feat(dashboard): Review replaces Learning and owns every pending decision

[kmonsoe]

Follow-ups from the stack-wide code review, stacked on #872.

Backend

• The conversation list pager's cursor now carries needsHumanAttention, so resuming across the attention boundary drops no rows; the ORDER BY tiebreaker is id to match the cursor. The queue pager strips the flag (its ordering never used it). New boundary-crossing test.
• list() and queue() share one query parser; the invalid-status error gains its conv_invalid: prefix.

Dashboard

• Detail-load failures render a translated error with a retry instead of an eternal spinner.
• The mobile full-screen composer declares role=dialog/aria-modal, closes on Escape, and focuses the textarea on open (see the draft exception below).
• The queue scroll fade is a CSS variable on the scroll container — scrolling no longer re-renders every row.
• Learning: failed decisions fetch is surfaced with a retry; published/dismissed counts are labelled as recent.
• One generic extendNavGroups behind the console/settings wrappers; agent settings gating keys off ACCOUNT_SETTINGS_HREF; the live-now dot halo derives from the accent token in both themes.

Runtime

• SetDraftReplyOpts exported once and reused by the in-process client.

---

Added in 6a848ad

Topic descriptions (migration 0089, nullable conv_topics.description)

Picking a topic is an automation decision — a topic mode overrides the conversation's — and the classifier was making it from a name and its kebab-case slug, the same word twice. Nothing told it where an org draws the line between Support, Technical and Security, or that an existing topic already covered the case it was about to duplicate.

• Returned by conv_list_topics and conv_list_topic_automation; set on conv_create_topic; edited by the new conv_update_topic tool or POST /v1/conversations/topics/:topicId. Slug stays immutable — it is how export/import addresses a topic.
• skill://conv/set-topic-and-title reads descriptions ahead of names, treats one that rules a case out as decisive, and writes a description for every topic it creates — but never rewrites one it did not author, since one conversation is not evidence enough to redraw an operator's boundary. setup-email-and-widget-channels' topic examples were missing the required slug; fixed while there.
• Editable in the automation page's policy dialog. Fixtures regenerated (219 tools).

Read-model contract for the auto-send gate

agentMode on a read is the gated effective mode; topicAgentMode is what an operator configured. The inbox row shows the first, the automation page the second. The old single test asserted an auto topic reads back as auto everywhere — the pre-gate contract, passing only because nothing evaluated the threshold yet. Now two tests: the override rule uses off (the one mode the gate cannot veto), and a second pins the fold.

Mobile oversight

• You can no longer act on a draft you were never shown. The draft lives in the composer textarea and pending drafts are filtered out of the thread, so on a phone Approve & send sat above a draft that appeared nowhere. The collapsed bar is now a single Review draft →; approve and reject both live in the editor it opens. Blind rejection was the quieter half — a reject counts against the topic's approved-unedited share, moving the auto-send gate on evidence nobody read.
• The no-draft bar is a button (Write reply → / Continue reply →) instead of markup styled as a text input, which read as somewhere to type when tapping it only ever opened the editor.
• Opening the editor onto an unedited draft no longer focuses the textarea: focus raises the keyboard over the text you opened the screen to read. Note tab, empty composer, and already-edited drafts focus as before.
• Editor gutters: the tab row and error banner were the only strips without a mobile value, so they kept the desktop 20px against a 16px header and body.
• The automation table stacks under md instead of scrolling sideways at every phone width — which put the policy, the reason you opened the page, off-screen. Weekly volume and the percentage bar are dropped there rather than shrunk. Desktop layout is unchanged; every rule is behind md:/max-md:.

Verified: workspace typecheck, backend-core conv 453, db 24 (journal guard included), dashboard-pages 101, en/nb key parity. Migration 0089 applied to a DB already at 0088 and exercised through the non-superuser munin_app role under RLS.

---

Added in 49fb890 — Review replaces Learning

/dashboard/learning becomes /dashboard/review, and the page stops being only about knowledge proposals. CMS drafts, CRM merge proposals, outreach drafts and forwarded feedback leave the admin overview and join the KB candidates, split by urgency:

• Blocking — everything that stops something shipping, sorted oldest first. A blocking queue sorted by recency buries exactly the item that has been ignored longest, which is the opposite of what the section is for.
• Improvements — knowledge proposals, newest first. No deadline: the base already answers customers, and a proposal only makes it answer better. Keeping them in one list with an unsent email made them look overdue; tinting or dimming the rows overcorrected, so the section label carries it.
• Decided · last 7 days stays as a third section.

The overview drops the queue concept entirely — no list, no hero count, no stat row. The sidebar badge on Review is the single signal that work is pending, and it counts all five kinds instead of only KB candidates.

Selecting a blocking item renders its detail in the split's right pane rather than a sheet, so the CMS field editor, outreach scheduling and the merge preview all work unchanged and are now deep-linkable at /dashboard/review/:id. DrawerHeader's close button became optional for that: in a persistent pane it pointed nowhere, because the list immediately re-selects the first row.

Outreach gets a purpose-built pane, ordered by what must not go wrong — the message as it will arrive, then why it is being sent, then when.

• The envelope is rendered, and so is everything the send path appends: the campaign's CTA URL and the fixed --- / Unsubscribe footer, as the literal text composeOutreachBody will append, under one "appended on send" marker. Note the CTA is a bare URL on its own line — not the styled button the design drew.
• "Why this, why now" now exists. evidence was already one call away on /v1/outreach/proposals/:id and the dashboard never fetched it. It is freeform jsonb whose shape varies per drafting agent, so it renders by shape: long values and reason-ish keys become prose, kb:// / kdoc_ references become chips, the rest become labelled chips. A parser keyed to the shape the drafting skills document rendered nothing for the proposals already in the dev database — that is what the new tests pin.
• Send timing is a control on the page, not a dialog behind a button. The cadence annotations the design asks for beside it need campaign cadenceRules, which are not on the DTO, so they are omitted rather than faked.
• SMS and voice are first-class: no subject row, no CTA or unsubscribe appends, a live segment count as you edit, and the sender as a number. A contact with no address on file blocks approval outright and says why, instead of tinting a note and failing at the service.
• The full diff sits behind a disclosure (N edits against the agent's original), with See what changed → in the revised-after-review banner. The final text is already rendered above it in full, so inline it was taller than the message and pushed the timing control off-screen.
• The old reply-thread quote is dropped, not carried over — it quoted the proposal's own snippet, not the inbound message it was replying to. Showing the real thread needs the conversation fetch.

DTO additions (all read-only, all off data the queries already had): ProposalCampaignSummary.ctaUrl (previously selected only to compute appendsCta, then discarded), ProposalDelivery.sender / senderName from the channel's own addressing config routed through publicChannelConfig so no credential can reach a DTO, and ProposalContactSummary.companyName.

Pill gains a marker variant so a pane's module pill carries the same glyph as its list row — square for CMS, diamond for CRM, hollow ring for KB — instead of the generic filled dot every tone rendered. BodyDiff gains wrap: these are prose bodies in a ~700px pane, where horizontal scrolling through an email diff is unusable.

Heads-up for cloud: the route rename is breaking for apps/web-cloud, which mirrors these routes itself and imports LearningPage. It needs dashboard/review/{page,[id]/page}.tsx and ReviewPage before it builds against the next @getmunin/dashboard-pages.

Verified: workspace typecheck, dashboard-pages 136 (11 new: review partition + outreach evidence parsing), backend-core outreach unit tests, en/nb key parity with no unused or missing keys. Every pane state exercised against the dev database in light, dark and at mobile width — email initial with a diff, email follow-up, SMS follow-up, and an SMS proposal for a contact with no phone.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NEtUitV5jQ1nbmwmtTML3u