v0.6.4.0 feat(assistant): action-intent routing through decision pipeline (#148 v1) by jayzalowitz · Pull Request #153 · jayzalowitz/skytwin

jayzalowitz · 2026-05-05T20:12:37Z

Summary

Closes #148 v1 — the last phase-2 piece for the assistant epic. Saying "archive that email" or "schedule a meeting with X" in chat now creates an `ApprovalRequest` on the existing `#/approvals` page. Conversational messages still go through the LLM chat path unchanged.

Phase 2 epic complete: #146 streaming + #147 context + #149 multi-turn + #148 action routing.

Conservative v1 design choice

Chat-driven actions ALWAYS land in approvals, never auto-execute. Even when `DecisionMaker.evaluate()` returns `autoExecute=true`, the chat router collapses that to `requiresApproval=true` for v1.

Why: chat is a free-text channel. An unintended intent match (regex firing on a message ABOUT scheduling instead of asking to schedule) cannot trigger a real send / spend / modify. The `/api/events` route has structured signals as ground truth; chat doesn't.

The user already has the `#/approvals` UI. Routing through it preserves the audit trail and feedback loop unchanged.

Phase 2 of #148 lifts this restriction once we have an LLM-classifier confidence score AND an explicit per-user opt-in.

Safety Invariants — all upheld

feat: bootstrap SkyTwin monorepo (v0.1.0.0) #1 (no auto-exec without policy check): every chat-driven action runs through `DecisionMaker.evaluate()` → `PolicyEvaluator.evaluate()`. No bypass.
Make SkyTwin usable by a non-technical person end-to-end #2 (always log explanations): `ExplanationGenerator.generate()` runs and persists for every chat-driven decision. Persist failure logs a warn but doesn't abort the user's request (audit-trail loss is recoverable; aborting isn't).
Make SkyTwin actually work end-to-end: adapter layer, execution persistence, safety gaps #3 (trust tiers): pulled from `users.trust_tier` row, never from chat input.
Scope Expansion: Full Build Plan (CEO Review 2026-04-01) #4 (spend limits hard): inherited from `DecisionMaker` — no chat-specific override.
feat: milestone 1.5 scope expansion (v0.2.0.0) #5 (reversible honest): inherited from candidate generation.
feat: real OpenClaw execution, 6 new domains, expanded onboarding #7 (every action has RiskAssessment): inherited from `DecisionMaker.evaluate`.

What landed

Rule-based intent classifier (`@skytwin/assistant`)

Pure regex/keyword matching for a small vocabulary:

Intent	Pattern	Maps to
`archive_email`	"archive that/this/the/it"	`email_triage`
`label_email`	"label/tag that as X" (captures label)	`email_triage`
`send_reply`	"reply to / respond to / send a reply"	`email_triage`
`create_event`	"schedule/book/set up a meeting/call"	`calendar_invite`
`decline_event`	"decline/skip/cancel that meeting"	`calendar_update`
`create_task`	"remind me to / add a task to"	`task_management`

False-positive guards (the load-bearing test surface):

Returns null for messages < 8 chars ("ok", "thanks", "sure")
Returns null for meta-discussion patterns ("how do you decide when to schedule things?")
Returns null for action verbs without an object phrase ("I have a large archive of files")

Pipeline wiring

`buildActionRouter()` constructs the same `TwinService + PolicyEvaluator + DecisionMaker + ExplanationGenerator` stack as `events.ts`. Reuses the issue #122 `LabelInferencePort` so chat-driven `label_email` candidates get the same learned-from-history confidence boost.

The route adapter:

Builds a synthetic `DecisionObject` (`chat_intent_`, rawData includes `triggerMessage`).
Persists the decision (FK target for downstream rows).
Builds full `DecisionContext` — same shape as events.ts (preferences + patterns + traits + temporalProfile, fetched in parallel).
Calls `DecisionMaker.evaluate(context)` — every safety gate fires here.
Generates + persists `ExplanationRecord`.
If `selectedAction`: persists `ApprovalRequest` (filters `accessToken` + `rawData` from parameters).
Emits `approval:new` SSE so the existing badge updates immediately.

Route branching

`POST /api/assistant/messages` now tries intent classification BEFORE the LLM call. If routed:

Persists a structured assistant message with `metadata.intentRoute` carrying the outcome.
Both sync JSON and SSE response shapes supported. SSE flushes `thread` + `user` + `done` in one shot — no chunk events because no streaming text was generated.

If no intent matches OR the router throws (graceful degradation), falls through to the existing LLM chat path.

Web

`pages/assistant.js:renderMessages` checks `m.metadata.intentRoute` and renders an action footer:
- `requires-approval` → "Open approval →" link to `#/approvals`
- `blocked` → muted "Action blocked by your safety policy" notice
`assistant.css` — small accent-colored footer styles using existing theme variables.
Empty-state copy updated to mention the new capability.

Test plan

`pnpm build` — 20 packages green
`pnpm test` — 40 packages green; 16 new tests (12 intent classifier + 4 routeIntent)
`pnpm lint` — clean
Manual: configure provider, navigate to `#/assistant`, type "archive that email". Verify (a) approval lands in `#/approvals`, (b) chat shows "Open approval" footer, (c) the bubble explains what was queued.
Manual: type "how does archiving work?" — verify it falls through to the LLM chat reply (false-positive guard working).
Manual: with trust tier OBSERVER (which blocks most actions), verify the chat surfaces "Action blocked by your safety policy" instead of an approval link.
Manual: stop CRDB, type an intent — verify chat falls through to LLM reply (graceful degradation), no crash.

Phase 2 epic — done

Issue	Title	Status
#146	SSE streaming	✅ shipped
#147	Twin profile + memory context	✅ shipped
#149	Multi-turn LlmClient API	✅ shipped
#148	Action-intent routing	✅ this PR (v1; phase 2 follow-ups noted)

🤖 Generated with Claude Code

…line (#148 v1) Closes #148 v1 — final phase-2 piece for the assistant epic. Chat now routes detected action intents ("archive that email", "schedule a meeting") through the existing decision pipeline. Saying an action in chat creates an ApprovalRequest on the existing #/approvals page; conversational messages still go through the LLM chat path unchanged. Conservative v1 — chat-driven actions ALWAYS land in approvals, never auto-execute, even when the engine returns autoExecute=true. Free-text is too ambiguous to bypass the approval step on the first cut. Phase 2 of #148 lifts this when we have an LLM-confidence score + per-user opt-in. Safety invariants — all upheld: - #1 (no auto-exec without policy): every intent runs through DecisionMaker.evaluate() → PolicyEvaluator.evaluate(). No bypass. - #2 (always log explanations): ExplanationGenerator.generate() persists for every chat-driven decision. Persist failure logs but doesn't abort. - #3 (trust tiers): pulled from user record, never from chat input. - #4-#7 inherited from DecisionMaker. Pieces: - @skytwin/assistant: detectIntent (rule-based regex/keyword classifier, tolerant to short/ambiguous messages, false-positive guarded) + ActionRouter port + AssistantService.routeIntent. Package stays free of decision-engine + db deps. - apps/api/src/routes/assistant.ts: buildActionRouter() factory wires TwinService + PolicyEvaluator + DecisionMaker + ExplanationGenerator + LabelInferencePort. Synthetic DecisionObject from chat intent. Persists ApprovalRequest. Emits approval:new SSE. - POST /messages branches on intent BEFORE the LLM call. Both sync JSON and SSE response paths supported. - pages/assistant.js: action footer renders approval-link or blocked notice based on metadata.intentRoute. CSS styled with theme variables. Tests: 16 new (12 intent classifier + 4 routeIntent). Full suite green across 40 packages; lint clean. Phase 2 epic is now complete: #146 streaming + #147 context + #149 multi-turn + #148 action routing. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Copilot

Pull request overview

This PR completes assistant phase 2c (#148 v1) by adding rule-based action-intent detection on chat messages and routing matched intents through the existing decision/policy/explanations pipeline, always creating an ApprovalRequest (never auto-executing) and surfacing the result in the assistant UI.

Changes:

Added a rule-based detectIntent() classifier and ActionRouter port to keep @skytwin/assistant decoupled from decision/db deps.
Wired POST /api/assistant/messages to attempt intent routing before the LLM call and persist an assistant message with metadata.intentRoute when routed.
Updated the web assistant page to render an “Open approval” footer (or “blocked” notice) and bumped version/changelog for 0.6.4.0.

Reviewed changes

Copilot reviewed 11 out of 11 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
`VERSION`	Bumps release version to `0.6.4.0`.
`CHANGELOG.md`	Documents assistant action-intent routing v1 and safety posture.
`packages/assistant/src/intent-classifier.ts`	Adds regex-based intent detection producing an `ActionIntent`.
`packages/assistant/src/action-router.ts`	Introduces `ActionRouter` port and `ActionRouteOutcome` contract.
`packages/assistant/src/assistant-service.ts`	Adds `routeIntent()` to classify + route before LLM reply (graceful degradation).
`packages/assistant/src/index.ts`	Re-exports intent/router API.
`packages/assistant/src/__tests__/intent-classifier.test.ts`	Adds positive + negative (false-positive guard) tests for classifier.
`packages/assistant/src/__tests__/assistant-service.test.ts`	Adds `routeIntent()` behavior tests.
`apps/api/src/routes/assistant.ts`	Implements `buildActionRouter()` and routes matched intents ahead of LLM.
`apps/web/public/js/pages/assistant.js`	Renders action footer based on `metadata.intentRoute`.
`apps/web/public/css/assistant.css`	Styles the action footer/link.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

+      const decision: DecisionObject = {
+        id: `chat_intent_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`,
+        situationType: intent.situationType as SituationType,
+        domain: intent.domain,
+        urgency: 'medium',
+        summary: intent.summary,
+        rawData: { ...intent.rawData, triggerMessage: intent.triggerMessage },
+        interpretedAt: new Date(),


+      // No selected action OR every candidate was denied → blocked.
+      if (!outcome.selectedAction) {
+        return {
+          kind: 'blocked',
+          reason: outcome.reasoning || 'No suitable action could be taken right now.',


+      const {
+        accessToken: _omitToken,
+        rawData: _omitRawData,
+        ...visibleParameters
+      } = (outcome.selectedAction.parameters ?? {}) as Record<string, unknown>;


+   * lookup failure). The `AssistantService.routeIntent` caller catches
+   * that and returns `{ kind: 'no-action' }` so the chat falls through
+   * to the LLM reply rather than blowing up the whole turn.


Copilot AI review requested due to automatic review settings May 5, 2026 20:12

Copilot started reviewing on behalf of jayzalowitz May 5, 2026 20:13 View session

Copilot AI reviewed May 5, 2026

View reviewed changes

jayzalowitz merged commit a4c74e0 into main May 5, 2026
12 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

v0.6.4.0 feat(assistant): action-intent routing through decision pipeline (#148 v1)#153

v0.6.4.0 feat(assistant): action-intent routing through decision pipeline (#148 v1)#153
jayzalowitz merged 1 commit into
mainfrom
jayzalowitz/assistant-actions

jayzalowitz commented May 5, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

jayzalowitz commented May 5, 2026

Summary

Conservative v1 design choice

Safety Invariants — all upheld

What landed

Rule-based intent classifier (`@skytwin/assistant`)

Pipeline wiring

Route branching

Web

Test plan

Phase 2 epic — done

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants