lp-do-ideas Trial Orchestrator

/lp-do-ideas generates actionable dispatch packets from standing-artifact deltas. It runs in mode: trial under Option B (queue-with-confirmation). Dispatches are enqueued and routed, but no downstream skill is auto-invoked in trial mode.

Operating Mode

TRIAL ONLY

This skill and its backing runtime (lp-do-ideas-trial.ts) operate exclusively in mode: trial. Invoking with mode: live is rejected fail-closed. Live-mode activation criteria are defined in docs/business-os/startup-loop/ideas/lp-do-ideas-go-live-seam.md.

Invocation

/lp-do-ideas [<artifact-path-or-id>] [--business <BIZ>] [--before <sha>] [--after <sha>]

Or provide a ArtifactDeltaEvent JSON payload directly. Or invoke with no arguments to use the intake path.

Intake Path (No Arguments or Free Text)

When invoked with no arguments, or with a plain description rather than a structured payload, run the intake flow.

Step 1 — Gather the idea

Ask:

What's the new information or idea? Describe it in plain language.

Step 2 — Determine trigger type

From the description, determine which of two trigger types applies:

artifact_delta — the user describes a change that has already happened to a specific standing document. The artifact reflects the new truth now.

operator_idea — the user describes new information, a signal, or an idea that has not yet been written into any standing artifact. The artifact update is a downstream output of the work this dispatch will trigger — not a precondition.

If unclear, ask one question: "Has this already been written into a doc, or is this something new you want to act on?"

Step 3 — Gather minimum context

For artifact delta:

• Which business?
• Which artifact changed? (show list from standing registry if needed)
• What changed? — user describes in plain language; agent infers changed_sections from description if SHAs are unavailable

For operator idea:

• Load and follow modules/operator-idea-structured-intake.md.
• Run the five core questions in order. Do not rely on broad free-form inference for area_anchor, location_anchors, domain, or evidence fields once the structured intake starts.
• If the operator reports multiple distinct gaps, split them first and produce one intake block per gap before routing.
• Use the module's deterministic assembly rules to populate:
  • area_anchor
  • location_anchors
  • domain
  • provisional_deliverable_family
  • evidence_refs
  • current_truth
  • next_scope_now
  • why
  • intended_outcome
• For likely code-bearing work (code-change, multi, infra, design), also load ../_shared/engineering-coverage-matrix.md and capture obvious coverage hints using the module's coverage-hint: evidence-ref convention. Do not turn this into a full planning interview at ideas stage.
• Routing remains a Step 4 judgment call. Do not ask the operator to decide the final route unless Step 4 is still genuinely ambiguous after the structured intake.

Step 4 — Apply routing intelligence, emit, and enqueue

Decomposition rule — one event, multiple narrow packets: When one incoming event contains multiple distinct gaps, emit one dispatch packet per gap. Do NOT produce one aggregate packet covering the entire event.

Each packet must be independently actionable — routable to a separate fact-find, direct build, or briefing without depending on the others.

Example: A PWRB strategy backfill touching 4 pending items should produce 4 packets:

• "PWRB IPEI agreement — document empty, needs drafting"
• "PWRB hardware SKU — no decision, needs supplier research"
• "PWRB venue shortlist — no selection made from IPEI customer base"
• "PWRB brand name — 'PWRB' is a code not a brand, ASSESSMENT-10 pending"

Apply routing intelligence (see below) to determine status and recommended_route. Emit a schema-valid dispatch packet and enqueue it in queue-state.json.

For operator idea packets:

• trigger: "operator_idea"
• artifact_id, before_sha, after_sha — omit
• location_anchors — must be populated from the structured intake; do not leave empty for likely fact_find_ready or micro_build_ready packets
• provisional_deliverable_family — populate from the structured intake enum
• evidence_refs — include operator-stated rationale using the format: "operator-stated: <one-line summary>"
• why and intended_outcome — preserve operator-authored values when explicitly stated; otherwise use the module's route-neutral auto fallback
• All other required fields apply as normal

Queue-with-confirmation policy (trial mode):

• fact_find_ready → enqueue (queue_state: "enqueued") and present summary to operator. Invoke /lp-do-fact-find only after explicit confirmation.
• micro_build_ready → enqueue (queue_state: "enqueued") and present summary to operator. Invoke /lp-do-build only after explicit confirmation.
• briefing_ready → enqueue (queue_state: "enqueued") and present summary to operator. Invoke /lp-do-briefing only after explicit confirmation.
• logged_no_action → record and report. No downstream invocation.
• auto_executed is reserved and must not be set in trial mode.

Required Inputs (Structured Invocation)

For structured / programmatic invocation with an artifact delta:

Optional but needed for classification:

Cutover Phase Behavior (Source-Trigger Migration)

Runtime admission behavior is phase-aware:

Safety rules:

• Unknown artifacts never auto-admit.
• Projection/read-model artifacts (projection_summary) do not auto-admit.
• trigger_policy: never cannot be bypassed by manual override.

Shared Workflow Telemetry

lp-do-ideas already owns queue admission and cycle telemetry in docs/business-os/startup-loop/ideas/trial/telemetry.jsonl.

DO stages (lp-do-ideas, lp-do-fact-find, lp-do-analysis, lp-do-plan, lp-do-build) append record_type: "workflow_step" lines to the same stream using:

pnpm --filter scripts startup-loop:lp-do-ideas-record-workflow-telemetry -- --stage <stage> --feature-slug <slug> ...

Use lp-do-ideas stage recording when a feature is first handed off into the DO chain to establish the feature baseline for later token-delta measurement.

Codex token usage is auto-captured from session metadata when CODEX_THREAD_ID is available. Claude token usage is auto-captured via project session logs (sessions-index.json → debug/latest fallback). Explicit --claude-session-id still takes priority when supplied.

This keeps ideas-stage telemetry and downstream hot-path telemetry in one append-only stream while allowing the existing queue rollups to ignore downstream records safely.

Progressive Handoff Packets

Once a dispatch enters the DO chain, downstream stages should use the canonical bounded handoff sidecars defined in docs/business-os/startup-loop/contracts/do-stage-handoff-packet-contract.md.

Rule:

• lp-do-fact-find emits fact-find.packet.json,
• lp-do-analysis consumes fact-find.packet.json first and emits analysis.packet.json,
• lp-do-plan consumes analysis.packet.json first and emits plan.packet.json,
• lp-do-build consumes plan.packet.json first before escalating to the full upstream markdown artifact.

Routing Intelligence

Any delta with changed_sections present is assessed by the agent. The agent reads the changed section content and exercises judgment to determine whether the change warrants planning investigation, understanding only, or no action.

No hard keyword lists for operator-idea routing. The agent must answer these questions using judgment. (Note: artifact-delta routing in the TS orchestrator uses the T1_SEMANTIC_KEYWORDS list plus a conservative direct-build heuristic in lp-do-ideas-trial.ts — those rules apply only to artifact_delta events, not to operator ideas handled here.)

1. Is the change material? Substantive edit, or a typo/formatting fix?
2. Does it open a planning gap? Does what's documented now differ from what's built, planned, or actively running in a way that requires investigation and tasks?
3. Is the area already covered? Check docs/plans/*/fact-find.md and docs/plans/*/plan.md for any active or in-flight work covering the same area. If yes, a duplicate fact-find is wasteful — prefer briefing_ready to review what's already in-flight.

Routing decision:

• fact_find_ready → change is material AND creates a planning gap AND no in-flight fact-find or active plan already covers the area. Route to lp-do-fact-find. This applies to any domain that affects strategy, offer, distribution, product, measurement, brand/visual direction, or supply/logistics — not just ICP/pricing/channel.
• micro_build_ready → change is material, directly executable, and trivially bounded: one surface, no architectural choice, no external research, no meaningful planning branch, and a clear validation path already exists. Route to lp-do-build.
• briefing_ready → change is material but the gap is informational rather than a planning gap (existing behaviour, context, background), OR in-flight work already covers the area and a briefing to review it is a better fit.
• logged_no_action → change is not material (formatting, typo, minor clarification), OR before_sha is null (first registration), OR the event describes an administrative startup-loop action rather than a knowledge or planning gap (see Admin non-idea suppression below).

No changed_sections or missing before_sha → logged_no_action (no dispatch emitted).

Admin non-idea suppression

Events that describe a startup-loop administrative action — not a gap in knowledge, strategy, or planning — must route logged_no_action. The test is:

"Does this event describe something the operator does, or something the operator needs to know or decide?"

If it describes an action (register, advance, complete), it is logged_no_action. If it describes a gap or uncertainty that requires investigation or a plan, it is fact_find_ready or briefing_ready.

Suppression examples:

• "PWRB startup loop not yet formally started" → logged_no_action. Redirect: run /startup-loop start --business PWRB.
• "Startup loop advanced to stage S4" → logged_no_action. No planning gap opened by the advancement itself.
• "Results review completed with no new findings" → logged_no_action.

Edge case — admin action that opens a planning gap: If a completed action reveals a gap (e.g. stage S3 complete → brand profiling now needed), suppress the admin action itself as logged_no_action and submit the revealed gap as a separate operator-idea dispatch with its own narrow area_anchor.

Evidence Fields for Classification

When gathering context in Step 3 for operator_idea, do not rely on passive "listen for signals" behavior alone. Run the conditional evidence prompts from modules/operator-idea-structured-intake.md after the five core questions. These fields are used downstream to determine how urgently an idea is acted on and which tier it lands in. They are advisory — if the operator does not have them, proceed without forcing a guess; the downstream classifier will apply automatic demotion to surface the gap.

Capture evidence fields during Step 3 and include any provided values in the dispatch packet under evidence_refs or as structured fields alongside operator-stated rationale.

Example: If the operator says "we have a live payment failure", ask for the incident reference and whether there is a deadline for resolution. Those two data points (incident_id, deadline_date) are enough to unlock the highest urgency tier.

If none of these signals are present in the operator's description, omit the evidence fields entirely and route based on the description alone.

Plain Language Writing Standard

All operator-facing fields (area_anchor, why) MUST be written for a business decision-maker, not a developer.

The operator reading these ideas may have no technical knowledge. They are making a business decision: approve, defer, or decline. The writing must help them make that decision confidently.

Non-negotiable rules

1. No technical jargon. No function names, file paths, schema names, API names, or internal tool names in area_anchor or why. If a technical concept must be referenced, describe what it does for users.

2. 14-year-old reader test. Before writing area_anchor or why, ask: "Could a smart 14-year-old read this and understand why the business should care?" If not, rewrite it.

3. Business consequence first. Every why must state what goes wrong for staff, customers, or the business — not which component is responsible.

4. Benefits, not mechanics. Describe what gets better when this is fixed. "Fixing this means every guest gets exactly one reply" is good. "Eliminates dedup gap at tool boundary" is not.

Self-check before writing

Ask yourself:

• Does area_anchor describe the business impact, or just the technical location of the problem?
• Does why explain what a real person experiences when this is broken?
• Could someone approve or decline this without knowing anything about the codebase?

If the answer to any of these is no, rewrite before emitting.

Worked contrast

Outputs

Each processed delta should produce a dispatch.v2 packet conforming to docs/business-os/startup-loop/ideas/schemas/lp-do-ideas-dispatch.v2.schema.json. dispatch.v1 remains compatibility-only for legacy packets (schema at docs/business-os/startup-loop/ideas/_deprecated/lp-do-ideas-dispatch.schema.json).

Key output fields:

Results are enqueued in docs/business-os/startup-loop/ideas/trial/queue-state.json by the queue layer (TASK-05: lp-do-ideas-trial-queue.ts).

Canonical queue lifecycle values are:

• enqueued
• processed
• skipped
• error

Legacy states in historical queue snapshots (auto_executed, completed, logged_no_action) are compatibility data and must not be emitted as new queue lifecycle values.

Guard: auto_executed is a reserved state and must never be hand-set in trial mode (Option B). Use completed, processed, skipped, or error only.

Idempotency

Duplicate events (same artifact_id + before_sha + after_sha) are suppressed. The seenDedupeKeys set is maintained across calls to the queue layer.

Failure Handling

Contract References

• Dispatch schema (primary): docs/business-os/startup-loop/ideas/schemas/lp-do-ideas-dispatch.v2.schema.json
• Dispatch schema (compat): docs/business-os/startup-loop/ideas/_deprecated/lp-do-ideas-dispatch.schema.json
• Standing registry: docs/business-os/startup-loop/ideas/schemas/lp-do-ideas-standing-registry.schema.json
• Trial contract: docs/business-os/startup-loop/ideas/lp-do-ideas-trial-contract.md
• Policy decision: docs/plans/lp-do-ideas-startup-loop-integration/artifacts/trial-policy-decision.md
• Routing adapter: scripts/src/startup-loop/ideas/lp-do-ideas-routing-adapter.ts (TASK-04)
• Queue + telemetry: scripts/src/startup-loop/ideas/lp-do-ideas-trial-queue.ts (TASK-05)

Known Issues

queue-state.json format divergence

The live queue file (docs/business-os/startup-loop/ideas/trial/queue-state.json) uses a hand-authored format that differs from what the TypeScript persistence layer writes:

The TS persistence layer (persistOrchestratorResult) has never been used to write to the live queue file. All current dispatches in the file are agent-authored directly.

The ideas.user.html viewer handles both formats via a conditional branch (else if (raw && Array.isArray(raw.dispatches))).

Do not attempt to migrate the live queue file to the TS format without a dedicated plan. The divergence is stable and intentional for this trial phase. The TS persistence infrastructure is ready for live-mode activation when the trial escalation criteria are met (trial contract Section 8).