(-) Fusion-Task-Id: FN-7023 Fusion-Task-Lineage: ea09c6c7-b3c3-4af6-9ca6-0ebbfec17139
This commit is contained in:
gsxdsm
@@ -570,6 +570,8 @@ Assignment-triggered heartbeats are completion-resilient: if an `agent:assigned`
|
||||
|
||||
Self-healing also covers abnormal run/session loss for assigned `in-progress` work. If the task remains assigned but the durable agent has no active heartbeat run and no active executor session after the orphan grace window, `reattach-orphaned-assigned-executions` re-dispatches the task forward via `executor.resumeTaskForAgent(agentId)` without pausing, failing, or moving the task backward.
|
||||
|
||||
Agent runtime sessions receive the project/global effective MCP server set when the selected runtime supports MCP (pi/Claude/ACP-compatible runtimes). Mock and unsupported runtimes skip MCP forwarding with content-free structured logs. MCP secret references are materialized only at session creation and are never stored on the agent or written to logs.
|
||||
|
||||
Heartbeat values are validated and minimum-clamped to 5 minutes (300,000 ms).
|
||||
Project setting `heartbeatMultiplier` (default `1`) scales resolved heartbeat timing globally: both the heartbeat interval (`pollIntervalMs`) and unresponsive timeout base (`heartbeatTimeoutMs`) are multiplied. Per-agent `heartbeatIntervalMs`/`heartbeatTimeoutMs` remain base values before multiplier scaling. This setting is configured from the **Agents** screen's **Controls** popup under "Heartbeat Speed".
|
||||
|
||||
|
||||
@@ -39,6 +39,7 @@ Threat-model baseline:
|
||||
- Ciphertext + nonce are persisted; plaintext exists only in process memory during create/reveal.
|
||||
- Secret values must never be logged.
|
||||
- MCP server settings store only secret references for sensitive env/header/token fields; imports surface plaintext as secret-creation descriptors instead of persisting it in settings.
|
||||
- MCP server secret references are materialized only at session/probe creation time for MCP-capable AI lanes and `POST /api/mcp/validate`; responses and structured logs include status/count metadata only, never resolved env/header values.
|
||||
|
||||
See also: [Storage](./storage.md), [Multi-project](./multi-project.md), [Architecture](./architecture.md), [Settings reference](./settings-reference.md).
|
||||
|
||||
|
||||
@@ -160,7 +160,11 @@ Each server is named and uses one transport:
|
||||
|
||||
Resolution uses project-over-global precedence by server name. The project-level `enabled` flag overrides the global flag when set; if the effective flag is false, no MCP servers are active. When enabled, global servers are loaded first, project servers with the same `name` replace them, and a project server with `enabled:false` removes the inherited server.
|
||||
|
||||
Secret rule: `env` and `headers` maps are sensitive. Values must be Fusion secret references such as `{ "secretRef": "sec_...", "scope": "project" }` or `{ "secretRef": "sec_...", "scope": "global" }`. Write-boundary sanitizers and validators reject plaintext strings in these fields. Claude Desktop-style imports return `secretsToCreate` descriptors for plaintext env/header values and replace those values with secret refs in the imported definitions.
|
||||
Enabled MCP servers are trusted once configured. Fusion materializes the effective server set at AI session creation and forwards it to every MCP-capable AI lane, including chat/planning, executor, reviewer, validator, merger, workflow model nodes, summarization, evaluator, research, cron/automation, mission, and reflection paths. Runtime support is guarded: Claude/pi/ACP-compatible runtimes receive MCP servers, while mock or unsupported runtimes skip forwarding and emit only a structured count/provider/runtime log entry, never server definitions or secret values.
|
||||
|
||||
Secret rule: `env` and `headers` maps are sensitive. Values must be Fusion secret references such as `{ "secretRef": "sec_...", "scope": "project" }` or `{ "secretRef": "sec_...", "scope": "global" }`. Write-boundary sanitizers and validators reject plaintext strings in these fields. Claude Desktop-style imports return `secretsToCreate` descriptors for plaintext env/header values and replace those values with secret refs in the imported definitions. At runtime, secret references are revealed through the scoped secrets store immediately before forwarding or validation, kept only in memory, and never echoed in API responses.
|
||||
|
||||
`POST /api/mcp/validate` validates an MCP server definition or configured server name against the current project context. The route resolves and materializes the target server with the same secret rules, then performs a bounded reachability probe (`stdio` supervised spawn, `sse`/`streamable-http` bounded fetch) and returns only `{ status, message? }` without resolved env/header contents.
|
||||
|
||||
### Notification providers (pluggable)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user