Consolidated Learnings Dashboard

Consolidated Learnings Dashboard

Master's window into Dream Mode's output. Browse every consolidated trajectory, see which devices contributed, track adoption, and manually trigger dream passes.

Accessible via Governance → Fleet Learning on any master-mode install.

---

What you see

Top — status card

• Last dream pass: version number + timestamp + elapsed time
• Total consolidated entries: count of consolidated_learnings rows
• Active version: what slaves are currently applying
• Per-stage failures from the last pass (if any)

Middle — incoming activity

• Per-device chart: rolling 30-day learnings-received chart, stacked by device
• Last push from each device: timestamp + size + ingest-reason (if rejected)

Bottom — consolidated entries browser

Main event. Searchable, filterable, paginated list of every consolidated trajectory.

---

Browsing entries

Each entry shows:

Column	What
Task shape	The distilled one-liner (from insight.taskShape)
Preferred / avoidance path	The distilled insight
Success score	Max across contributing devices
Fleet-wide usage	Sum of local usage counts
Contributing devices	Count + optional hover for device IDs
Adoption	How many times slaves have used this entry (and succeeded)
Workspace	Workspace scope (NULL = fleet-wide)
Failure pattern	Badge if this is an avoidance rule
Version introduced	Which dream pass first consolidated this

Filters

• Category — trajectory / memory summary / template patch
• Workspace — pick a specific workspace or "all"
• Success pattern / Failure pattern
• Date range — by first consolidated version
• Adoption — show only entries with > N adoption count

Search

Free-text search over taskDescription, taskShape, and preferredPath.

Sort

By any column. Default is rank (highest score × usage × adoption).

---

Drilling into an entry

Click any row → full detail panel:

• Full insight JSON — {taskShape, preferredPath, avoidancePath, triggerCondition}
• Source trajectories — the top 3 raw trajectories that contributed (anonymized, post-redaction)
• Contributing device breakdown — per-device usage + success count
• Adoption history — how many slaves have used this entry via retrieval, and their turn outcomes
• Timeline — which dream pass introduced it, any updates, adoption trend

---

Actions

Delete an entry

[Entry] → Delete → confirm admin re-auth

Removes from consolidated_learnings + propagates on next config bundle pull (slaves delete matching rows locally).

Use cases:

• Entry contains bad advice (LLM distillation produced misleading insight)
• Sensitive pattern leaked through redaction
• Pattern is obsolete (technology changed)

Override workspace scope

[Entry] → Change workspace scope → pick

Useful for promoting a workspace-specific pattern to fleet-wide (set workspace to NULL), or restricting a pattern to a specific tenant.

Export

[Entry] → Export JSON or Export all → CSV.

Useful for incident reports or offline analysis.

---

Run Dream Now

Near the top of the dashboard, a "Run Dream Now" button triggers a manual pass outside the normal schedule.

• Admin re-auth required (same as destructive commands)
• Pass runs immediately, returns a DreamPassResult after completion
• Useful for:
  • Testing a fresh config
  • Catching up after an outage that skipped the nightly slot
  • Rushing high-frequency clusters to consolidation before they pile up

The button is disabled while a pass is already running.

---

Template patches browser

Toggle the view to Template Patches:

• Per-template entries (Backend Engineer, QA Engineer, etc.)
• fleet_instructions_md preview — what gets appended to the template's persona
• Cluster count that contributed
• Max rank score
• Rollback button — remove the patch, templates go back to unpatched state

Useful for verifying Dream Mode isn't injecting bad rules into your templates (rare but possible if the distillation misfires).

---

Memory summaries browser

Toggle to Memory Summaries:

• Per-slot aggregations (grouped by agentSlotHash)
• Contributing device count
• Top N summaries per slot (default 5)

These are the domain-knowledge slices that slaves will receive for their equivalent agent slots.

---

Health indicators

Traffic-light status at the top:

• 🟢 Healthy — last pass ≤ 48h old, zero stage failures, consolidated count > 0
• 🟡 Warning — last pass > 48h old OR ≥1 non-critical stage failure
• 🔴 Attention needed — pass failed 3× retry, or critical stage (write/rank) failed

Click the indicator for actionable diagnostics.

---

Cost + observability

Each dream pass logs its LLM cost in activity_log:

• Input tokens consumed by distillation
• Output tokens generated
• Cost estimate in cents

Dashboard shows rolling 30d cost of dream passes. Useful for validating you're within the cap and for budget forecasting as the fleet scales.

---

Audit

Every user interaction with the dashboard is audited:

• dream.dashboard.entry_deleted — with entry ID + reason
• dream.dashboard.scope_changed — with from/to workspace
• dream.dashboard.manual_pass_triggered — with user ID
• dream.dashboard.patch_rolled_back — with template ID

Chain + signed. See Audit Logging.

---

Common workflows

Debugging "my agents aren't getting any hints"

1. Slave's Reasoning Bank panel: is fleet-consolidated count > 0?
   • If 0: slave hasn't pulled a bundle with learnings. Check slave's config version.
   • If > 0: retrieval is filtering them out. Check workspace scope matches.
2. Master's consolidated entries: do entries exist for workspaces the slave's agents use?
3. If workspaces don't match: adjust scope or run tasks in the matching workspace to build signal there.

Auditing a questionable pattern

1. Search the dashboard for the task shape
2. Drill into the entry
3. Inspect source trajectories — are they from devices you expected?
4. Check adoption — if adoption is low, low-risk. If adoption is high, possibly propagated bad advice.
5. If bad: delete entry, then consider a cache.clear command to slaves that have adopted it.

Rolling back a template patch

1. Dashboard → Template Patches view
2. Find the patch
3. Click Rollback
4. Next config bundle pull: slaves clear fleet_instructions_md for that template
5. Next agent spawn: agents get the base instructionsMd only

---

Related pages

• Dream Mode Overview — the big picture
• Dream Pass Internals — the distillation pipeline detail
• Enabling Dream Mode — turning it on
• Privacy and Redaction — what's protected
• Audit Logging — every dashboard action is audited