werk
Repositories Projects Packages

@loom/mcp (0.1.1)

Published 2026-07-26 20:22:43 +02:00 by amelie

Installation

@loom:registry=https://gitlab.amandy.love/api/packages/werk/npm/
npm install @loom/mcp@0.1.1
"@loom/mcp": "0.1.1"

About this package

@loom/mcp

MCP is the plug. Our policy is the security.

External agents request tools; @loom/mcp enforces policy, path guards, plan gates, session scope, and audit logging before anything runs.

Agent (Cursor / Gemini / Codex)
  → tool request
@loom/mcp (McpSafetyGate)
  → tool policy (read / dry-run-write / write / dangerous)
  → path guard (realpath, symlinks, blocked segments)
  → secret guard (.env, keys, credentials)
  → plan gate (server-registered planId)
  → session gate (createdPlans per connection)
  → audit log
  → execute adapter (workspace / analysis / nerve / deck)

Safety layers

Layer Module Role
Tool policy toolPolicy.ts Risk class per tool; default deny write/dangerous
Permissions permissions.ts Config + session flags, apply: true
Path guard pathGuards.ts allowedRoots, realpath, symlink escape, node_modules/.git/…
Secret guard secretGuards.ts Block .env, keys, credentials (read and write)
Plan gate planGate.ts planEdit registers planId; applyEditPlan only accepts session plans
Session session.ts Per-client createdPlans, scoped roots
Audit auditLog.ts Every call logged (memory + .loom/mcp-audit.jsonl)
Sandbox @loom/sandbox Validation execution uses safe profiles; no host shell execution
Gate McpSafetyGate.ts Orchestrates all checks

Config defaults

{
  allowWrite: false,
  allowDangerous: false,
  allowLiveAi: false,
  requirePlanBeforeApply: true,
  requireHumanApprovalForWrite: true,
  validationSandboxProfile: "none",
  maxFileBytes: 2_000_000,
  maxResponseBytes: 4_000_000,
  timeoutMs: 120_000
}

Environment: LOOM_MCP_ALLOW_WRITE, LOOM_MCP_REQUIRE_PLAN, LOOM_MCP_REQUIRE_HUMAN_APPROVAL, LOOM_MCP_SANDBOX_PROFILE, LOOM_MCP_AUDIT_PATH, LOOM_MCP_CONFIG

Plan → apply flow (Phase 2)

workspace_planEdit / deck_refactorSymbol    → returns planId + riskReport (registered in session)
[user / Deck approval]
workspace_applyEditPlan / deck_refactorSymbol → { planId, apply: true, humanApproved: true }

Agents must not invent plans or skip riskReport. Human approval (humanApproved: true) in the tool payload is ignored; approval must come from a trusted server-side source (like Deck UI or local CLI via mcp_approvePlan). High-risk writes remain strictly blocked unless approved.

Agent documentation: docs/AGENT_GUIDE.md (tooling guide, symbol IDs, MCP ↔ workspace mapping). Chat meta prompt guide (EN): docs/METAPROMPT.md (how to instruct normal chats to produce safe MCP plans). Chat meta prompt guide (DE): docs/METAPROMPT.de.md (wie man normale Chats fuer sichere MCP-Plaene anleitet). Reliability status: docs/REFACTOR_RELIABILITY_STATUS.md (implemented vs partial vs pending refactor/write guarantees).

Tool risk levels

Tool Mode
workspace_planEdit, deck_refactorSymbol (dry-run), deck_addTestCase (dry-run), deck_issueList, deck_issueView, deck_issueNext, deck_fileGuidanceView read / dry-run-write
workspace_applyEditPlan, deck_refactorSymbol (apply), deck_addTestCase (apply), deck_issueComment, deck_issueClose, deck_fileGuidanceAnnotate write
shell / git / live AI dangerous (not exposed by default)

Sandbox Validation

Validation during workspace_applyEditPlan, deck_refactorSymbol, or deck_addTestCase is safely routed through @loom/sandbox.

  • Argv Only: workspace_validate and other validation fields strictly require an array of strings (argv). Command strings are rejected to prevent shell-injection.
  • No Shell Parsing: No shell metacharacters or redirection operators are permitted in any argument segment.
  • Default Profile: none (execution blocked).
  • Other Profiles: shadow-only, mock, bwrap-fast, podman-test.
  • Policy: No host shell execution is allowed. Validation commands are executed via sandbox containers or lightweight process boundaries if explicitly allowed via validationSandboxProfile.
  • Allowlist: Only a predefined set of safe commands (e.g., bun test, tsc --noEmit) are permitted by default. Use allowDangerous: true to override.

Test-hang safety defaults (promote post-verify)

workspace_promoteExecutionTarget supports optional postVerify with compact and hang-safe behavior:

  • resultDetail: full | failures-only | summary (default: failures-only)
  • timeoutMs: per-command timeout
  • strictTestMode (default: true) for framework-safe args
  • testProfile: bun | vitest | jest | playwright

When strict mode is active, anti-watch/anti-hang flags are injected for known frameworks. Timed out commands are tagged with an explicit likely-hang warning.

Execution targets (isolated write/refactor loop)

Execution targets provide long-lived shadow copies under .loom/targets/<id> for safer mutation cycles:

  1. workspace_createExecutionTarget
  2. run checks in target (workspace_targetRunCommand)
  3. refactor/apply in target (workspace_refactorSymbol with options.executionTargetId)
  4. workspace_promoteExecutionTarget
  5. workspace_disposeExecutionTarget

Promote hardening:

  • dryRun preview (wouldCopyFiles, wouldDeleteFiles, conflictFiles)
  • strategy: copy-only | mirror-delete
  • conflict policy: abort | overwrite
  • baseline hash guards for source drift detection

Quickstart payload sequence

Minimal safe pipeline:

  1. workspace_createExecutionTarget
  2. workspace_refactorSymbol with options.executionTargetId (apply:true)
  3. workspace_promoteExecutionTarget with dryRun:true
  4. workspace_promoteExecutionTarget with apply:true + postVerify
  5. workspace_disposeExecutionTarget

Recommended promote postVerify shape:

{
  "postVerify": {
    "commands": [["bun", "run", "typecheck"]],
    "resultDetail": "failures-only",
    "profile": "mock",
    "timeoutMs": 60000,
    "strictTestMode": true,
    "testProfile": "bun"
  }
}

Failure playbook (promote + verify)

  1. dryRun shows conflictFiles:
  • keep conflictPolicy:"abort" default
  • repair in target and retry preview
  • use overwrite only intentionally
  1. Promote blocked by conflicts:
  • inspect target via workspace_targetRunCommand
  • either preserve source and rework target, or intentionally overwrite
  1. Post-verify timed out:
  • treat as likely hanging suite
  • narrow command scope
  • keep strictTestMode:true
  • prefer resultDetail:"failures-only"
  1. Post-verify failed:
  • use workspace_nextToolInfo
  • apply targeted repair flow
  • re-run targeted validation before broader checks
  1. Need rollback:
  • workspace_undoApply with partial modes (only / exclude)
  • then workspace_verifyRefactor

Decision matrix (fast operator mode)

Situation Tool Defaults
First orientation workspace_decideNext include compact decision signals
Dense one-call context workspace_capsule keep to minimum useful scope
Single mutation + feedback workspace_editWithFeedback reportMode:"compact-agent"
Multi-step automation workspace_executeActionTree stopOnFailure:true
Risky refactor execution target flow always dryRun before promote apply
Verify after promote workspace_promoteExecutionTarget.postVerify resultDetail:"failures-only", strictTestMode:true, timeout set
Recovery workspace_undoApply partial undo first

Commands

bun run mcp:start
bun run mcp:list
bun run mcp:config --target cursor
bun run skills:sync --package @loom/mcp

Phase roadmap

  1. Now — read + plan; apply blocked by default
  2. Now — apply with allowWrite + planId + humanApproved
  3. Next — Deck UI shows riskReport, user clicks Apply
  4. Later — dangerous tools in separate MCP profile only

Dependencies

Dependencies

ID Version
@deck/runtime ^0.1.0
@loom/analysis ^0.1.0
@loom/package ^0.1.0
@loom/sandbox ^0.1.0
@loom/workspace ^0.1.0
@nerve/pulse ^0.1.0
@nerve/search ^1.0.0

Development Dependencies

ID Version
@types/bun latest
@types/node ^22.15.30
typescript ^6.0.3
Details
npm
2026-07-26 20:22:43 +02:00
0
87 KiB
Assets (1)
mcp-0.1.1.tgz 87 KiB
Versions (3) View all
0.1.2 2026-07-26
0.1.1 2026-07-26
0.1.0 2026-07-26