BINDING.schema.json: machine-readable agent contracts

BINDING.schema.json is the JSON Schema that defines the structure of a Continuity binding — a finite, machine-readable authorization for a recursive computation. A binding specifies who runs a computation, what the objective is, which tools are allowed, what resource budgets apply, and when the computation must stop. For agents that delegate work recursively — spawning child agents, subagents, or bounded subtasks — a binding is the contract that governs that delegation. This file lives at the workspace root and is an Advanced-tier feature.

Full schema

{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "title": "Binding",
  "description": "Finite authorization for a recursive computation",
  "type": "object",
  "required": [
    "binding_id",
    "who",
    "what",
    "budget",
    "allowed_tools",
    "stop_conditions"
  ],
  "properties": {
    "binding_id": {
      "type": "string",
      "description": "Unique identifier"
    },
    "who": {
      "type": "object",
      "required": ["name", "role"],
      "properties": {
        "name": { "type": "string" },
        "role": { "type": "string" },
        "parent": { "type": "string" }
      }
    },
    "what": {
      "type": "object",
      "required": ["objective"],
      "properties": {
        "objective": { "type": "string" },
        "success_criteria": { "type": "string" }
      }
    },
    "budget": {
      "type": "object",
      "properties": {
        "max_seconds": { "type": "integer" },
        "max_tokens": { "type": "integer" },
        "max_turns": { "type": "integer" },
        "max_depth": { "type": "integer" },
        "max_disk_mb": { "type": "integer" }
      }
    },
    "allowed_tools": {
      "type": "array",
      "items": { "type": "string" }
    },
    "checkpoint": {
      "type": "object",
      "properties": {
        "type": {
          "enum": ["manual", "scripted", "evidence"]
        },
        "triggers": {
          "type": "array",
          "items": { "type": "string" }
        }
      }
    },
    "evidence": {
      "type": "object",
      "properties": {
        "required_evidence": {
          "type": "array",
          "items": { "type": "string" }
        },
        "validation_method": { "type": "string" }
      }
    },
    "stop_conditions": {
      "type": "array",
      "items": { "type": "string" }
    },
    "writeback": {
      "type": "object",
      "properties": {
        "to": { "type": "string" },
        "format": { "type": "string" }
      }
    }
  }
}

Field reference

Top-level required fields

binding_id

string

required

Unique identifier for this binding. Used to reference the binding in logs, audits, and parent-agent state files. Should be descriptive enough to identify the task at a glance.Example: "bind-auth-refactor-2026-06-27"

who

object

required

Identifies the agent running this binding — its name, operational role, and optional parent agent.

Show who fields

who.name

string

required

The name of the agent executing this binding.Example: "ResearchSubagent"

who.role

string

required

The operational role for this binding — what kind of work this agent performs.Example: "research agent", "backend coder", "security reviewer"

who.parent

string

The binding ID or name of the parent agent that delegated this work. Establishes the delegation chain for tracing and safety inheritance.Example: "bind-orchestrator-root", "OrchestratorAgent"

what

object

required

Defines the objective of this binding and the criteria for considering it successfully completed.

Show what fields

what.objective

string

required

The specific goal this binding authorizes. Should be concrete and bounded — not “improve the codebase” but “refactor the JWT auth module to use RS256 tokens.”Example: "Refactor src/auth/jwt.ts to use RS256 signing."

what.success_criteria

string

The observable condition that indicates the objective has been met. Used by the verification phase and by parent agents evaluating the result.Example: "All existing auth tests pass. No RS256-specific test regressions. PR created."

allowed_tools

string[]

required

The explicit list of tools this agent is authorized to use. Any tool not listed here is implicitly denied. This is the capability boundary of the binding.Example: ["read_file", "write_file", "run_tests", "git_commit"]

stop_conditions

string[]

required

Conditions under which the agent must halt and report back to its parent, regardless of whether the objective is complete. Used to prevent runaway computation.Example: ["budget exhausted", "test suite regression detected", "scope exceeds auth module"]

budget

object

required

Resource limits for this binding. All limits are enforced by the runtime — the agent must not exceed them regardless of whether the objective is complete.

Show budget fields

budget.max_seconds

integer

Maximum wall-clock seconds allowed for this binding.

budget.max_tokens

integer

Maximum tokens the agent may consume across all turns.

budget.max_turns

integer

Maximum number of reasoning turns the agent may take.

budget.max_depth

integer

Maximum recursion depth — how many levels of child bindings may be spawned. Prevents unbounded delegation chains.

budget.max_disk_mb

integer

Maximum disk space in megabytes the agent may write.

Optional fields

checkpoint

object

Configures the approval gate that governs when the agent must pause and seek human or parent-agent confirmation.

Show checkpoint fields

checkpoint.type

string

The checkpoint type. One of:

manual — agent must pause and wait for explicit human approval
scripted — a script or automated check runs and either clears or blocks
evidence — the agent must produce a specified artifact before proceeding

checkpoint.triggers

string[]

Conditions that activate the checkpoint. When any trigger is met, the agent halts until the checkpoint is cleared.Example: ["before any destructive file write", "before database migration"]

evidence

object

Defines what artifacts or outputs must exist before the binding is considered complete. Enforces the Continuity evidence requirement: a task is complete only when a reviewable artifact exists.

Show evidence fields

evidence.required_evidence

string[]

List of artifacts or conditions that must exist when the binding concludes.Example: ["passing test suite", "git commit with SHA", "SNAPSHOT.md updated"]

evidence.validation_method

string

How the evidence is validated — a command, a script path, or a description of the review process.Example: "run npm test and verify zero failures", "./scripts/verify-binding.sh"

writeback

object

Specifies where and in what format the agent should write its result when the binding concludes.

Show writeback fields

writeback.to

string

The path where the result should be written.Example: "completed/2026-06-27-auth-refactor.md", "STATE.md"

writeback.format

string

The format of the written result.Example: "completed-entry", "state-update", "json"

Example binding

A complete binding for a bounded auth-refactor subtask:

{
  "binding_id": "bind-auth-refactor-2026-06-27",
  "who": {
    "name": "BackendSubagent",
    "role": "backend coder",
    "parent": "bind-orchestrator-root"
  },
  "what": {
    "objective": "Refactor src/auth/jwt.ts to use RS256 signing instead of HS256.",
    "success_criteria": "All existing auth tests pass. No RS256-specific regressions. PR branch created."
  },
  "budget": {
    "max_seconds": 1800,
    "max_tokens": 50000,
    "max_turns": 30,
    "max_depth": 1,
    "max_disk_mb": 10
  },
  "allowed_tools": [
    "read_file",
    "write_file",
    "run_tests",
    "git_diff",
    "git_commit"
  ],
  "checkpoint": {
    "type": "evidence",
    "triggers": [
      "before modifying any file outside src/auth/",
      "before committing"
    ]
  },
  "evidence": {
    "required_evidence": [
      "passing test suite (npm test exits 0)",
      "git commit SHA recorded in writeback"
    ],
    "validation_method": "run npm test and verify zero failures"
  },
  "stop_conditions": [
    "budget exhausted",
    "test suite regression with no clear fix path",
    "scope expands outside src/auth/"
  ],
  "writeback": {
    "to": "completed/2026-06-27-auth-refactor.md",
    "format": "completed-entry"
  }
}

Binding lifecycle

A binding moves through seven phases from creation to release:

POTENTIAL → BIND → RECURSE → COMPOSE → VERIFY → PERSIST → UNBIND

Phase	What happens
`POTENTIAL`	Capability exists but is not yet authorized
`BIND`	Parent creates the binding; agent receives authorization
`RECURSE`	Agent executes; may spawn child bindings (up to `max_depth`)
`COMPOSE`	Results from child bindings are assembled
`VERIFY`	Evidence is validated against `evidence.required_evidence`
`PERSIST`	Results written to `writeback.to`
`UNBIND`	Authorization released; binding is closed

When to use

BINDING.schema.json is an Advanced-tier feature. Most agents do not need it. Use it when:

An orchestrator agent spawns child agents with bounded scopes
You need machine-readable enforcement of capability boundaries
You need a verifiable audit trail of what each subagent was authorized to do
You need max_depth controls to prevent unbounded recursive delegation

For single agents, CHECKPOINT.md and WARNING.md are sufficient. The binding schema is for recursive multi-agent delegation.

For the full binding lifecycle, safety inheritance rules, and runtime integration details, see Runtime Contracts.

File Taxonomy

Schema

BINDING.schema.json: machine-readable agent contracts

Full schema

Field reference

Top-level required fields

Optional fields

Example binding

Binding lifecycle

When to use

Build docs developers (and LLMs) love

File Taxonomy

Schema

Documentation Index

​Full schema

​Field reference

​Top-level required fields

​Optional fields

​Example binding

​Binding lifecycle

​When to use

Build docs developers (and LLMs) love

Full schema

Field reference

Top-level required fields

Optional fields

Example binding

Binding lifecycle

When to use