feat(readmymind): phase 3 PR 1: alternates, rethink steering, phone surfaces

- Alternate suggestions render as tappable rows that swap into the
  editable field, with their rationale as a visible second line (phones
  have no hover for a title tip). The predictor already returned up to 3
  kind-diverse suggestions; the modal showed only the first.
- Rethink now records every displayed suggestion as rejected and carries
  an optional free-text steer note (the API accepted steer since phase 2;
  the UI never collected it). Both reset on each open. The steer field
  stays available in the error phase: steering a failed run's retry is
  exactly when a note helps.
- The modal header names the target session: overview rows and the
  accessory key can open it for a session that is not the active tab.
- Phone surfaces: a keyboard-accessory 🧠 key (hidden unless
  readMyMindEnabled is ON and the active session is claude mode,
  re-derived after innerHTML rebuilds, settings applies, and session
  switches) and a Suggest strip on the phone overview's yellow waiting
  rows (waiting only: on red rows a dialog is on screen and input text
  would land in its menu; answer routing is phase 3 PR 2)
- The modal renders as a compact sheet on phones with 16px inputs (iOS
  zoom guard) and 44px tap targets
- Send/Insert/Rethink freeze during the loading phase so a stale
  suggestion cannot be sent mid-rethink
- test/readmymind-phase3-surfaces.test.ts pins the load-bearing facts in
  CI: key in BOTH bar layouts, [hidden] re-assertion over inline-flex,
  header button off phones, the waiting-only row gate, no terminal
  refocus from the readmymind action, steer outside the result div

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

docs/readmymind-plan.md

@@ -122,14 +122,15 @@ Agent use cases this unlocks: a lead session records intentions as the user stat
 
 ## Phases
 
-1. **Intent store + capture + intent endpoints + skill docs.** Immediately useful to agents even before any UI exists.
-2. **Context assembler + predictor + predict endpoint + desktop button/modal.** The feature as pitched. The assembler ships with all collectors it can serve from day one (transcript, intent, git, run-summary, siblings); the approvals collector activates when PR #245 lands.
-3. **Phone accessory key, rethink steering, alternates row.**
-4. Explicitly later: proactive predict-on-idle (ghost suggestion chip), auto-compaction of `recentPrompts` into `goals` via a cheap model, codex/gemini capture, cross-case "global" intent.
+1. **Intent store + capture + intent endpoints + skill docs.** Immediately useful to agents even before any UI exists. Shipped 1.16.1 (PR #253).
+2. **Context assembler + predictor + predict endpoint + desktop button/modal.** The feature as pitched. The assembler ships with all collectors it can serve from day one (transcript, intent, git, run-summary, siblings); the approvals collector activates when PR #245 lands. Shipped 1.16.2 (PR #256).
+3. **Phase 3, PR 1: alternates row, rethink steering, phone surfaces.** Alternate suggestions as tappable rows that swap into the field; Rethink records everything displayed as rejected and carries the optional steer note; keyboard-accessory 🧠 key (`refreshReadMyMind()`, re-derived after every innerHTML rebuild, settings apply, and session switch); `🧠 Suggest` strip on the phone overview's YELLOW waiting rows only. Red rows deliberately get no shortcut yet: a dialog is on screen there and text sent via `POST /input` would land in its menu.
+4. **Phase 3, PR 2: the approvals fusion** (the Cloudflare OS learnings tie-in, items 1-2 of `cloudflare-os-learnings-plan.md`). `GET /api/sessions/:id/recap`: a deterministic "what was done, simplified" catch-up (last assistant tail + recent tool one-liners + git state) reusing the phase-2 collectors verbatim; no model call, effectively an observation-ledger v0 whose data source can later swap to a real ledger. Surfaces: collapsible "What happened" on approval cards (lazy-fetched), recap in the modal's loading phase (read while opus thinks), 🧠 on approval cards and red overview rows. Answer-aware Send: the predictor output gains an optional `answer` option number (validated against the pending dialog's parsed options, dropped when invalid) and the modal routes dialog answers through `POST /api/approvals/:id/answer` (option digits / idle text), keeping `POST /input` only for dialog-free sessions.
+5. Explicitly later: proactive predict-on-idle (ghost suggestion chip), auto-compaction of `recentPrompts` into `goals` via a cheap model, codex/gemini capture, cross-case "global" intent, a model-written prose recap (deterministic-only in v1).
 
 ## Open questions
 
-- Should Rethink's rejected-suggestion memory persist across modal closes, or reset each open?
+- ~~Should Rethink's rejected-suggestion memory persist across modal closes, or reset each open?~~ Decided in phase 2 and kept: reset each open (a fresh open is a fresh question); the steer note resets with it.
 - Is a composer-adjacent placement (next to the toolbar Run controls) better than the header for discoverability?
 - Pending-dialog input (source #1) consumes the approvals-inbox store (PR #245, merged): the phase-2 collector reads pending items directly from `src/approval-inbox.ts`.
 

docs/readmymind.md

@@ -23,11 +23,11 @@ Add `-u user:password` if your install has `CODEMAN_PASSWORD` set, and drop `-k`
 
 ## The 🧠 button
 
-On a Claude session, press the brain button in the header (desktop; the phone surface is a planned keyboard-accessory key). Codeman assembles everything it already knows: your goals, your recent prompts (with your voice: length, tone, shorthand), the tail of the last assistant reply, recent tool activity, git state (branch, dirty files, pending changesets), how long you have been away and what happened meanwhile, sibling sessions in the same case, and any dialog the session is currently waiting on. A one-shot model call (opus by default, `readMyMindModel` to override) turns that into 1-3 suggestions; the top one lands in an editable field with its rationale.
+On a Claude session, press the brain button in the header (desktop). On phones the same modal opens from the 🧠 key on the keyboard accessory bar, or from the `🧠 Suggest` strip under a yellow waiting row on the phone overview home screen (steer a session without opening its tab). Codeman assembles everything it already knows: your goals, your recent prompts (with your voice: length, tone, shorthand), the tail of the last assistant reply, recent tool activity, git state (branch, dirty files, pending changesets), how long you have been away and what happened meanwhile, sibling sessions in the same case, and any dialog the session is currently waiting on. A one-shot model call (opus by default, `readMyMindModel` to override) turns that into 1-3 kind-diverse suggestions (continue / verify / redirect); the top one lands in an editable field with its rationale, and the others render as tappable alternate rows that swap into the field.
 
 - **Send** submits it to the session (with Enter).
 - **Insert** drops it on the CLI composer *without* Enter, so you can edit it in the terminal before sending.
-- **Rethink** re-runs with the shown suggestion recorded as rejected.
+- **Rethink** re-runs with everything currently displayed recorded as rejected. The optional steer field above the buttons ("no, I meant the mobile bug") tells the re-run what you actually meant; it is your own words and outranks everything the predictor observed.
 - **Dismiss** closes; nothing happens.
 
 A prediction takes 5-90 seconds and costs real tokens; one runs per session at a time. If the session is sitting on a permission/question dialog, the suggestion is usually an answer to that dialog: that is intentional.
@@ -87,13 +87,14 @@ The `codeman` agent skill documents the same verbs (SKILL.md §3 plus `reference
 
 ## What comes next (phase 3+)
 
-Phone keyboard-accessory 🧠 key, a steer-note input on Rethink, and tappable alternate suggestions. Explicitly later: proactive predict-on-idle, auto-compaction of the prompt history into goals, non-Claude capture. See the phases section of [`readmymind-plan.md`](readmymind-plan.md).
+The approvals fusion: a per-session catch-up recap ("what was done, simplified", derived from the same collectors the predictor uses), shown on Approvals Inbox cards next to a 🧠 button, plus answer-aware Send that routes dialog answers through the approvals endpoint instead of typing into a menu. Explicitly later: proactive predict-on-idle, auto-compaction of the prompt history into goals, non-Claude capture. See the phases section of [`readmymind-plan.md`](readmymind-plan.md).
 
 ## Troubleshooting
 
 | Symptom | Cause / fix |
 | ------- | ----------- |
-| No 🧠 button in the header | `readMyMindEnabled` is OFF (App Settings → Panels), you are on a phone (desktop-only in this phase), or the active session is not claude-mode |
+| No 🧠 button in the header | `readMyMindEnabled` is OFF (App Settings → Panels), you are on a phone (the header button is desktop-only; phones use the keyboard-accessory 🧠 key and the overview waiting rows), or the active session is not claude-mode |
+| No 🧠 key on the phone keyboard bar | `readMyMindEnabled` is OFF, or the active session is not claude-mode (the key hides itself for codex/gemini/opencode/antigravity/shell sessions) |
 | Prediction feels generic | The profile is thin: record goals (PUT or ask your agent to), and let capture accumulate a few real prompts first |
 | "A prediction is already running" (409) | One per session at a time; wait for the current one (up to 90 s) |
 | Prediction fails (502) | The model returned no usable JSON, or the CLI could not start; retry. Check `readMyMindModel` if you overrode it |