improvisation_digest

Weekly advisory digest scanning VP records for fleet/state claims missing the VP-Sources footer.

improvisation_digest

Scan a rolling time window of VP tasks, messages, and memories for records that carry durable-artifact fleet/state tokens (commit SHA, PR number, VP document ID, or decisive verb such as merged, deployed, approved) but have no VP-Sources footer. This is the Eta heuristic proxy for an orchestrator having made a fleet-state claim without a prior recall upstream.

ADVISORY-only — pure read query. improvisation_digest never blocks any action. Results are informational: a high improvisation rate indicates a team should increase VP-Sources citation hygiene, but the tool itself takes no automated action and has no side effects.

Args

Arg	Type	Default	Description
`windowDays`	number	`7`	Number of days to look back.
`orchestrators`	string[]	—	Scope to these orchestrator roles only (e.g. `["sigma","pi"]`). Omit to scan all orchestrators.

Returns

{
  countsByOrch: Record<string, number>,      // hit count per orchestrator
  countsByCategory: Record<string, number>,  // hit count per record type: "task" | "message" | "memory"
  samples: Array<{                           // up to 50 representative snippets
    id: string,
    category: string,
    orchestrator: string,
    snippet: string
  }>
}

countsByOrch and countsByCategory are both zero-initialized for all observed orchestrators/categories — entries with zero hits are omitted from the returned map. samples is sorted newest-first and capped at 50 entries.

Examples

Default 7-day window across all orchestrators

// call
{ "windowDays": 7 }

// response (illustrative)
{
  "countsByOrch": { "sigma": 3, "pi": 1 },
  "countsByCategory": { "task": 2, "message": 2 },
  "samples": [
    {
      "id": "k17abc...",
      "category": "task",
      "orchestrator": "sigma",
      "snippet": "completionNote: merged PR #954 into main — VP-MCP top level..."
    },
    {
      "id": "k97def...",
      "category": "message",
      "orchestrator": "pi",
      "snippet": "deployed vantage-peers-mcp@2.13.0 to Railway at commit ef91f6f"
    }
  ]
}

Scoped to a single orchestrator

// call — audit sigma's last 14 days
{ "windowDays": 14, "orchestrators": ["sigma"] }

// response — only sigma records are evaluated
{
  "countsByOrch": { "sigma": 5 },
  "countsByCategory": { "task": 3, "memory": 2 },
  "samples": [
    {
      "id": "k17xxx...",
      "category": "memory",
      "orchestrator": "sigma",
      "snippet": "approved PR-C — list_repo_mappings envelope safety shipped at 4ddca2b"
    }
  ]
}

Detection heuristic (Eta A5 scope filter)

A record is flagged when both conditions hold simultaneously:

Condition	Check
Durable-artifact token present	Body contains at least one of: 7–40 hex commit SHA; `#NNN` PR/issue ref; Convex document ID (`k1…` or `j…` prefix); decisive verb (`merged`, `deployed`, `approved`, `shipped`, `released`, `fixed`).
VP-Sources footer absent	Body does NOT contain the `VP-Sources:` substring.

A5 scope exclusions — the following are never flagged regardless of content:

Records authored by system
Records where createdBy matches /^cron-/i (dash mandatory)
Records originating from webhook ingestion paths

The A5 exclusions prevent false positives from automated infrastructure records that legitimately reference SHAs or PR numbers without a VP-Sources footer obligation.

V1 scope and V2 roadmap

V1 (current) scans VP records only — tasks, messages, and memories stored in VantagePeers (Option C). Per Pi Day-113 arbitration (msg k97a0pp6kq1axkj6cmc4pecpy989ce1w), the fallback if V1 misses too many improvisations is Option B — a new dedicated sessions Convex table — not Option A (transcript-replay from JSONL conversation logs).

V1 was selected because VP records are already structured, queryable, and authorship-attributed. Monitor countsByOrch trends over 2–4 weeks; if V1 coverage proves insufficient (because agents do not store all fleet-state claims as VP records), the next iteration introduces a dedicated sessions table (Option B) where the digest can pull richer per-session context without the transcript ingestion pipeline complexity of Option A.

Cross-reference

VP-Sources answer-footer doctrine — full reference including worked examples, none-needed acceptable cases, and advisory-only rationale.
Convex query: improvisationDigest:scanWindow
Mission: k571gcctka8mq5jbkgpj0a0b2n892ctg (VP-MCP top level Bloc A, PR-I)
T-RED cd6cda3 · T-GREEN b9414dc

improvisation_digest

On this page