Run Growth Loop

Run Growth Loop

Coordinate state and handoffs. Do not replace the specialist judgment contained in the other eleven skills.

Inputs

Require:

• workspace or artifact directory;
• user objective;
• operating mode: learning/preview or apply;
• selected policy set;
• current run-state.json, if one exists;
• approval and blocker context.

Default to learning/preview when the mode is omitted, the request is educational, or the authority to modify files is unclear.

Operating modes

Learning / preview

Use this mode on the first turn. Read and validate available artifacts without creating, editing, appending, sending, publishing, or deploying anything. Return these six labeled parts in plain language:

1. inputs read;
2. model interpretation;
3. deterministically verified facts;
4. unverified or blocked facts;
5. decisions that require a person;
6. proposed file changes and exactly one next skill with its reason.

If the deterministic CLI cannot run, keep working from readable artifacts but label deterministic validation not verified. Do not require a learner to install Python merely to receive the preview.

Apply

Enter apply only after the person reviews the preview and explicitly confirms the named files and authority in a later user turn. A prior generic approval, an APR--looking model string, or the existence of an output file is not that confirmation. External writes additionally require a valid scoped human record in approvals.jsonl.

Workflow

1. Set the operating mode and default to learning/preview.
2. Validate repository and available artifacts when the deterministic runtime is available.
3. Read the run objective, completed skills, approvals, and blockers.
4. Reconcile stated completion with files and reference integrity.
5. Identify the first missing or invalid gate.
6. Route to exactly one specialist skill unless independent work is explicitly requested.
7. Preserve partial outputs and explain why the run cannot advance.
8. Ask for human approval at decision and external-action transitions.
9. In apply mode only, update run-state.json as a new auditable state.
10. In apply mode only, create a handoff that distinguishes completed, locally validated, externally executed, and outcome-recorded work.

A schema-valid primary file is not a completed skill by itself. Require every artifact named in that skill's output contract. In particular, evidence with strength=awaiting_human_tag is a valid intermediate record but blocks signals, decisions, and outcomes until a person records both strength and approved_by.

Routing graph

first-five / direct-seeding objective, no prior evidence or metric
  → design-first-user-loop directly (states a working hypothesis and proposes
    candidate metrics inline; evidence and a decision record can follow)

research objective
  → plan-customer-reach → run-switch-interview → synthesize-interviews

configured CS source
  → connect-customer-channels → triage-customer-signals

validated evidence or signal already on hand
  → define-growth-metrics → record-growth-decision → design-first-user-loop
  (a reasonable default when evidence exists, not a precondition for the skill)

completed first-five direct seeding + introduction/referral objective
  → define-growth-metrics (introduction recipe) → record-growth-decision
  → design-first-user-loop (only after a new bounded decision)

approved decision needing build review
  → export-hplan → hplan gates

optional content branch
  → audit-answer-visibility → draft-evidence-content
  → design-first-user-loop → draft-evidence-content
  → draft-evidence-content → osmu-fanout

optional release branch
  → record-growth-decision → announce-release-to-customers

Route from the stated objective and valid available artifacts, not from a mandatory universal sequence. Keep the manual signal path when no connector artifact exists. Route to connect-customer-channels when any connector artifact exists but the required connection, event, or state artifact is incomplete. Keep visibility and content work optional unless the objective asks for them.

When the objective asks for a product introduction page or other answer-first content, route to draft-evidence-content only after design-first-user-loop is complete, so the page describes a loop that was actually validated. While that loop is incomplete, route to design-first-user-loop first and say why.

When the objective asks for an introduction or referral loop, route to design-first-user-loop until the first-five direct-seeding output is complete. Then route to define-growth-metrics to define qualified introduction, referred first value, reuse, and the activation-based loop coefficient. Do not route from that request to a send, reward, or new batch; record-growth-decision must capture the HOLD/resume and reward choice before any later bounded iteration.

When the objective asks to reuse existing content on other surfaces, route to osmu-fanout only after draft-evidence-content is complete and a content-brief.json validates against contracts/content-brief.schema.json. The brief is the shared input both surfaces read; without it each surface would re-decide the topic on its own.

When the objective asks for a release note or customer-facing change announcement, route to announce-release-to-customers only after record-growth-decision is complete, so each shipped change can be described against the decision log rather than against a model's reading of the code. That skill drafts only; announcing to customers stays an external write.

Boundaries

• Let the model interpret blockers and propose the next specialist.
• Use deterministic checks for artifact existence, schema, references, state transitions, and approval.
• Require a person to approve important decisions and every external write.
• Keep learning/preview read-only and wait for a later user turn before apply.
• Do not recreate interview, metric, content, or channel analysis inside the router.
• Do not call provider APIs directly or treat an unconfigured connector as a blocker for manual signal input.
• Do not mark a run complete because files exist; required validation and approval must also pass.

Outputs

In learning/preview, return only the six-part preview and do not write files.

In apply, create or update:

• run-state.json
• next-action.md
• blocked-items.md
• handoff.md

Read references/output-contract.md before writing them.

Stop conditions

Stop when a required artifact is invalid, an approval is missing, policies conflict, restricted data cannot be handled safely, or the next action would expand authority.

Verification

Run:

python3 scripts/stg.py validate-artifacts artifacts/
python3 scripts/stg.py validate-connectors artifacts/
python3 scripts/stg.py next-step artifacts/ --objective "현재 사용자 목표"

Treat the CLI result as routing evidence. The user still owns prioritization and approval.