docs: add unified confirm/revise with plan card to editable plan spec

Per-message Confirm buttons now auto-fill the corresponding plan card
field when a plan is visible, using field_agents mapping and keyword
detection from gap analysis categories.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

docs/superpowers/specs/2026-04-12-editable-plan-card-design.md

@@ -82,6 +82,38 @@ When the user clicks "Ask Agent" on a field:
 
 No new API endpoints — reuses existing `POST /api/chat` with `@mentions` routing.
 
+### 6. Unified Confirm/Revise with Plan Card
+
+Currently there are three disconnected approval mechanisms:
+- **Plan card** — Approve / Reject (whole plan)
+- **Wizard step 7** — Approve & Generate (same as plan card)
+- **Per-message** — Confirm / Revise on every agent response (sends chat message, updates design state indirectly via `extract_decisions()`)
+
+The per-message Confirm/Revise buttons must be aware of the plan card.
+
+**When a plan card is visible (phase = "planning"):**
+
+`confirmRecommendation()` behavior changes:
+1. Sends the confirmation message to chat (same as today — keeps design state updated)
+2. **Auto-fills the corresponding plan card field** — detects which field the agent message relates to using the `field_agents` mapping + keyword matching from `GapAnalysisConfig.category_keywords`
+3. Recalculates the plan card's current score
+
+Field detection logic (reuses existing gap analysis categories):
+- Agent is `engineering` + message contains dimension keywords → update `dimensions`
+- Agent is `engineering` + message contains material keywords → update `material`
+- Agent is `design` + message contains feature keywords → update `features`
+- Agent is `cnc` + message contains constraint keywords → update `constraints`
+- Agent is `cnc` + message contains axis/machining keywords → update `axis_recommendation`
+- If no field match is found, skip the auto-fill (just send the chat message as today)
+
+Value extraction: parse the confirmed message content for the relevant value (e.g., extract "Al 6061" from "I recommend Al 6061 for its machinability"). Uses simple heuristics — first quoted value, or value after "recommend"/"suggest"/"use".
+
+`reviseRecommendation()` — no change. Works as today (pre-fills input with context).
+
+**When no plan card is visible (phase = "exploring"):**
+
+Confirm/Revise works exactly as it does today — no plan card interaction.
+
 ## Data Model Changes
 
 ### `DesignPlan` (`agents/design_state.py`)
@@ -148,9 +180,13 @@ field_agents: dict[str, str] = {
 - `askAgentForField(fieldName)` — constructs mention-targeted message, calls `sendMessage()`, switches to chat tab
 - `savePlanAndContinue()` — merges edited values into `designState`, saves to localStorage, removes plan card, resets phase to exploring
 - `recalcPlanScore()` — reads current field values from plan card inputs, runs `computeScore()`, updates current score display
+- `detectPlanField(agentId, content)` — maps an agent message to a plan field using `field_agents` mapping + `category_keywords`. Returns field name or null.
+- `extractFieldValue(fieldName, content)` — parses agent message content to extract a value for the given field (quoted values, values after "recommend"/"suggest"/"use")
+- `updatePlanCardField(fieldName, value)` — programmatically sets a plan card field's value and triggers `recalcPlanScore()`
 
 ### Modified functions
 - `approvePlanCard()` — reads edited field values + notes from plan card inputs before sending to API
+- `confirmRecommendation()` — when phase is "planning", also calls `detectPlanField()` → `extractFieldValue()` → `updatePlanCardField()` to auto-fill the plan card
 - `computeScore()` — no logic change, just called by `recalcPlanScore()`
 
 ## Testing
@@ -162,3 +198,8 @@ field_agents: dict[str, str] = {
 - Verify Save & Continue preserves edits in designState and resets phase
 - Verify Reject discards all edits
 - Verify Approve sends edited values (not original values)
+- Verify Confirm on agent message auto-fills corresponding plan card field when plan is visible
+- Verify Confirm on agent message works as before when no plan card is visible
+- Verify field detection correctly maps agent + keyword to plan field
+- Verify value extraction handles common patterns (quoted, after "recommend", etc.)
+- Verify plan card score recalculates after Confirm auto-fill