Repository navigation
Description
⚠️ OpenSpec-managed issue — this content is automatically synced
from theopenspec/directory. Manual edits will be overwritten on next sync.
Artifacts
Summary
Product framing: every user action an app offers should also exist as an MCP tool, so any
action can in principle be automated by an AI agent — and chat becomes a way of commanding
the app. A chair should be able to say "open today's board meeting and put the budget
overrun on next week's draft agenda" and have decidiq execute it, with exactly the same
guards a click would hit, plus the agent-specific ones a click does not need.
What makes that sane now — and made it insane before — is hermiq's grant model: every tool
is classified on scope (read/create/update/delete, from OpenRegister's MCP
annotations) × reach (self/user/instance/external, ToolReachResolver), writes
are default-deny (ToolGrantResolver::requiresGrant(), fail-closed on anything it cannot
positively classify as a low-reach read), users grant tools per agent, per tool, optionally
argument-scoped, high-impact invocations pass a human approval gate
(ApprovalService), and dry-runs neutralise side-effecting calls
(ToolClassificationService::isSideEffecting()). decidiq-mcp-adoption withdrew
startMeeting precisely because, pre-model, it was failing open. This change enumerates
decidiq's real user actions from its controllers/services, ships one curated #[McpTool]
per action on the owning service, annotates every write honestly for the grant model, and
reinstates the meeting-lifecycle capability behind the approval gate — the safety layer
whose absence justified the withdrawal.
Specs
Tasks
- Implement
- Test
- Implement
- Test
- Implement
- Test
- Implement
- Test
- Implement
- Test
- Implement
- Test
- Implement
- Test
- Implement
- Test
- Grep checks: every
#[McpTool]underlib/declares an explicitscope(grep -rn "McpTool(" lib/ | grep -vc "scope:"= 0); noMcpToolattribute exists underlib/Mcp/; the D2 matrix row count equals the number of write attributes. -
php -lon every touched PHP file;python3 -m json.tool lib/Settings/decidesk_register.json. -
composer check:strictexits zero. - Hydra gates pass (
hydra-gatesskill / shared runner) — semantic-auth and no-admin-idor must show no new findings on the annotated services. -
openspec validate hermiq-ai-tooling --type change --strictpasses. - Live catalogue check on
/api/mcp: inventory ✅/♻ tool ids present;decidesk.castVoteabsent; derived surface unchanged (still 10 schemas × search/get). - PHPUnit unit tests — attribute/matrix conformance (Task 3), scannable-services two-way agreement (Task 5), withdrawal pins (Task 7), audit exactly-one-entry (Task 6), draft-pin and published-agenda refusal (Task 2), notification-rule condition (Task 1, JSON-level).
- Newman/Postman tests — N/A: no decidiq HTTP endpoint changes; the MCP surface is served by OpenRegister's
/api/mcp. - Browser tests (Playwright MCP) — the hermiq approval flow for
transitionMeeting(approve + reject branches) and chat flow (1); N/A for decidiq's own UI (unchanged). - All tests pass (
composer test), zero new failures against a self-measured baseline. - Feature documentation updated in
docs/— the full tool table with scope × reach × approval, the withdrawn actions and their reasons, and the granting walkthrough ("what your assistant may do, per agent"). - Screenshot captured and committed to
docs/images/— the approval card for a meeting transition. - N/A for tool
descriptionfields (agent-facing prose, not UI copy — same ruling as decidiq-mcp-adoption). Approval-card strings live in hermiq, not decidiq.
Design
See design.md for technical design details.
Synced from openspec/changes/hermiq-ai-tooling by OpenSpec workflow
App: decidiq
Activity