{
  "$schema": "https://json-schema.org/draft/2020-12/schema",
  "$id": "https://deepworkplan.com/schema/benchmark-record/v1.json",
  "description": "Deep Work Plan benchmark record, v1 (spec/BENCHMARK.md). One closed JSON object per completed v6 plan, emitted into the plan's analysis_results/ by shared/benchmark.py when the repository opted into benchmark mode (.dwp/config.json or ~/.dwp/config.json). The record is derived, never imputed: every field comes from records the plan already owns (journal.ndjson events, manifest.json, contract.json, state.json, best-effort git diff stats). Quantities with no recording source are null - most notably tokens and spend, which exist only when resource_sample events were journaled by a metering host. The record carries no emission timestamp: two runs over identical plan bytes produce identical bytes (spec section 8). v5-generation plans are never measured. The subsystem is opt-in and never a conformance gate. The optional timing.task_spans and context_accounting fields are a pre-release v1 amendment (ADR-0005 in the skill repository): no released pack minted this $id before they landed, optional-only so every earlier committed record still validates.",
  "type": "object",
  "additionalProperties": false,
  "required": [
    "schema",
    "plan",
    "title",
    "generation",
    "contract_id",
    "status",
    "versions",
    "timing",
    "shape",
    "friction",
    "gates",
    "metered",
    "environment",
    "diff_stats"
  ],
  "properties": {
    "schema": {
      "const": "https://deepworkplan.com/schema/benchmark-record/v1.json"
    },
    "plan": {
      "type": "string",
      "pattern": "^PLAN_([0-9]{3,}_)?[a-z0-9]+(_[a-z0-9]+){1,4}$",
      "description": "Plan folder name, same grammar as the journal's plan field."
    },
    "title": {
      "type": "string",
      "minLength": 1,
      "description": "Human plan title from the manifest."
    },
    "generation": {
      "enum": ["v6", "v7"],
      "description": "Only v6- and v7-generation plans are measured; the v5 line is frozen."
    },
    "contract_id": {
      "type": "string",
      "pattern": "^[0-9a-f]{64}$",
      "description": "The plan's materialized outcome contract id."
    },
    "status": {
      "enum": [
        "ready",
        "in_progress",
        "blocked",
        "completed"
      ],
      "description": "Plan status projected from state.json at derivation time."
    },
    "versions": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "dwp_skill",
        "spec",
        "agent_tool"
      ],
      "properties": {
        "dwp_skill": {
          "type": "string",
          "pattern": "^[0-9]+\\.[0-9]+\\.[0-9]+$",
          "description": "Version of the emitting pack (its SKILL.md frontmatter version)."
        },
        "spec": {
          "type": "string",
          "description": "DWP standard the plan recorded (for example 6.0.0)."
        },
        "agent_tool": {
          "type": "string",
          "description": "Detected agent tool name, mirroring context.sh's families and labels (claude-code, codex-cli, cursor, openclaw, gemini-cli, windsurf), or 'unknown'."
        }
      }
    },
    "timing": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "first_event_ts",
        "last_event_ts",
        "span_seconds",
        "task_count_spanned"
      ],
      "properties": {
        "first_event_ts": {
          "type": "string",
          "format": "date-time",
          "description": "Timestamp of the first journal event (the approval)."
        },
        "last_event_ts": {
          "type": "string",
          "format": "date-time",
          "description": "Timestamp of the last journal event at derivation time."
        },
        "span_seconds": {
          "type": "integer",
          "minimum": 0,
          "description": "Calendar span between first and last event. NOT compute time: sessions idle, hosts restart, humans review (spec section 3)."
        },
        "task_count_spanned": {
          "type": "integer",
          "minimum": 0,
          "description": "Number of distinct tasks with at least one task_start event - interrupted and retried tasks count once."
        },
        "task_spans": {
          "type": "array",
          "description": "Per-task calendar spans, one entry per distinct task_start event (an interrupted-and-restarted task carries one entry per start). Optional field: a pre-release v1 amendment; absent when the journal records no task_start events. Calendar span, not compute time (spec section 3).",
          "items": {
            "type": "object",
            "additionalProperties": false,
            "required": [
              "task",
              "start_ts",
              "end_ts",
              "span_seconds"
            ],
            "properties": {
              "task": {
                "type": "string",
                "pattern": "^T-[a-z0-9]+(-[a-z0-9]+)*$",
                "description": "Task id from the contract."
              },
              "start_ts": {
                "type": "string",
                "format": "date-time",
                "description": "Timestamp of the task's task_start event."
              },
              "end_ts": {
                "type": [
                  "string",
                  "null"
                ],
                "format": "date-time",
                "description": "Timestamp of the task's completion evidence; null when the window holds no completion evidence for that start - never zero, never imputed."
              },
              "span_seconds": {
                "type": [
                  "integer",
                  "null"
                ],
                "minimum": 0,
                "description": "Calendar distance start_ts to end_ts; null exactly when end_ts is null."
              }
            }
          }
        }
      }
    },
    "shape": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "tasks",
        "criteria",
        "invariants",
        "gate_intents",
        "events"
      ],
      "properties": {
        "tasks": {
          "type": "integer",
          "minimum": 1,
          "description": "Task count from the contract."
        },
        "criteria": {
          "type": "integer",
          "minimum": 1,
          "description": "Acceptance criteria declared in the contract."
        },
        "invariants": {
          "type": "integer",
          "minimum": 0,
          "description": "Invariants declared in the contract."
        },
        "gate_intents": {
          "type": "integer",
          "minimum": 0,
          "description": "Sum of declared gate-intent criteria across tasks."
        },
        "events": {
          "type": "integer",
          "minimum": 1,
          "description": "Total journal events at derivation time."
        }
      }
    },
    "friction": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "adaptations",
        "amendments",
        "interventions",
        "refusals",
        "retries"
      ],
      "properties": {
        "adaptations": {
          "type": "integer",
          "minimum": 0,
          "description": "Count of adaptation events."
        },
        "amendments": {
          "type": "integer",
          "minimum": 0,
          "description": "Count of amendment events (contract revisions)."
        },
        "interventions": {
          "type": "integer",
          "minimum": 0,
          "description": "Count of intervention events (human or host break-ins)."
        },
        "refusals": {
          "type": "integer",
          "minimum": 0,
          "description": "Count of refusal events (every refusal is itself recorded)."
        },
        "retries": {
          "type": "integer",
          "minimum": 0,
          "description": "Count of gate_run events sharing a criterion with an earlier failing gate_run for the same task."
        }
      }
    },
    "gates": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "runs",
        "exit_0",
        "exit_nonzero",
        "evidence_histogram"
      ],
      "properties": {
        "runs": {
          "type": "integer",
          "minimum": 1,
          "description": "Total gate_run events."
        },
        "exit_0": {
          "type": "integer",
          "minimum": 0,
          "description": "Gate runs that exited 0."
        },
        "exit_nonzero": {
          "type": "integer",
          "minimum": 0,
          "description": "Gate runs that exited nonzero, including 124 timeouts."
        },
        "evidence_histogram": {
          "type": "object",
          "additionalProperties": false,
          "required": [
            "observed",
            "imported",
            "asserted"
          ],
          "properties": {
            "observed": {
              "type": "integer",
              "minimum": 0,
              "description": "Evidence minted by a shipped executor itself (A1)."
            },
            "imported": {
              "type": "integer",
              "minimum": 0,
              "description": "Evidence carried in from outside the journal's own execution."
            },
            "asserted": {
              "type": "integer",
              "minimum": 0,
              "description": "Claims by a mediating actor - never behavior."
            }
          }
        }
      }
    },
    "metered": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "flag",
        "tokens",
        "spend_usd"
      ],
      "properties": {
        "flag": {
          "type": "boolean",
          "description": "False unless a token or spend resource_sample was observed; samples for other limit ids are advisory and never set the flag. Never synthesized."
        },
        "tokens": {
          "type": [
            "integer",
            "null"
          ],
          "minimum": 0,
          "description": "Latest observed token resource_sample (highest seq, selection unit \"tokens\" - AGENT_PROTOCOL 8.4, never a sum of samples); null when not metered (spec section 3 - missing stays missing)."
        },
        "spend_usd": {
          "type": [
            "number",
            "null"
          ],
          "minimum": 0,
          "description": "Latest observed spend resource_sample (highest seq, selection limit_id \"spend_usd\", unit USD - never a sum of samples); null when not metered."
        }
      }
    },
    "environment": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "repo",
        "branch"
      ],
      "properties": {
        "repo": {
          "type": "string",
          "minLength": 1,
          "description": "Repository name (basename only - the privacy boundary forbids absolute user paths, spec section 7)."
        },
        "branch": {
          "type": "string",
          "minLength": 1,
          "description": "Branch name of the plan's repository at derivation time; 'unknown' outside git."
        }
      }
    },
    "diff_stats": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "available",
        "files",
        "insertions",
        "deletions"
      ],
      "properties": {
        "available": {
          "type": "boolean",
          "description": "False whenever the git window could not be computed; the three counts are then null, never zero (spec section 3)."
        },
        "files": {
          "type": [
            "integer",
            "null"
          ],
          "minimum": 0,
          "description": "Files changed in the plan window."
        },
        "insertions": {
          "type": [
            "integer",
            "null"
          ],
          "minimum": 0,
          "description": "Insertions in the plan window."
        },
        "deletions": {
          "type": [
            "integer",
            "null"
          ],
          "minimum": 0,
          "description": "Deletions in the plan window."
        }
      }
    },
    "context_accounting": {
      "type": "object",
      "additionalProperties": false,
      "required": [
        "available",
        "instruction_bytes",
        "provider_tokens",
        "cost_usd",
        "wall_clock_hours"
      ],
      "description": "The four-quantity context accounting block (spec section 3), each quantity from its own source, missing as missing. Optional field: a pre-release v1 amendment. Populated only when the plan's context accounting is recoverable at emission time; otherwise available is false and all four values are null - never synthesized.",
      "properties": {
        "available": {
          "type": "boolean",
          "description": "False when the plan's context accounting could not be recovered at emission time; the four values are then null, never zero."
        },
        "instruction_bytes": {
          "type": [
            "integer",
            "null"
          ],
          "minimum": 0,
          "description": "Measured static instruction bytes (the context manifest plus any declared bundle files); measured, not estimated."
        },
        "provider_tokens": {
          "type": [
            "integer",
            "null"
          ],
          "minimum": 0,
          "description": "Latest observed token meter sample; null without resource_sample events. The same recording as metered.tokens, carried once more so the four quantities read as one unit."
        },
        "cost_usd": {
          "type": [
            "number",
            "null"
          ],
          "minimum": 0,
          "description": "Latest observed spend sample in USD, from real rates only - never bytes-to-money; null without resource_sample events. The same recording as metered.spend_usd."
        },
        "wall_clock_hours": {
          "type": [
            "number",
            "null"
          ],
          "minimum": 0,
          "description": "Recorded event ts span in hours; the same recording as timing.span_seconds, carried once more so the four quantities read as one unit."
        }
      }
    }
  }
}