@yemi33/minions 0.1.1949 → 0.1.1951

package/docs/completion-reports.md

@@ -19,6 +19,28 @@ Path shape (resolved by `shared.dispatchCompletionReportPath()` in `engine/share
 
 The agent must write the JSON to that exact path before exiting. Any character outside `[a-zA-Z0-9._-]` in the dispatch id is replaced with `-` by the engine when computing the path.
 
+## Trust boundary
+
+Each spawn also receives a per-dispatch cryptographic value via the `MINIONS_COMPLETION_NONCE` environment variable. The engine generates this with `crypto.randomBytes(16).toString('hex')` in `engine.js:spawnAgent()` and stores it on the in-memory active-process record. The agent is required to copy the value verbatim into the report's `nonce` field. On parse, `engine/lifecycle.js:runPostCompletionHooks()` compares `report.nonce` against the in-memory value:
+
+- **Match** — the report is trusted and processed normally.
+- **Mismatch** — the report is treated as forged (a prompt-injected agent or a stale process writing into a sibling dispatch's completion path). Every signal it carries — `status`, `pr`, `noop`, `failure_class`, `retryable`, `needs_rerun`, fenced/summary fallbacks — is discarded. The dispatch is failed with `failure_class: 'completion-nonce-mismatch'` and the work item is marked failed (no auto-retry honors the agent's `retryable` claim).
+- **Missing** — by default, the engine logs `[security] completion-nonce-missing dispatch=… required=false (degraded — report honored)` and still honors the report. Flip `ENGINE_DEFAULTS.completionNonceRequired` (or `engine.completionNonceRequired` in `config.json`) to `true` to hard-fail missing nonces too. Default is `false` for one release so older runtime caches and agents that haven't picked up the prompt change degrade with a warning instead of breaking.
+
+Security event log lines are emitted on the `error` channel and are designed to be greppable:
+
+```
+[security] completion-nonce-missing dispatch=<id> agent=<name> wi=<wiId> required=<bool>
+[security] completion-nonce-mismatch dispatch=<id> agent=<name> wi=<wiId> expected=<8char> got=<8char>
+```
+
+**Migration path.** Roll out in two phases:
+
+1. **Phase 1 (default):** Ship the engine + playbook change with `completionNonceRequired: false`. Agents that echo the nonce are validated; agents that don't continue to work but emit a warning per dispatch. Watch the `[security] completion-nonce-missing` warnings drain to zero across runtimes.
+2. **Phase 2:** Once warnings are quiet for at least one release window, flip `ENGINE_DEFAULTS.completionNonceRequired` to `true` (or set per-deployment via `config.json` → `engine.completionNonceRequired`). Missing nonces then hard-fail like mismatched ones.
+
+Do **not** invent, regenerate, or share the nonce across dispatches — each spawn gets a unique value.
+
 ## Top-level schema
 
 ```json
@@ -31,6 +53,7 @@ The agent must write the JSON to that exact path before exiting. Any character o
   "retryable": false,
   "needs_rerun": false,
   "noop": false,
+  "nonce": "<value of MINIONS_COMPLETION_NONCE env var>",
   "artifacts": [
     {"type": "pr", "path": "https://github.com/owner/repo/pull/123", "title": "PR-123"}
   ]
@@ -49,6 +72,7 @@ The agent must write the JSON to that exact path before exiting. Any character o
 | `retryable` | boolean | `true` if the engine should auto-retry the dispatch on failure. Overrides the default per-class retry policy when present. |
 | `needs_rerun` | boolean | `true` if the same work needs to be re-dispatched (vs. retried). Used by build-fix and review-fix loops. |
 | `artifacts` | array | Durable artifacts the agent created or updated; surfaces in the dashboard work-item detail modal. See [Artifacts](#artifacts). |
+| `nonce` | string | Per-spawn cryptographic value the engine injects via `MINIONS_COMPLETION_NONCE`. Copy the env var value verbatim; the engine validates it on read. See [Trust boundary](#trust-boundary). |
 
 ### Optional fields
 
@@ -76,6 +100,7 @@ Defined in `engine/shared.js` as `FAILURE_CLASS`. Use the canonical hyphenated s
 | `network-error` | API rate limit, DNS, connectivity | Default retry logic |
 | `out-of-context` | Context window exhausted | Flag for human review |
 | `max-turns` | Claude CLI `error_max_turns` — work in progress | Retry same agent |
+| `completion-nonce-mismatch` | Completion JSON missing or mismatched `nonce` (forged completion). See [Trust boundary](#trust-boundary). | Never retry (untrusted) |
 | `unknown` | Unclassified failure | Default retry logic |
 
 Use `"N/A"` when `status` is `success` or `partial` without a failure.

package/docs/design-state-storage.md

@@ -46,7 +46,7 @@ release .lock file
 Key properties:
 - **Synchronous blocking** — `withFileLock` spins with `sleepMs(25)` until lock acquired or 5s timeout (source: `engine/shared.js:175-231`)
 - **Whole-file granularity** — updating one field in one work item rewrites all 180 items (370 KB)
-- **Stale lock recovery** — locks older than 60s are force-removed (source: `engine/shared.js:173`, `LOCK_STALE_MS`)
+- **Stale lock recovery** — locks older than 5 min (`LOCK_STALE_MS = 300_000`) are force-removed; holders that recorded a `{pid, ts}` payload are kept alive past the threshold while `process.kill(pid, 0)` succeeds, with a hard last-resort cap at 5×LOCK_STALE_MS (source: `engine/shared.js`, P-b7d4e8f2)
 - **Read caching** — only `dispatch.json` has a 2s TTL cache (source: `engine/queries.js:82-91`)
 
 ### 1.3 Read vs Write Ratio

package/docs/slim-ux/architecture-suggestions.md

@@ -0,0 +1,467 @@
+# Slim UX — Architectural Suggestions
+
+> Written after Round 1 (Phase A concept mapping + Phase B layout rebuild). These are **proposals**
+> for engineering changes that would make the slim UX cheaper to maintain and more honest about
+> what's happening under the hood. **None of these are implemented in Round 1.** They're for
+> human discussion before any of them get prioritized.
+>
+> Format per section: **Problem → Proposal → Cost → Risk**.
+>
+> Cost is rough person-days. Risk is what could go wrong.
+
+---
+
+## 1. A `/api/cockpit` aggregate endpoint
+
+**Problem.** Right now the slim cockpit has to hit `/api/status` every 5 seconds and pluck six
+fields out of a ~60 KB blob (`dispatch.active`, `dispatch.pending`, `pullRequests`, `watches`,
+`engine`, `dispatch.completed`). The full status payload includes `agents`, `inbox`, `notes`,
+`metrics`, `prdProgress`, `verifyGuides`, `archivedPrds`, `skills`, `mcpServers`, `schedules`,
+`pipelines`, `pinned`, `projects`, `autoMode`, `version` — all of which the cockpit ignores.
+Polling that whole thing every 5 s on every open tab is wasteful.
+
+**Proposal.** A new `/api/cockpit` that returns *only* the cockpit-relevant aggregates:
+
+```jsonc
+{
+  "engine":   { "running": true, "mode": "running", "startedAt": "…" },
+  "dispatches": { "active": 2, "pending": 5, "workingAgents": ["dallas", "ralph"] },
+  "prs":      { "active": 3, "failingBuilds": 1 },
+  "watches":  { "active": 4, "triggeredRecently": 1 },
+  "schedules":{ "dueWithinHour": 2 },
+  "lastEvents": [ { "kind": "completion", "ts": "…", "title": "…" }, … ]
+}
+```
+
+Compute it from the existing fast-state cache in `dashboard.js` (`_fastState`) so the cost is
+near-zero — just a few field projections. Slim polls this; full dashboard keeps `/api/status`.
+
+**Cost.** ~0.5 day. New handler + a couple of unit tests covering aggregate math.
+
+**Risk.** Low. Pure read-side; doesn't touch any state. The only failure mode is the slim
+showing stale numbers if the cache pruner gets confused, but that's the same risk `/api/status`
+already has.
+
+---
+
+## 2. SSE for cockpit updates instead of polling
+
+**Problem.** 5-second polling is fine for Round 1 but feels laggy when an agent finishes — you
+see "Sending…" disappear in the chat 30 s before "Active dispatches" updates from 1 to 0. The
+full dashboard already has `/api/status-stream` (SSE); slim just doesn't use it.
+
+**Proposal.** Slim subscribes to `/api/status-stream` (or a new `/api/cockpit-stream` that emits
+deltas only). Polling becomes a fallback for browsers without SSE. Visibility-change pause
+already exists in slim (round-1 code), so background tabs cost nothing.
+
+**Cost.** ~0.5 day if reusing `/api/status-stream`; ~1.5 days if writing a delta-encoded
+`/api/cockpit-stream`.
+
+**Risk.** Medium. SSE connections occasionally drop and need exponential-backoff reconnect.
+Multi-tab leaks are a real concern (each tab opens its own EventSource). Need a small connection
+manager on the client.
+
+---
+
+## 3. A unified event stream for the History panel
+
+**Problem.** The History panel currently merges three sources client-side: `dispatch.active`,
+`dispatch.completed`, and `pullRequests`. Each has its own timestamp field (`startedAt`,
+`completedAt`, `updatedAt`), its own shape, and its own truthiness rules. The merge logic is
+fragile — adding a fourth source (e.g. consolidation runs, watch fires, schedule executions,
+pipeline state changes) means more glue.
+
+**Proposal.** A canonical `engine/events.jsonl` (append-only, rotated weekly) with a uniform
+shape:
+
+```jsonc
+{ "ts": "2026-05-08T15:23:11Z", "kind": "completion", "agent": "dallas",
+  "title": "fix slim button", "workItemId": "W-…", "pr": "#42",
+  "summary": "…", "level": "info" }
+```
+
+Plus `/api/events?since=…&limit=…` for paged reads. The engine already writes most of these
+moments to `engine/log.json` (the audit ring buffer); we just need to enrich them and expose
+them. The History panel becomes one fetch.
+
+**Cost.** ~1.5 days. The hard part is catching every event-emit site — dispatch start, dispatch
+complete, PR sync, watch fire, schedule run, pipeline transition, meeting round advance,
+consolidation run. Mostly mechanical.
+
+**Risk.** Medium-low. Events are reads; the risk is missing one. Add a unit test that asserts
+each lifecycle path emits an event.
+
+---
+
+## 4. Should Work Items + Pipelines merge in the data model?
+
+**Problem.** Carlos called out "Work Items" and "Pipelines" as conceptually overlapping in the
+existing UI. Looking at the data: a Pipeline is a sequence of stages where each stage *is*
+either a Work Item, a Meeting, or a Plan. So a Pipeline isn't really a separate concept — it's
+a Work Item with a stage list. Yet they have entirely separate stores
+(`engine/dispatch.json` vs `engine/pipeline-runs.json`), separate API surfaces
+(`/api/work-items*` vs `/api/pipelines*`), and separate UIs.
+
+**Proposal.** Don't merge the *storage* — that's a deep migration. Instead, introduce a
+"workspace" view at the API layer that returns Work Items and Pipeline runs in one paginated
+list, tagged with `kind: "work-item" | "pipeline-run" | "pipeline-stage"`, sorted by recency.
+Slim's "Work" button opens this unified view; it can drill down into stages when the user
+expands a pipeline.
+
+**Cost.** ~2 days. Read-only aggregator + a UI that knows how to render the three flavors. No
+storage changes.
+
+**Risk.** Low at the engine layer (read-only); medium at the UX layer (pipelines have richer
+state — wait conditions, retriggers — that don't map cleanly to a flat list).
+
+---
+
+## 5. Plan + PRD: collapse to one document?
+
+**Problem.** Carlos asked "What is a PRD? Should I generate one to get started?" The answer is
+"no, PRDs are auto-generated from approved plans by the `plan-to-prd` agent" — but that's
+opaque. The two-stage flow (plan → PRD → work items) exists because the plan is a human
+discussion artifact and the PRD is the machine-readable execution contract.
+
+**Proposal.** Don't drop the distinction; *unify the surface*. Treat the plan as the primary
+document and embed the PRD JSON as a fenced block inside the plan markdown:
+
+```markdown
+# Plan: Refresh the Settings sidebar
+
+## Goal
+Three small commits, no big-bang refactor.
+
+## Acceptance criteria
+- …
+
+```prd
+{ "items": [ … ] }
+```
+```
+
+The materializer reads the fenced block instead of a separate `prd/*.json` file. Plan and PRD
+have the same lifecycle, the same `source_plan`, and the same archive path. The slim shows one
+"Plans" button instead of two concepts.
+
+**Cost.** ~3 days. Migration script for existing PRDs + materializer changes + verify flow.
+Risky because plan resume logic depends on reading the PRD JSON twin.
+
+**Risk.** Medium-high. The plan-to-prd diff-aware update logic is non-trivial and would need to
+be re-verified. **Recommendation: defer until at least one round of slim UX shows the unified
+"Plans" button is the right primitive.**
+
+---
+
+## 6. Notes + KB + Pinned: which survives?
+
+**Problem.** Carlos couldn't tell Notes vs KB vs Pinned Context apart. The reality:
+
+- **Notes inbox** = raw, unsorted (one file per agent per task).
+- **KB** = consolidated, classified into 5 categories.
+- **Pinned** = hand-picked subset that gets prepended to every agent prompt.
+- **`notes.md`** = a *separate* consolidated team-decisions blob, also injected into agent
+  prompts.
+
+Four distinct concepts for "the team's memory" is at least one too many.
+
+**Proposal.** Collapse to two concepts the human sees: **"Knowledge"** (everything durable)
+and **"Pinned"** (the always-prepended subset). Hide the inbox/`notes.md` distinction from the
+UI — they're consolidation pipeline internals.
+
+The data layer keeps four stores under the hood; the UI presents two:
+
+| UI label  | Data sources                                                  |
+|-----------|---------------------------------------------------------------|
+| Knowledge | `notes/inbox/*.md` + `notes/archive/**/*.md` + `notes.md`     |
+| Pinned    | `pinned.md`                                                   |
+
+Pinning a Knowledge entry promotes it into `pinned.md`; un-pinning returns it. Inbox vs archived
+becomes a "freshness" badge on the Knowledge entry rather than a separate tab.
+
+**Cost.** ~1 day for the API aggregator + new UI tab. No engine changes.
+
+**Risk.** Low. Pure presentation.
+
+---
+
+## 7. Schedule vs Watch: is one a special case of the other?
+
+**Problem.** Schedule = "fire on a cron pattern." Watch = "fire when a condition flips." They
+share most of the lifecycle (definition, fire history, pause/resume, expire/stopAfter). The
+two state files (`schedule-runs.json` / `watches.json`) and parallel CRUD APIs add maintenance
+without adding real concepts.
+
+**Proposal.** Generalize to **Trigger**, with two `kind`s: `cron` and `event`. Backend stays
+in two files for now (avoid the migration), but the API and UI present them as one.
+
+```jsonc
+{
+  "id": "nightly-tests",
+  "kind": "cron",
+  "cron": "0 2 *",
+  "action": { "type": "work-item", "title": "Nightly tests", "agent": "dallas" }
+}
+```
+
+```jsonc
+{
+  "id": "watch-pr-42",
+  "kind": "event",
+  "target": "PR-42",
+  "condition": "merged",
+  "action": { "type": "notify", "channel": "inbox" }
+}
+```
+
+Slim's "Trigger" button opens one creator that branches on `kind`.
+
+**Cost.** ~2 days. New aggregator endpoint, one creator UI, normalized response shape.
+
+**Risk.** Medium. Watches have a richer condition vocabulary
+(`new-comments`, `status-change`, `vote-change`, `build-fail`, …) than Schedules. The unified
+surface can't paper over that — the creator UI still needs branch-per-kind rendering. So this
+is mostly a mental-model win, not a surface-area-savings win.
+
+---
+
+## 8. "Two unrelated tasks side-by-side on the same project"
+
+**Problem.** Carlos asked how to run two unrelated work streams in parallel on the same project.
+Today the answer is: just queue two work items; the dispatcher will pick two agents (up to
+`engine.maxConcurrent = 5`) and they'll work in separate worktrees. **But** there's no UI
+affordance for separating them — both show up in the same flat dispatch queue. There's no
+"context A vs context B" boundary.
+
+**Proposal.** Add a **"thread" / "stream" tag** to Work Items. Optional string, free-form. The
+slim's Status and History panels group by thread when present. Default thread = `default`.
+Threads aren't in the engine at all — they're a presentation grouping. CC accepts them on
+dispatch.
+
+```jsonc
+{ "title": "fix login bug", "thread": "auth-rewrite" }
+{ "title": "rename buttons", "thread": "ux-cleanup" }
+```
+
+This costs almost nothing technically, and gives the human a way to keep two storylines
+mentally distinct.
+
+**Cost.** ~1 day. Schema field on work items, group-by in queries, slim UX wiring.
+
+**Risk.** Low. Backward compat by treating missing thread as `default`.
+
+---
+
+## 9. Settings sprawl: split fast vs advanced
+
+**Problem.** Carlos said Settings has too many options. The full dashboard's Settings page
+mixes:
+
+- Agents (per-agent CLI/model/skill/budget)
+- Projects (paths, repo hosts, work sources)
+- Engine knobs (timeouts, retries, concurrency)
+- Runtime defaults (default CLI, default model)
+- Feature flags
+- ADO/GitHub auth
+- MCP servers
+- Cache controls / reset
+
+Most users only ever touch feature flags + the slim-ux toggle. Most settings are either
+*infra* (set once at install) or *fleet* (rarely changed).
+
+**Proposal.** Two-tier settings:
+
+- **Slim "Quick settings"**: just feature flags + a button to open the full settings.
+- **Full "Advanced settings"**: everything else. Stays as is in the existing dashboard. The
+  slim deep-links to it.
+
+Round 1 already does this. The next step is to *split* the full dashboard's Settings page into
+"Quick" (the things `/api/settings/reset` fixes) and "Advanced" (the things you tune once).
+
+**Cost.** ~1 day on the full dashboard side. Zero on the slim side (already done).
+
+**Risk.** None.
+
+---
+
+## 10. Recent Completions: less metadata, more context
+
+**Problem.** Carlos said the completions widget shows too much (ID column, agent column, no
+click-through to source). His mental model is: "I want to know what shipped, when, and what
+prompted it."
+
+**Proposal.** Restructure each completion entry as a **prompt → result** pair:
+
+```
+12:04  Carlos: "Fix the dashboard typo on the Plans page"
+       → Dallas: ✅ shipped PR #43 (2 files, 4 lines) · 1m ago
+                 view diff · view PR · view chat thread
+```
+
+The "what prompted it" piece requires the engine to remember the originating CC message — which
+it doesn't today. Either:
+
+1. **Cheap:** When CC dispatches, stash the originating message text in the work item's
+   `description` field (it already does this loosely).
+2. **Right:** Add `_originPrompt` (originating CC turn) and `_originSession` (CC session id) to
+   the work item. Slim's history panel can then render the prompt above the result and link
+   back to the chat.
+
+**Cost.** ~1 day for the cheap version, ~2 days for the right version.
+
+**Risk.** Low. Read-side change to the dispatch creation path.
+
+---
+
+## 11. "Skills + MCPs probably belong off the main screen"
+
+**Problem.** Carlos called these out as "off the main screen." Skills and MCPs are
+power-user/runtime concerns; they pollute the slim's mental model.
+
+**Proposal.** Drop the "Tools" page entirely from slim. Keep them in the full dashboard.
+Surface a single "X skills, Y MCPs available" line in slim's Settings dialog with a deep link.
+
+**Cost.** Zero (already done in Round 1 — slim has no Tools page).
+
+**Risk.** None.
+
+---
+
+## 12. Dashboard build pipeline: extract slim assets
+
+**Problem.** The full dashboard SPA is built once at startup from `dashboard/layout.html` +
+fragments and gzipped. The slim deliberately bypasses this so iteration is hot — read fresh on
+every request. As slim grows past the 1000-line mark, that single-file model will hurt:
+
+- CSS, action-prompt JS, cockpit JS, history JS all in one file.
+- No component reuse between slim panels.
+- Hot-reload works for the file but not for individual sections.
+
+**Proposal.** Extract slim assets into `dashboard/slim/` once it crosses a complexity threshold
+(round 3 or 4):
+
+```
+dashboard/slim/
+  index.html        ← references the assets below
+  cockpit.js
+  history.js
+  actions.js
+  chat.js
+  styles.css
+```
+
+`serveSlimUx` reads `index.html` and inlines or links the others. Hot-reload still works. Each
+file < 300 lines. **Don't do this yet** — the seam is wrong below 1500 lines.
+
+**Cost.** ~1 day.
+
+**Risk.** Low. CSP relaxation already in place; just needs to extend to include
+`/dashboard/slim/*` if assets become cross-origin-y. (They won't for same-origin assets.)
+
+---
+
+## 13. CC actions: the "create a watch from chat" path needs a visual confirm
+
+**Problem.** Round 1 wires the four Action buttons through CC by typing
+"Create a work item: …" into the chat. CC parses, emits `===ACTIONS===`, and the action
+executes. But the user has no easy way to confirm "yes, that's the work item I meant" before
+it's queued. CC's free-text understanding is good but not perfect.
+
+**Proposal.** Add an optional `dryRun: true` flag on CC actions. When set, CC echoes the parsed
+action as a structured preview and *does not* enqueue. The slim modal sets `dryRun: true` by
+default; the user clicks "Confirm" to submit a second turn that drops the flag.
+
+Or, simpler: skip CC entirely for the four buttons, and call `/api/work-items` /
+`/api/plans/create` / `/api/notes` / `/api/schedules` directly with form fields. Round 2 task.
+
+**Cost.** ~1 day for direct API path. ~1.5 days for the dryRun preview flow.
+
+**Risk.** Low for the direct path; medium for dryRun (needs CC system prompt edits + tests).
+**Recommendation: do the direct API path; skip dryRun.**
+
+---
+
+## 14. Charter + Routing should be one page in slim
+
+**Problem.** Charter is "who the agent is." Routing is "what work goes to whom." They're tightly
+coupled (you can't route review to an agent whose charter says "doesn't review") but live in
+totally separate places.
+
+**Proposal.** Add a slim sub-page (off the gear menu) called **"Team"** that shows each agent
+side-by-side with: charter excerpt, routing rules they own, recent completions, current
+dispatch (if any). One screen for "what is the team doing?"
+
+**Cost.** ~1.5 days. Read-only aggregator + simple grid UI.
+
+**Risk.** Low.
+
+---
+
+## 15. The slim-ux feature flag itself: ramp plan
+
+**Problem.** `slim-ux` is currently the only entry point; if it breaks, the user can't get back
+to the full dashboard except via the `?fullDashboard=1` deep link (round-1 added this) or the
+settings dialog. That's fine for early dogfooding, but at GA we'll want a smoother fallback.
+
+**Proposal.** Two feature flags instead of one:
+
+- `slim-ux` — show the slim *option* (renders a "Try slim UX" link from the full dashboard).
+- `slim-ux-default` — slim becomes the root route; full moves to `/full`.
+
+Round 1's behavior corresponds to `slim-ux: true && slim-ux-default: true`. Once slim is GA,
+default flips on; `slim-ux` registry entry gets deleted as redundant.
+
+**Cost.** ~0.5 day.
+
+**Risk.** None. Simple gating.
+
+---
+
+## 16. Surprises found during Phase A
+
+A few things were genuinely surprising while building the concept dictionary. Capturing them so
+the team can decide whether they're features or footguns:
+
+**a. `notes.md` and the KB are *separate* memory stores.** Both feed agent prompts; they don't
+share content. Consolidation either writes to `notes.md` (decisions / pinned-style) or
+classifies into a KB category. There's no "this fact lives in both" path. **Probably fine; just
+worth a docstring somewhere.**
+
+**b. CC sessions are *non-expiring*.** The CC system intentionally never prunes sessions
+(per `CLAUDE.md` line ~XYZ). So a long-lived tab keeps growing the prompt forever. The
+prompt-hash-mismatch invalidator is the only safety net. **Consider a soft "session age" warning
+in the UI when context approaches the cache limit.**
+
+**c. Build-failure cache can go stale.** `_buildStatusStale` is honored by the auto-fix
+dispatcher but not by the cockpit display. Slim's PR tile could show "1 failing build" when the
+build is actually green-after-rerun. **Add `_buildStatusStale` consideration to the cockpit
+display.**
+
+**d. `dispatchCompletionReportPath` writes to `engine/completions/<dispatchId>.json` but the
+oversized-prompt sidecar writes to `engine/contexts/<dispatchId>.json`.** Two directories with
+similar purposes. The dashboard widget for "open completion report" sometimes hits the wrong
+one if you copy-paste a path. **Either merge the directories or make the API explicit about
+which kind of artifact you want.**
+
+**e. `WORK_TYPE` has 13 entries but `routing.md` only routes ~7 of them.** The unrouted types
+(`MEETING`, `EXPLORE`, `ASK`, …) are routed by other code paths. **Either consolidate routing
+or document the gap clearly.**
+
+---
+
+## Suggested order of operations
+
+If we were prioritizing for a Round 2 sprint, this is roughly the order I'd go:
+
+1. **§1** `/api/cockpit` (½ day, immediate slim perf win)
+2. **§13** Direct API calls for the four Action buttons (1 day, removes the TEMP layer)
+3. **§3** `/api/events` + history feed (1.5 days, replaces three client-side merges)
+4. **§6** Knowledge / Pinned UI consolidation (1 day, big mental-model simplification)
+5. **§2** SSE for cockpit (½–1.5 days, removes the polling loop)
+6. **§10** Origin prompt linkage on completions (1 day, makes "what prompted this?" answerable)
+
+Total: ~6 days for a meaningful Round 2 that earns real shipping confidence.
+
+The bigger model-merger questions (§4 Work Items + Pipelines, §5 Plan + PRD, §7 Schedule +
+Watch) should wait until at least Round 3 — we should let the slim UX expose where the
+*model* friction actually is, not assume it from the current UX friction.