Work Summary

Work Summary — Current (v3)

Attention-first project dashboard. The default (lean) mode answers four questions — what's moving, what's stuck, what's next, what's queued — in a bounded view (~45 lines) regardless of workspace size. The exhaustive inventory is available as an explicit mode, not the default: mature workspaces accumulate hundreds of terminal items, and rendering them all buries the actionable content.

Supplement the data with brief observations only when there is a genuine anomaly or actionable insight.

Step 0 — Mode Selection

Step 0.5 — Project Scope Resolution

Before running Lean, Full, or Multi-Project mode, resolve the scope:

• Check session context from the SessionStart hook. A ## Project Scope section carries a project rootId, which scopes reads (ancestorId/anchorId) and anchors. A ## Personal Scope section carries a personal root, which only anchors new items: in personal scope every call below runs unscoped (whole database) and the header says personal scope — whole database.
• Otherwise, read .taskorchestrator/config.yaml's top-level project.rootId field (a file read, not an MCP call).
• If no rootId is found by either path, behavior is exactly as before this feature existed — every call below runs unscoped (whole-DB), and no ancestorId/anchorId param is passed. This is the common case for single-project workspaces.

When a project rootId is known, Lean and Full mode calls pass it as ancestorId="<rootId>" — except the overview calls, which use anchorId="<rootId>" (all noted inline below). Scoped mode is unaffected — it already resolves its own scope UUID per invocation via FTS.

Lean Mode (default)

Data collection (all parallel)

1. query_items(operation="overview", excludeTerminal=true) — non-terminal roots with their childCounts role roll-ups only. includeChildren is deliberately omitted: it is a no-op on the anchored (project-scoped) path — the anchored overview never returns per-child arrays, only childCounts — and unnecessary on the global path, since the step-5 queue-list is the canonical source for individual children. Add anchorId="<rootId>" when a rootId is known, to anchor the overview to this project. Detecting an older server: MCP servers silently ignore unknown parameters (they return 200 with the full unfiltered payload) rather than rejecting them — so a server predating excludeTerminal will not error. Detect the no-op by inspecting the payload: if any returned root has role: "terminal" AND its childCounts show only terminal descendants (no queue/work/review/blocked), the filter was not applied. Note the qualifier: a current server intentionally retains a terminal-role root that still has non-terminal descendants (it is active work parked under a done container), so a terminal root with open descendants is expected, not evidence of an old server. When the no-op is detected, (a) drop client-side only those terminal roots whose childCounts show no open descendants during enrichment, and (b) if the response is also truncated: true, re-issue with limit=<total> (still passing anchorId if known) so non-terminal roots in the tail (e.g. active containers) are not silently dropped — a truncated unfiltered payload can hide real workstreams. Headline counts never come from this payload regardless (see the role-total queries below), so correctness holds either way; this only recovers the workstream/backlog completeness the filter was meant to guarantee.
2. get_context() — health-check: active items, blocked items, stalled items, claim summary. Add ancestorId="<rootId>" when known.
3. get_context(mode="session-resume", since="<now minus 48h, ISO 8601>") — recent role transitions (the "where did I leave off" signal). Add ancestorId="<rootId>" when known — but note the known limitation: recentTransitions is NOT scoped by ancestorId even when passed; it always reflects the whole tree, not just this project.
4. get_next_item(limit=5, includeDetails=true) — ranked recommendations with parentId/tags. Add ancestorId="<rootId>" when known.
5. query_items(operation="search", role="queue", limit=100) — the canonical source for queue children. Group its rows by parentId to get each workstream's "next up" (the queue children the anchored overview omits); each row carries parentId, depth, priority, and type/tags. total gives the true queued count. Add ancestorId="<rootId>" when known. Exclude shelves from the headline: subtract rows whose type or tag is container or project — a freshly-created container/anchor sits in queue role until its first advance and would otherwise inflate N queued. If total exceeds the returned page (queue backlog >100), note the headline count is approximate.
6. query_items(operation="search", role="terminal", limit=1) — read total for the true terminal count. Add ancestorId="<rootId>" when known.

Headline counts must come from these sources, never from eyeballing the overview payload — overview covers the roots set with childCounts roll-ups only and may be truncated.

Enrichment

• Containers are shelves, not work. Items with tag or type = container never count as active; exclude them from the active/In Flight lists.
• Active = health-check activeItems minus containers. Blocked/stalled = health-check lists as-is.
• Workstreams = non-terminal roots with children. Progress fraction = terminal / total from childCounts, rendered N/M terminal (not done — the terminal bucket includes cancelled items, so labeling it "done" overstates completion); "next up" = the highest-priority queue row from the step-5 queue-list whose parentId is that root.
• Standalone queue roots group by kind: the type field if set; else the first tag not in {agent-observation, feature, container}; else "Uncategorized".
• Rollup threshold: any group with more than 5 items renders as a count line broken down by kind, listing individual items only when priority ≥ medium or tagged action-item.
• Hygiene candidates: empty non-terminal containers; evident test artifacts (probe/smoke/depth-test titles, mtest-* tags).

Dashboard format

Render in this order; omit any section with no data.

## ◆ Work Summary

[One sentence: health and momentum.]

**N active · N blocked · N stalled · N queued · N terminal**

### ◉ In Flight
| ID | Item | Container | Status | Pri |
|----|------|-----------|--------|-----|
| `xxxxxxxx` | <title> | <parent title or —> | ◉ work | high |

↳ Recent: <up to 3 one-line transition summaries from session-resume, newest first — omit line if none in 48h>

### ⊘ Attention Required
| ID | Item | Container | Issue |
|----|------|-----------|-------|
| `xxxxxxxx` | <title> | <parent or —> | Blocked by: `<blocker-id>` <blocker-title> |
| `yyyyyyyy` | <title> | <parent or —> | Stalled: missing `<note-key>` (trait: <trait-name> if applicable) |

### ▸ Recommended Next
| ID | Title | Container | Pri |
|----|-------|-----------|-----|

### ○ Backlog
**Workstreams**
| ID | Workstream | Progress | Next up | Pri |
|----|-----------|----------|---------|-----|
| `xxxxxxxx` | <root title> | 2/5 terminal | <queue child title> | high |

**<Kind>** (N): `id` <title> (pri) · `id` <title> (pri) · ...        ← groups of ≤5
**<Kind>** (N): X bug · Y optimization · Z friction — highest: `id` <title> (pri)   ← groups of >5

### ✓ Terminal — N terminal items

↳ Hygiene: <N> test artifacts, <N> empty containers — candidates for /batch-complete

Lean-mode rules:

• Total output target is ~45 lines; rollups are how you hold that budget.
• The Terminal section is a single line (count from the role-total query). No titles, no done/cancelled split in lean mode — the terminal statusLabels aren't fetched. Use full mode for the breakdown.
• The Hygiene line appears only when candidates exist.
• If all of In Flight / Attention / Recommended Next are empty, say so in the health sentence ("Quiet board — nothing in flight...") rather than rendering empty sections.

Full Inventory Mode ($ARGUMENTS = full)

Everything lean mode shows, plus the complete per-container inventory.

Data collection

1. query_items(operation="overview", limit=<total>) — the compact roots list, without includeChildren. Add anchorId="<rootId>" when a rootId is known, to anchor the roots list to this project. It carries each root's role, statusLabel, childCounts, tags, and type but pulls no subtree, so it stays under the tool-result token ceiling even on mature workspaces. If the first call reports truncated: true, re-issue once with limit=<total>.
2. Per active container (a non-terminal root whose childCounts show non-terminal children), in parallel: query_items(operation="overview", itemId="<root-id>", excludeTerminal=true) — returns that container plus its open children, plus any terminal-role child that itself still has non-terminal descendants (scoped excludeTerminal filters the children array by role but retains done-labeled sub-containers holding open work; childCounts stays full). Each call is bounded to one container's open work. These are already itemId-scoped and need no ancestorId.
3. get_context() — health-check. Add ancestorId="<rootId>" when known.
4. get_next_item(limit=5, includeDetails=true). Add ancestorId="<rootId>" when known.
5. Role-total queries as in lean mode: the terminal count via limit=1; the queued headline via the role="queue" list with container/project exclusion (lean step 5). Same ancestorId rule.

Never fetch every root's children in one call (overview includeChildren=true across the whole tree). On a mature workspace the combined terminal subtrees exceed the tool-result token ceiling — the per-container scoped fetch in step 2 is what keeps full mode bounded. Terminal children are counted, not enumerated: the ✓ N terminal footer comes from childCounts, never from pulling the terminal subtree. To see a specific container's completed items, run Scoped mode on that container (one bounded subtree).