Agent operating contract

Answer first

What agents may do

Read approved context. Propose bounded signals, decisions, and work orders. Execute only inside an explicitly approved scope.

Completed work requires a durable receipt — run ID, actor, scope, inputs, outputs, gate results, artifact links, side-effect status, and terminal state. A plan, tool call, or model response is not execution.

Core command test

Before editing any product surface, read the current operating file and pass this check.

Cursor, read Launchfiles before editing.

An agent passes only when it can identify:

Current approved decision
Exact work order
Allowed scope
Forbidden actions
Approval boundary
Required proof
Latest receipt or explicit not_run
Unresolved gates
Launchfiles vs Content OS boundary

Scope model

Three scopes govern every agent interaction. Proposals and plans never cross into execution without a separate approval and receipt.

Launchfiles agent scope model
Scope	Agent may	Agent may not	Required proof	Blocked behavior
`read`	Inspect approved context, doctrine refs, decisions, work orders, receipts, gates, readouts, and approved memory.	Change state, claim execution, or treat proposals as approvals.	Source mode label, object IDs, and approval boundary remain visible.	Inferring execution from plans, tool calls, or model output.
`propose`	Draft signals, decisions, work orders, or memory updates for founder review.	Treat a proposal as approved truth, expand scope, or publish.	Proposal labeled proposed; supersession and cooldown when applicable.	Auto-approving or mutating production systems.
`execute_with_approval`	Perform the exact approved action inside the granted scope and return proof.	Expand scope, publish, mutate unrelated systems, or skip the receipt.	Durable run ledger with terminal state and side-effect truth.	Unrestricted autonomous execution.

Object rendering law

Every launchfile object must expose exact state in the interface. Missing proof stays missing; proposals are not approvals.

decision

Exact choice, authority, rationale/why now, status, and proof.

work_order

Owner, scope, approval boundary, forbidden actions, status/blocker, required proof.

run_ledger

Did execution happen, terminal result, side effects, receipt/artifacts.

gate_result

Gate ID, outcome, severity, reason, repair/proof.

approval

Approver, exact action/scope, artifact hash, time/expiry.

readout

Window, source, observed result or explicit missing result, recommendation.

memory_update

Proposed lesson, sources, confidence, approval state/supersession.

Receipt requirement

No agent execution is complete without a durable receipt adjacent to the claim it supports.

run_id: Durable workflow or agent run identifier.
actor: Human, agent, or workflow actor that initiated the run.
scope: Exact approved product, paths, and systems.
inputs: Commands, artifacts, and hashes consumed.
outputs: Artifacts written and verification results.
gate_results: Deterministic or judgment gates that permitted or blocked progression.
artifact_links: Paths or URLs to durable proof objects.
side_effect_status: Whether publish, queue, schedule, or public mutation occurred.
terminal_state: completed, blocked, failed, or not_run — never inferred.

A model response, tool call, or green status is not a receipt.

Deterministic agent context example

Fixture export only — secret-free, scoped, no mutable runtime state, no unrestricted execution language. Product-repo agents consume approved exports; they do not browse private orchestration source.

Agent export

launchfile_agent_export_v0

{
  "schemaVersion": "launchfile_agent_export_v0",
  "generatedAt": "2026-06-19T19:15:00+07:00",
  "sourceMode": "fixture",
  "scope": {
    "workspaceId": "workspace_launchfile_fixture",
    "productId": "product_launchfile_fixture"
  },
  "currentApprovedDecision": {
    "id": "decision_visual_direction_v2",
    "title": "Use the Verified Operating File visual direction",
    "decision": "Adopt the supplied Launchfile brand board as the V0 visual source for page and component protocol work.",
    "rationale": [
      "The file-first mark and restrained palette express operating-file, verified, controlled, and forward principles.",
      "The selected direction removes the previous visual-source blocker without authorizing production behavior."
    ],
    "authority": "fixture_founder"
  },
  "workOrder": {
    "id": "work_order_frontend_protocol_v2",
    "objective": "Make future Launchfile pages contract-led, visually coherent, proof-oriented, responsive, and agent-readable.",
    "owner": "fixture_codex",
    "allowedPaths": [
      "docs/LAUNCHFILE_*.md",
      "docs/page-contracts/**",
      "apps/launchfile-web/**",
      "__tests__/launchfile/**"
    ],
    "allowedSystems": [
      "local filesystem",
      "local app",
      "local browser",
      "local tests"
    ],
    "acceptanceCriteria": [
      "No route exists without a Page Contract.",
      "The Today route exposes fixture mode, blockers, proof, and not_run truth.",
      "The component demo shows all ten approved primitives.",
      "Typecheck, focused tests, build, and visual QA pass."
    ],
    "forbiddenActions": [
      "Deploy or provision infrastructure.",
      "Mutate a database or production system.",
      "Wire approval, execution, publishing, or Content OS runtime behavior.",
      "Rename Content OS runtime identifiers."
    ],
    "requiredProof": [
      "Launchfile app typecheck passes.",
      "Focused Launchfile tests pass.",
      "Launchfile production build passes.",
      "Desktop and mobile screenshots pass visual QA."
    ]
  },
  "doctrineRefs": [
    "AGENTS.md",
    "docs/LAUNCHFILE_PAGE_CREATION_PROTOCOL_V0.md",
    "docs/LAUNCHFILE_UI_COMPONENT_SPEC_V0.md",
    "docs/LAUNCHFILE_VISUAL_QA_SCORECARD_V0.md"
  ],
  "executionAuthority": {
    "mode": "execute_with_approval",
    "approvalId": "approval_visual_direction_v2",
    "allowedActions": [
      "Edit approved local docs and fixture UI files",
      "Run local verification and visual QA"
    ]
  },
  "executionTruth": {
    "state": "not_run",
    "reason": "This fixture contains no workflow run ledger; local UI work is not runtime execution proof.",
    "latestReceipt": null
  },
  "unresolvedGates": [
    {
      "id": "gate_behavior_wiring_v2",
      "gateId": "BLOCK_BEHAVIOR_WIRING",
      "outcome": "fail",
      "severity": "blocking",
      "reason": "No separate product approval, authorization model, idempotency contract, or receipt-producing mutation path exists."
    }
  ]
}

Integration path

Progressive maturity from read-only context to scoped execution tools.

V0
Read approved context
Agents consume the operating file, doctrine refs, and deterministic exports like the example below.
Next
Propose signals, decisions, and work orders
Draft bounded objects for founder review without treating proposals as approved truth.
Later
Execute-with-approval tools
Run only the exact approved action with idempotency keys, expected side effects, and rollback rules.
Future
MCP resources and tools
Expose read/propose/execute surfaces after scopes, receipts, idempotency, and rollback are explicit.

Launchfiles / Content OS boundary

Launchfiles

The master operating file and founder/agent product surface across decisions, scoped work, proof, readouts, and memory.

Content OS

The content-specific module for evidence, strategy, generation, quality gates, review, and content readouts. Internal commands and audit paths are not public naming guidance.

Repository agents

Engineering agents opened in this repository may patch Launchfiles product code. Product-repo agents consume exported artifacts and approved execution packets only.

No mimicry

Agents do not browse or mimic private orchestration source unless explicitly opened in the repo. Content OS can generate and evaluate page packets; it is not the public master brand.

What agents may do

Core command test

Scope model

Object rendering law

Receipt requirement

Deterministic agent context example

Integration path

V0

Read approved context

Next

Propose signals, decisions, and work orders

Later

Execute-with-approval tools

Future

MCP resources and tools

Launchfiles / Content OS boundary

Launchfiles

Content OS

Repository agents

No mimicry