@xoxoai/checkmate 0.6.7 → 0.7.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +54 -257
- package/dist/checkmate.js +13375 -0
- package/package.json +30 -94
- package/plugin/.claude-plugin/plugin.json +10 -0
- package/plugin/bin/checkmate.ts +3 -0
- package/plugin/contracts/finding.schema.json +78 -0
- package/plugin/contracts/gap.schema.json +56 -0
- package/plugin/contracts/plan-candidate.schema.json +103 -0
- package/plugin/contracts/plan-delta.schema.json +63 -0
- package/plugin/contracts/plan-patch.schema.json +116 -0
- package/plugin/contracts/plan-request.schema.json +108 -0
- package/plugin/contracts/plan.schema.json +313 -0
- package/plugin/contracts/profile.schema.json +154 -0
- package/plugin/contracts/response.schema.json +25 -0
- package/plugin/contracts/result.schema.json +410 -0
- package/plugin/dist/checkmate.js +13012 -0
- package/plugin/package.json +4 -0
- package/plugin/references/browser.md +35 -0
- package/plugin/references/http.md +30 -0
- package/plugin/references/workflow.md +122 -0
- package/plugin/skills/execute/SKILL.md +92 -0
- package/plugin/skills/plan/SKILL.md +94 -0
- package/plugin/skills/report/SKILL.md +28 -0
- package/plugin/skills/revise/SKILL.md +43 -0
- package/plugin/skills/setup/SKILL.md +82 -0
- package/plugin/skills/status/SKILL.md +20 -0
- package/plugin/skills/ui/SKILL.md +26 -0
- package/plugin/src/actor.ts +21 -0
- package/plugin/src/cli.ts +1133 -0
- package/plugin/src/contracts.ts +342 -0
- package/plugin/src/index.ts +52 -0
- package/plugin/src/plans.ts +1028 -0
- package/plugin/src/render.ts +544 -0
- package/plugin/src/runs.ts +868 -0
- package/plugin/src/store.ts +384 -0
- package/plugin/src/targets.ts +205 -0
- package/plugin/src/types.ts +319 -0
- package/plugin/src/ui/assets/client.js +501 -0
- package/plugin/src/ui/assets/favicon.svg +1 -0
- package/plugin/src/ui/assets/fonts/inter-OFL.txt +93 -0
- package/plugin/src/ui/assets/fonts/inter.woff2 +0 -0
- package/plugin/src/ui/assets/fonts/jetbrains-mono-OFL.txt +93 -0
- package/plugin/src/ui/assets/fonts/jetbrains-mono.woff2 +0 -0
- package/plugin/src/ui/assets/fonts/space-grotesk-OFL.txt +93 -0
- package/plugin/src/ui/assets/fonts/space-grotesk.woff2 +0 -0
- package/plugin/src/ui/assets/style.css +1926 -0
- package/plugin/src/ui/edit.ts +305 -0
- package/plugin/src/ui/forms.ts +209 -0
- package/plugin/src/ui/html.ts +130 -0
- package/plugin/src/ui/layout.ts +198 -0
- package/plugin/src/ui/lists.ts +645 -0
- package/plugin/src/ui/plan.ts +577 -0
- package/plugin/src/ui/run.ts +489 -0
- package/plugin/src/ui/server.ts +684 -0
- package/plugin/src/validators.d.ts +13 -0
- package/plugin/src/validators.js +1 -0
- package/plugin/src/views.ts +839 -0
- package/bin/checkmate.js +0 -13
- package/dist/ai/client.d.ts +0 -64
- package/dist/ai/client.d.ts.map +0 -1
- package/dist/ai/client.js +0 -313
- package/dist/ai/client.js.map +0 -1
- package/dist/ai/message-handler.d.ts +0 -9
- package/dist/ai/message-handler.d.ts.map +0 -1
- package/dist/ai/message-handler.js +0 -50
- package/dist/ai/message-handler.js.map +0 -1
- package/dist/ai/message-history.d.ts +0 -17
- package/dist/ai/message-history.d.ts.map +0 -1
- package/dist/ai/message-history.js +0 -45
- package/dist/ai/message-history.js.map +0 -1
- package/dist/ai/prompts.d.ts +0 -4
- package/dist/ai/prompts.d.ts.map +0 -1
- package/dist/ai/prompts.js +0 -24
- package/dist/ai/prompts.js.map +0 -1
- package/dist/ai/rate-limit-policy.d.ts +0 -9
- package/dist/ai/rate-limit-policy.d.ts.map +0 -1
- package/dist/ai/rate-limit-policy.js +0 -34
- package/dist/ai/rate-limit-policy.js.map +0 -1
- package/dist/ai/tool-response-handler.d.ts +0 -16
- package/dist/ai/tool-response-handler.d.ts.map +0 -1
- package/dist/ai/tool-response-handler.js +0 -85
- package/dist/ai/tool-response-handler.js.map +0 -1
- package/dist/ai/turn-processor.d.ts +0 -34
- package/dist/ai/turn-processor.d.ts.map +0 -1
- package/dist/ai/turn-processor.js +0 -107
- package/dist/ai/turn-processor.js.map +0 -1
- package/dist/api/allocate-run-identity.d.ts +0 -8
- package/dist/api/allocate-run-identity.d.ts.map +0 -1
- package/dist/api/allocate-run-identity.js +0 -10
- package/dist/api/allocate-run-identity.js.map +0 -1
- package/dist/api/execute-prepared-run.d.ts +0 -22
- package/dist/api/execute-prepared-run.d.ts.map +0 -1
- package/dist/api/execute-prepared-run.js +0 -184
- package/dist/api/execute-prepared-run.js.map +0 -1
- package/dist/api/index.d.ts +0 -25
- package/dist/api/index.d.ts.map +0 -1
- package/dist/api/index.js +0 -69
- package/dist/api/index.js.map +0 -1
- package/dist/api/prepare-run.d.ts +0 -44
- package/dist/api/prepare-run.d.ts.map +0 -1
- package/dist/api/prepare-run.js +0 -261
- package/dist/api/prepare-run.js.map +0 -1
- package/dist/cli/diagnostic-relay.d.ts +0 -6
- package/dist/cli/diagnostic-relay.d.ts.map +0 -1
- package/dist/cli/diagnostic-relay.js +0 -6
- package/dist/cli/diagnostic-relay.js.map +0 -1
- package/dist/cli/main.d.ts +0 -21
- package/dist/cli/main.d.ts.map +0 -1
- package/dist/cli/main.js +0 -166
- package/dist/cli/main.js.map +0 -1
- package/dist/cli/protocol.d.ts +0 -57
- package/dist/cli/protocol.d.ts.map +0 -1
- package/dist/cli/protocol.js +0 -116
- package/dist/cli/protocol.js.map +0 -1
- package/dist/cli/reconcile-result.d.ts +0 -13
- package/dist/cli/reconcile-result.d.ts.map +0 -1
- package/dist/cli/reconcile-result.js +0 -300
- package/dist/cli/reconcile-result.js.map +0 -1
- package/dist/cli/run-parent.d.ts +0 -37
- package/dist/cli/run-parent.d.ts.map +0 -1
- package/dist/cli/run-parent.js +0 -491
- package/dist/cli/run-parent.js.map +0 -1
- package/dist/cli/signals.d.ts +0 -36
- package/dist/cli/signals.d.ts.map +0 -1
- package/dist/cli/signals.js +0 -85
- package/dist/cli/signals.js.map +0 -1
- package/dist/cli/static-commands.d.ts +0 -11
- package/dist/cli/static-commands.d.ts.map +0 -1
- package/dist/cli/static-commands.js +0 -83
- package/dist/cli/static-commands.js.map +0 -1
- package/dist/cli/worker-stream-guard.d.ts +0 -20
- package/dist/cli/worker-stream-guard.d.ts.map +0 -1
- package/dist/cli/worker-stream-guard.js +0 -41
- package/dist/cli/worker-stream-guard.js.map +0 -1
- package/dist/cli/worker.d.ts +0 -3
- package/dist/cli/worker.d.ts.map +0 -1
- package/dist/cli/worker.js +0 -111
- package/dist/cli/worker.js.map +0 -1
- package/dist/config/ingestion.d.ts +0 -58
- package/dist/config/ingestion.d.ts.map +0 -1
- package/dist/config/ingestion.js +0 -283
- package/dist/config/ingestion.js.map +0 -1
- package/dist/config/manifest.d.ts +0 -20
- package/dist/config/manifest.d.ts.map +0 -1
- package/dist/config/manifest.js +0 -91
- package/dist/config/manifest.js.map +0 -1
- package/dist/config/model-egress.d.ts +0 -10
- package/dist/config/model-egress.d.ts.map +0 -1
- package/dist/config/model-egress.js +0 -62
- package/dist/config/model-egress.js.map +0 -1
- package/dist/config/package-resolution.d.ts +0 -4
- package/dist/config/package-resolution.d.ts.map +0 -1
- package/dist/config/package-resolution.js +0 -24
- package/dist/config/package-resolution.js.map +0 -1
- package/dist/config/policy.d.ts +0 -44
- package/dist/config/policy.d.ts.map +0 -1
- package/dist/config/policy.js +0 -141
- package/dist/config/policy.js.map +0 -1
- package/dist/config/record.d.ts +0 -4
- package/dist/config/record.d.ts.map +0 -1
- package/dist/config/record.js +0 -10
- package/dist/config/record.js.map +0 -1
- package/dist/config/secrets.d.ts +0 -11
- package/dist/config/secrets.d.ts.map +0 -1
- package/dist/config/secrets.js +0 -42
- package/dist/config/secrets.js.map +0 -1
- package/dist/contracts/diagnostics.d.ts +0 -4
- package/dist/contracts/diagnostics.d.ts.map +0 -1
- package/dist/contracts/diagnostics.js +0 -31
- package/dist/contracts/diagnostics.js.map +0 -1
- package/dist/contracts/serialize.d.ts +0 -2
- package/dist/contracts/serialize.d.ts.map +0 -1
- package/dist/contracts/serialize.js +0 -90
- package/dist/contracts/serialize.js.map +0 -1
- package/dist/contracts/types.d.ts +0 -316
- package/dist/contracts/types.d.ts.map +0 -1
- package/dist/contracts/types.js +0 -74
- package/dist/contracts/types.js.map +0 -1
- package/dist/contracts/validator.d.ts +0 -8
- package/dist/contracts/validator.d.ts.map +0 -1
- package/dist/contracts/validator.js +0 -71
- package/dist/contracts/validator.js.map +0 -1
- package/dist/driver.d.ts +0 -97
- package/dist/driver.d.ts.map +0 -1
- package/dist/driver.js +0 -24
- package/dist/driver.js.map +0 -1
- package/dist/drivers/descriptor.d.ts +0 -15
- package/dist/drivers/descriptor.d.ts.map +0 -1
- package/dist/drivers/descriptor.js +0 -122
- package/dist/drivers/descriptor.js.map +0 -1
- package/dist/drivers/loader.d.ts +0 -18
- package/dist/drivers/loader.d.ts.map +0 -1
- package/dist/drivers/loader.js +0 -97
- package/dist/drivers/loader.js.map +0 -1
- package/dist/drivers/web/checkmate-driver.json +0 -47
- package/dist/drivers/web/index.d.ts +0 -6
- package/dist/drivers/web/index.d.ts.map +0 -1
- package/dist/drivers/web/index.js +0 -27
- package/dist/drivers/web/index.js.map +0 -1
- package/dist/drivers/web/session.d.ts +0 -25
- package/dist/drivers/web/session.d.ts.map +0 -1
- package/dist/drivers/web/session.js +0 -132
- package/dist/drivers/web/session.js.map +0 -1
- package/dist/drivers/web/tools/network-request-recorder.d.ts +0 -46
- package/dist/drivers/web/tools/network-request-recorder.d.ts.map +0 -1
- package/dist/drivers/web/tools/network-request-recorder.js +0 -164
- package/dist/drivers/web/tools/network-request-recorder.js.map +0 -1
- package/dist/drivers/web/tools/screenshot-service.d.ts +0 -10
- package/dist/drivers/web/tools/screenshot-service.d.ts.map +0 -1
- package/dist/drivers/web/tools/screenshot-service.js +0 -23
- package/dist/drivers/web/tools/screenshot-service.js.map +0 -1
- package/dist/drivers/web/tools/snapshot-filter/index.d.ts +0 -4
- package/dist/drivers/web/tools/snapshot-filter/index.d.ts.map +0 -1
- package/dist/drivers/web/tools/snapshot-filter/index.js +0 -4
- package/dist/drivers/web/tools/snapshot-filter/index.js.map +0 -1
- package/dist/drivers/web/tools/snapshot-filter/semantic-scorer.d.ts +0 -15
- package/dist/drivers/web/tools/snapshot-filter/semantic-scorer.d.ts.map +0 -1
- package/dist/drivers/web/tools/snapshot-filter/semantic-scorer.js +0 -99
- package/dist/drivers/web/tools/snapshot-filter/semantic-scorer.js.map +0 -1
- package/dist/drivers/web/tools/snapshot-filter/snapshot-filter.d.ts +0 -5
- package/dist/drivers/web/tools/snapshot-filter/snapshot-filter.d.ts.map +0 -1
- package/dist/drivers/web/tools/snapshot-filter/snapshot-filter.js +0 -56
- package/dist/drivers/web/tools/snapshot-filter/snapshot-filter.js.map +0 -1
- package/dist/drivers/web/tools/snapshot-filter/tree-reconstructor.d.ts +0 -3
- package/dist/drivers/web/tools/snapshot-filter/tree-reconstructor.d.ts.map +0 -1
- package/dist/drivers/web/tools/snapshot-filter/tree-reconstructor.js +0 -85
- package/dist/drivers/web/tools/snapshot-filter/tree-reconstructor.js.map +0 -1
- package/dist/drivers/web/tools/snapshot-service.d.ts +0 -31
- package/dist/drivers/web/tools/snapshot-service.d.ts.map +0 -1
- package/dist/drivers/web/tools/snapshot-service.js +0 -96
- package/dist/drivers/web/tools/snapshot-service.js.map +0 -1
- package/dist/drivers/web/tools/tool.d.ts +0 -76
- package/dist/drivers/web/tools/tool.d.ts.map +0 -1
- package/dist/drivers/web/tools/tool.js +0 -562
- package/dist/drivers/web/tools/tool.js.map +0 -1
- package/dist/drivers/web/tools/transient-state-tracker.d.ts +0 -26
- package/dist/drivers/web/tools/transient-state-tracker.d.ts.map +0 -1
- package/dist/drivers/web/tools/transient-state-tracker.js +0 -301
- package/dist/drivers/web/tools/transient-state-tracker.js.map +0 -1
- package/dist/drivers/web/tools/types.d.ts +0 -11
- package/dist/drivers/web/tools/types.d.ts.map +0 -1
- package/dist/drivers/web/tools/types.js +0 -2
- package/dist/drivers/web/tools/types.js.map +0 -1
- package/dist/evidence/atomic-file.d.ts +0 -23
- package/dist/evidence/atomic-file.d.ts.map +0 -1
- package/dist/evidence/atomic-file.js +0 -104
- package/dist/evidence/atomic-file.js.map +0 -1
- package/dist/evidence/layout.d.ts +0 -30
- package/dist/evidence/layout.d.ts.map +0 -1
- package/dist/evidence/layout.js +0 -152
- package/dist/evidence/layout.js.map +0 -1
- package/dist/evidence/retention.d.ts +0 -5
- package/dist/evidence/retention.d.ts.map +0 -1
- package/dist/evidence/retention.js +0 -10
- package/dist/evidence/retention.js.map +0 -1
- package/dist/evidence/store.d.ts +0 -101
- package/dist/evidence/store.d.ts.map +0 -1
- package/dist/evidence/store.js +0 -375
- package/dist/evidence/store.js.map +0 -1
- package/dist/evidence/terminal-finalizer.d.ts +0 -19
- package/dist/evidence/terminal-finalizer.d.ts.map +0 -1
- package/dist/evidence/terminal-finalizer.js +0 -56
- package/dist/evidence/terminal-finalizer.js.map +0 -1
- package/dist/index.d.ts +0 -9
- package/dist/index.d.ts.map +0 -1
- package/dist/index.js +0 -12
- package/dist/index.js.map +0 -1
- package/dist/logging/invocation-logger.d.ts +0 -4
- package/dist/logging/invocation-logger.d.ts.map +0 -1
- package/dist/logging/invocation-logger.js +0 -25
- package/dist/logging/invocation-logger.js.map +0 -1
- package/dist/logging/level-logger.d.ts +0 -11
- package/dist/logging/level-logger.d.ts.map +0 -1
- package/dist/logging/level-logger.js +0 -82
- package/dist/logging/level-logger.js.map +0 -1
- package/dist/logging/types.d.ts +0 -8
- package/dist/logging/types.d.ts.map +0 -1
- package/dist/logging/types.js +0 -7
- package/dist/logging/types.js.map +0 -1
- package/dist/redaction/diagnostic-sanitizer.d.ts +0 -7
- package/dist/redaction/diagnostic-sanitizer.d.ts.map +0 -1
- package/dist/redaction/diagnostic-sanitizer.js +0 -42
- package/dist/redaction/diagnostic-sanitizer.js.map +0 -1
- package/dist/redaction/redactor.d.ts +0 -40
- package/dist/redaction/redactor.d.ts.map +0 -1
- package/dist/redaction/redactor.js +0 -176
- package/dist/redaction/redactor.js.map +0 -1
- package/dist/redaction/scrub.d.ts +0 -29
- package/dist/redaction/scrub.d.ts.map +0 -1
- package/dist/redaction/scrub.js +0 -54
- package/dist/redaction/scrub.js.map +0 -1
- package/dist/runtime/config.d.ts +0 -16
- package/dist/runtime/config.d.ts.map +0 -1
- package/dist/runtime/config.js +0 -2
- package/dist/runtime/config.js.map +0 -1
- package/dist/runtime/driver-boundary.d.ts +0 -16
- package/dist/runtime/driver-boundary.d.ts.map +0 -1
- package/dist/runtime/driver-boundary.js +0 -71
- package/dist/runtime/driver-boundary.js.map +0 -1
- package/dist/runtime/driver-session.d.ts +0 -9
- package/dist/runtime/driver-session.d.ts.map +0 -1
- package/dist/runtime/driver-session.js +0 -64
- package/dist/runtime/driver-session.js.map +0 -1
- package/dist/runtime/internal-step-evidence.d.ts +0 -30
- package/dist/runtime/internal-step-evidence.d.ts.map +0 -1
- package/dist/runtime/internal-step-evidence.js +0 -82
- package/dist/runtime/internal-step-evidence.js.map +0 -1
- package/dist/runtime/result-adapter.d.ts +0 -9
- package/dist/runtime/result-adapter.d.ts.map +0 -1
- package/dist/runtime/result-adapter.js +0 -77
- package/dist/runtime/result-adapter.js.map +0 -1
- package/dist/runtime/runner.d.ts +0 -43
- package/dist/runtime/runner.d.ts.map +0 -1
- package/dist/runtime/runner.js +0 -72
- package/dist/runtime/runner.js.map +0 -1
- package/dist/runtime/scenario-control.d.ts +0 -40
- package/dist/runtime/scenario-control.d.ts.map +0 -1
- package/dist/runtime/scenario-control.js +0 -94
- package/dist/runtime/scenario-control.js.map +0 -1
- package/dist/runtime/scenario-runner.d.ts +0 -41
- package/dist/runtime/scenario-runner.d.ts.map +0 -1
- package/dist/runtime/scenario-runner.js +0 -274
- package/dist/runtime/scenario-runner.js.map +0 -1
- package/dist/runtime/scenario-state.d.ts +0 -55
- package/dist/runtime/scenario-state.d.ts.map +0 -1
- package/dist/runtime/scenario-state.js +0 -128
- package/dist/runtime/scenario-state.js.map +0 -1
- package/dist/runtime/step-execution.d.ts +0 -27
- package/dist/runtime/step-execution.d.ts.map +0 -1
- package/dist/runtime/step-execution.js +0 -143
- package/dist/runtime/step-execution.js.map +0 -1
- package/dist/runtime/text.d.ts +0 -10
- package/dist/runtime/text.d.ts.map +0 -1
- package/dist/runtime/text.js +0 -25
- package/dist/runtime/text.js.map +0 -1
- package/dist/runtime/types.d.ts +0 -53
- package/dist/runtime/types.d.ts.map +0 -1
- package/dist/runtime/types.js +0 -2
- package/dist/runtime/types.js.map +0 -1
- package/dist/runtime/usage-tracker.d.ts +0 -37
- package/dist/runtime/usage-tracker.d.ts.map +0 -1
- package/dist/runtime/usage-tracker.js +0 -110
- package/dist/runtime/usage-tracker.js.map +0 -1
- package/dist/tools/define-agent-tool.d.ts +0 -20
- package/dist/tools/define-agent-tool.d.ts.map +0 -1
- package/dist/tools/define-agent-tool.js +0 -30
- package/dist/tools/define-agent-tool.js.map +0 -1
- package/dist/tools/dispatcher.d.ts +0 -19
- package/dist/tools/dispatcher.d.ts.map +0 -1
- package/dist/tools/dispatcher.js +0 -91
- package/dist/tools/dispatcher.js.map +0 -1
- package/dist/tools/loop-detector.d.ts +0 -27
- package/dist/tools/loop-detector.d.ts.map +0 -1
- package/dist/tools/loop-detector.js +0 -76
- package/dist/tools/loop-detector.js.map +0 -1
- package/dist/tools/registry.d.ts +0 -19
- package/dist/tools/registry.d.ts.map +0 -1
- package/dist/tools/registry.js +0 -54
- package/dist/tools/registry.js.map +0 -1
- package/dist/tools/step/result-tool.d.ts +0 -7
- package/dist/tools/step/result-tool.d.ts.map +0 -1
- package/dist/tools/step/result-tool.js +0 -34
- package/dist/tools/step/result-tool.js.map +0 -1
- package/dist/tools/tool-contract.d.ts +0 -3
- package/dist/tools/tool-contract.d.ts.map +0 -1
- package/dist/tools/tool-contract.js +0 -2
- package/dist/tools/tool-contract.js.map +0 -1
- package/dist/tools/types.d.ts +0 -210
- package/dist/tools/types.d.ts.map +0 -1
- package/dist/tools/types.js +0 -4
- package/dist/tools/types.js.map +0 -1
- package/docs/CLI.md +0 -59
- package/docs/CONFIGURATION.md +0 -105
- package/docs/DRIVERS.md +0 -79
- package/docs/EVIDENCE.md +0 -64
- package/schemas/checkmate-config.v1.json +0 -176
- package/schemas/describe-result.v1.json +0 -414
- package/schemas/driver-descriptor.v1.json +0 -66
- package/schemas/run-request.v1.json +0 -67
- package/schemas/run-result.v1.json +0 -333
- package/schemas/validation-result.v1.json +0 -86
|
@@ -1,333 +0,0 @@
|
|
|
1
|
-
{
|
|
2
|
-
"$schema": "https://json-schema.org/draft/2020-12/schema",
|
|
3
|
-
"$id": "https://checkmate.dev/schemas/run-result.v1.json",
|
|
4
|
-
"title": "Checkmate run result version 1",
|
|
5
|
-
"oneOf": [
|
|
6
|
-
{ "$ref": "#/$defs/executionResult" },
|
|
7
|
-
{ "$ref": "#/$defs/invalidResult" },
|
|
8
|
-
{ "$ref": "#/$defs/preExecutionOperationalResult" },
|
|
9
|
-
{ "$ref": "#/$defs/containmentResult" }
|
|
10
|
-
],
|
|
11
|
-
"$defs": {
|
|
12
|
-
"diagnostic": {
|
|
13
|
-
"type": "object",
|
|
14
|
-
"additionalProperties": false,
|
|
15
|
-
"required": ["code", "path", "message"],
|
|
16
|
-
"properties": {
|
|
17
|
-
"code": { "type": "string", "minLength": 1 },
|
|
18
|
-
"path": { "type": "string", "pattern": "^(?:/(?:[^~/]|~[01])*)*$" },
|
|
19
|
-
"message": { "type": "string", "minLength": 1 }
|
|
20
|
-
}
|
|
21
|
-
},
|
|
22
|
-
"effectiveLimits": {
|
|
23
|
-
"type": "object",
|
|
24
|
-
"additionalProperties": false,
|
|
25
|
-
"required": [
|
|
26
|
-
"scenarioTimeoutMs",
|
|
27
|
-
"stepTimeoutMs",
|
|
28
|
-
"turnsPerStep",
|
|
29
|
-
"requestTimeoutMs",
|
|
30
|
-
"maxRetries",
|
|
31
|
-
"loopMaxRepetitions",
|
|
32
|
-
"cleanupTimeoutMs"
|
|
33
|
-
],
|
|
34
|
-
"properties": {
|
|
35
|
-
"scenarioTimeoutMs": { "type": "integer", "minimum": 1 },
|
|
36
|
-
"stepTimeoutMs": { "type": "integer", "minimum": 1 },
|
|
37
|
-
"turnsPerStep": { "type": "integer", "minimum": 1 },
|
|
38
|
-
"requestTimeoutMs": { "type": "integer", "minimum": 1 },
|
|
39
|
-
"maxRetries": { "type": "integer", "minimum": 0 },
|
|
40
|
-
"loopMaxRepetitions": { "type": "integer", "minimum": 1 },
|
|
41
|
-
"cleanupTimeoutMs": { "type": "integer", "minimum": 1 },
|
|
42
|
-
"budgetTokens": { "type": "integer", "minimum": 1 }
|
|
43
|
-
}
|
|
44
|
-
},
|
|
45
|
-
"usage": {
|
|
46
|
-
"type": "object",
|
|
47
|
-
"additionalProperties": false,
|
|
48
|
-
"required": ["promptTokens", "cachedPromptTokens", "completionTokens", "totalTokens"],
|
|
49
|
-
"properties": {
|
|
50
|
-
"promptTokens": { "type": "integer", "minimum": 0 },
|
|
51
|
-
"cachedPromptTokens": { "type": "integer", "minimum": 0 },
|
|
52
|
-
"completionTokens": { "type": "integer", "minimum": 0 },
|
|
53
|
-
"totalTokens": { "type": "integer", "minimum": 0 }
|
|
54
|
-
}
|
|
55
|
-
},
|
|
56
|
-
"stepReason": {
|
|
57
|
-
"enum": [
|
|
58
|
-
"met-expectation",
|
|
59
|
-
"failed-expectation",
|
|
60
|
-
"loop-detected",
|
|
61
|
-
"turn-cap-exceeded",
|
|
62
|
-
"step-timeout",
|
|
63
|
-
"scenario-timeout",
|
|
64
|
-
"tool-error",
|
|
65
|
-
"provider-error",
|
|
66
|
-
"token-budget-exceeded",
|
|
67
|
-
"interrupted",
|
|
68
|
-
"internal-error"
|
|
69
|
-
]
|
|
70
|
-
},
|
|
71
|
-
"runReason": {
|
|
72
|
-
"enum": [
|
|
73
|
-
"scenario-complete",
|
|
74
|
-
"failed-expectation",
|
|
75
|
-
"loop-detected",
|
|
76
|
-
"turn-cap-exceeded",
|
|
77
|
-
"step-timeout",
|
|
78
|
-
"scenario-timeout",
|
|
79
|
-
"tool-error",
|
|
80
|
-
"provider-error",
|
|
81
|
-
"token-budget-exceeded",
|
|
82
|
-
"interrupted",
|
|
83
|
-
"driver-load-failed",
|
|
84
|
-
"driver-start-failed",
|
|
85
|
-
"driver-teardown-failed",
|
|
86
|
-
"evidence-write-failed",
|
|
87
|
-
"result-write-failed",
|
|
88
|
-
"internal-error"
|
|
89
|
-
]
|
|
90
|
-
},
|
|
91
|
-
"toolCall": {
|
|
92
|
-
"type": "object",
|
|
93
|
-
"additionalProperties": false,
|
|
94
|
-
"required": ["turn", "driverId", "name", "arguments", "status"],
|
|
95
|
-
"properties": {
|
|
96
|
-
"turn": { "type": "integer", "minimum": 1 },
|
|
97
|
-
"driverId": { "type": "string", "minLength": 1 },
|
|
98
|
-
"name": { "type": "string", "minLength": 1 },
|
|
99
|
-
"arguments": true,
|
|
100
|
-
"status": { "enum": ["ok", "error"] }
|
|
101
|
-
}
|
|
102
|
-
},
|
|
103
|
-
"executedStep": {
|
|
104
|
-
"type": "object",
|
|
105
|
-
"additionalProperties": false,
|
|
106
|
-
"required": ["id", "status", "category", "reason", "turns", "durationMs", "usage", "toolCalls"],
|
|
107
|
-
"properties": {
|
|
108
|
-
"id": { "type": "string", "minLength": 1 },
|
|
109
|
-
"status": { "enum": ["passed", "failed"] },
|
|
110
|
-
"category": { "enum": ["app", "model", "infra"] },
|
|
111
|
-
"reason": { "$ref": "#/$defs/stepReason" },
|
|
112
|
-
"actual": { "type": "string" },
|
|
113
|
-
"turns": { "type": "integer", "minimum": 0 },
|
|
114
|
-
"durationMs": { "type": "number", "minimum": 0 },
|
|
115
|
-
"usage": { "$ref": "#/$defs/usage" },
|
|
116
|
-
"toolCalls": { "type": "array", "items": { "$ref": "#/$defs/toolCall" } }
|
|
117
|
-
}
|
|
118
|
-
},
|
|
119
|
-
"notRunAfterFailure": {
|
|
120
|
-
"type": "object",
|
|
121
|
-
"additionalProperties": false,
|
|
122
|
-
"required": ["id", "status", "reason", "blockedBy"],
|
|
123
|
-
"properties": {
|
|
124
|
-
"id": { "type": "string", "minLength": 1 },
|
|
125
|
-
"status": { "const": "not-run" },
|
|
126
|
-
"reason": { "const": "prior-step-failed" },
|
|
127
|
-
"blockedBy": { "type": "string", "minLength": 1 }
|
|
128
|
-
}
|
|
129
|
-
},
|
|
130
|
-
"notRunAtStart": {
|
|
131
|
-
"type": "object",
|
|
132
|
-
"additionalProperties": false,
|
|
133
|
-
"required": ["id", "status", "reason"],
|
|
134
|
-
"properties": {
|
|
135
|
-
"id": { "type": "string", "minLength": 1 },
|
|
136
|
-
"status": { "const": "not-run" },
|
|
137
|
-
"reason": { "const": "scenario-not-started" }
|
|
138
|
-
}
|
|
139
|
-
},
|
|
140
|
-
"stepResult": {
|
|
141
|
-
"oneOf": [
|
|
142
|
-
{ "$ref": "#/$defs/executedStep" },
|
|
143
|
-
{ "$ref": "#/$defs/notRunAfterFailure" },
|
|
144
|
-
{ "$ref": "#/$defs/notRunAtStart" }
|
|
145
|
-
]
|
|
146
|
-
},
|
|
147
|
-
"evidenceReference": {
|
|
148
|
-
"type": "object",
|
|
149
|
-
"additionalProperties": false,
|
|
150
|
-
"required": ["kind", "mediaType", "path", "producer"],
|
|
151
|
-
"properties": {
|
|
152
|
-
"kind": { "type": "string", "minLength": 1 },
|
|
153
|
-
"mediaType": { "type": "string", "minLength": 1 },
|
|
154
|
-
"path": { "type": "string", "minLength": 1 },
|
|
155
|
-
"producer": { "type": "string", "minLength": 1 },
|
|
156
|
-
"stepId": { "type": "string", "minLength": 1 }
|
|
157
|
-
}
|
|
158
|
-
},
|
|
159
|
-
"executionResult": {
|
|
160
|
-
"type": "object",
|
|
161
|
-
"additionalProperties": false,
|
|
162
|
-
"required": [
|
|
163
|
-
"kind",
|
|
164
|
-
"schemaVersion",
|
|
165
|
-
"runId",
|
|
166
|
-
"scenarioId",
|
|
167
|
-
"status",
|
|
168
|
-
"category",
|
|
169
|
-
"reason",
|
|
170
|
-
"targetMutation",
|
|
171
|
-
"startedAt",
|
|
172
|
-
"durationMs",
|
|
173
|
-
"driver",
|
|
174
|
-
"policy",
|
|
175
|
-
"usage",
|
|
176
|
-
"steps",
|
|
177
|
-
"evidence",
|
|
178
|
-
"diagnostics"
|
|
179
|
-
],
|
|
180
|
-
"properties": {
|
|
181
|
-
"kind": { "const": "run-result" },
|
|
182
|
-
"schemaVersion": { "const": 1 },
|
|
183
|
-
"runId": { "type": "string", "pattern": "^[0-9a-f]{16}$" },
|
|
184
|
-
"scenarioId": { "type": "string", "minLength": 1 },
|
|
185
|
-
"status": { "enum": ["passed", "failed", "interrupted"] },
|
|
186
|
-
"category": { "enum": ["passed", "app", "model", "infra"] },
|
|
187
|
-
"reason": { "$ref": "#/$defs/runReason" },
|
|
188
|
-
"targetMutation": { "enum": ["not-attempted", "possibly-mutated"] },
|
|
189
|
-
"startedAt": { "type": "string", "format": "date-time" },
|
|
190
|
-
"durationMs": { "type": "number", "minimum": 0 },
|
|
191
|
-
"driver": {
|
|
192
|
-
"type": "object",
|
|
193
|
-
"additionalProperties": false,
|
|
194
|
-
"required": ["id", "contractVersion"],
|
|
195
|
-
"properties": {
|
|
196
|
-
"id": { "type": "string", "minLength": 1 },
|
|
197
|
-
"contractVersion": { "const": 1 }
|
|
198
|
-
}
|
|
199
|
-
},
|
|
200
|
-
"policy": {
|
|
201
|
-
"type": "object",
|
|
202
|
-
"additionalProperties": false,
|
|
203
|
-
"required": ["id", "effectiveLimits"],
|
|
204
|
-
"properties": {
|
|
205
|
-
"id": { "type": "string", "minLength": 1 },
|
|
206
|
-
"effectiveLimits": { "$ref": "#/$defs/effectiveLimits" }
|
|
207
|
-
}
|
|
208
|
-
},
|
|
209
|
-
"usage": {
|
|
210
|
-
"type": "object",
|
|
211
|
-
"additionalProperties": false,
|
|
212
|
-
"required": ["promptTokens", "cachedPromptTokens", "completionTokens", "totalTokens", "state"],
|
|
213
|
-
"properties": {
|
|
214
|
-
"promptTokens": { "type": "integer", "minimum": 0 },
|
|
215
|
-
"cachedPromptTokens": { "type": "integer", "minimum": 0 },
|
|
216
|
-
"completionTokens": { "type": "integer", "minimum": 0 },
|
|
217
|
-
"totalTokens": { "type": "integer", "minimum": 0 },
|
|
218
|
-
"state": { "enum": ["complete", "partial", "unavailable"] }
|
|
219
|
-
}
|
|
220
|
-
},
|
|
221
|
-
"steps": { "type": "array", "minItems": 1, "items": { "$ref": "#/$defs/stepResult" } },
|
|
222
|
-
"evidence": {
|
|
223
|
-
"type": "object",
|
|
224
|
-
"additionalProperties": false,
|
|
225
|
-
"required": ["state", "references"],
|
|
226
|
-
"properties": {
|
|
227
|
-
"state": { "enum": ["complete", "partial"] },
|
|
228
|
-
"references": { "type": "array", "items": { "$ref": "#/$defs/evidenceReference" } }
|
|
229
|
-
}
|
|
230
|
-
},
|
|
231
|
-
"diagnostics": { "type": "array", "items": { "$ref": "#/$defs/diagnostic" } }
|
|
232
|
-
}
|
|
233
|
-
},
|
|
234
|
-
"invalidResult": {
|
|
235
|
-
"type": "object",
|
|
236
|
-
"additionalProperties": false,
|
|
237
|
-
"required": ["kind", "schemaVersion", "status", "category", "reason", "targetMutation", "diagnostics"],
|
|
238
|
-
"properties": {
|
|
239
|
-
"kind": { "const": "run-result" },
|
|
240
|
-
"schemaVersion": { "const": 1 },
|
|
241
|
-
"status": { "const": "invalid" },
|
|
242
|
-
"category": { "const": "invalid" },
|
|
243
|
-
"reason": { "const": "invalid-invocation" },
|
|
244
|
-
"targetMutation": { "const": "not-attempted" },
|
|
245
|
-
"diagnostics": {
|
|
246
|
-
"type": "array",
|
|
247
|
-
"minItems": 1,
|
|
248
|
-
"items": { "$ref": "#/$defs/diagnostic" }
|
|
249
|
-
}
|
|
250
|
-
}
|
|
251
|
-
},
|
|
252
|
-
"preExecutionOperationalResult": {
|
|
253
|
-
"type": "object",
|
|
254
|
-
"additionalProperties": false,
|
|
255
|
-
"required": ["kind", "schemaVersion", "source", "status", "category", "reason", "targetMutation", "diagnostics"],
|
|
256
|
-
"properties": {
|
|
257
|
-
"kind": { "const": "run-result" },
|
|
258
|
-
"schemaVersion": { "const": 1 },
|
|
259
|
-
"source": { "const": "parent" },
|
|
260
|
-
"status": { "const": "error" },
|
|
261
|
-
"category": { "const": "infra" },
|
|
262
|
-
"reason": { "const": "pre-execution-error" },
|
|
263
|
-
"targetMutation": { "const": "not-attempted" },
|
|
264
|
-
"diagnostics": {
|
|
265
|
-
"type": "array",
|
|
266
|
-
"minItems": 1,
|
|
267
|
-
"items": { "$ref": "#/$defs/diagnostic" }
|
|
268
|
-
}
|
|
269
|
-
}
|
|
270
|
-
},
|
|
271
|
-
"containmentResult": {
|
|
272
|
-
"type": "object",
|
|
273
|
-
"additionalProperties": false,
|
|
274
|
-
"required": [
|
|
275
|
-
"kind",
|
|
276
|
-
"schemaVersion",
|
|
277
|
-
"source",
|
|
278
|
-
"status",
|
|
279
|
-
"category",
|
|
280
|
-
"reason",
|
|
281
|
-
"executionState",
|
|
282
|
-
"durability",
|
|
283
|
-
"scenarioId",
|
|
284
|
-
"declaredStepIds",
|
|
285
|
-
"targetMutation",
|
|
286
|
-
"startedAt",
|
|
287
|
-
"durationMs",
|
|
288
|
-
"containment",
|
|
289
|
-
"diagnostics"
|
|
290
|
-
],
|
|
291
|
-
"properties": {
|
|
292
|
-
"kind": { "const": "run-result" },
|
|
293
|
-
"schemaVersion": { "const": 1 },
|
|
294
|
-
"source": { "const": "parent" },
|
|
295
|
-
"status": { "const": "contained" },
|
|
296
|
-
"category": { "const": "infra" },
|
|
297
|
-
"reason": { "const": "parent-containment" },
|
|
298
|
-
"executionState": { "const": "unavailable" },
|
|
299
|
-
"durability": { "const": "uncommitted" },
|
|
300
|
-
"runId": { "type": "string", "pattern": "^[0-9a-f]{16}$" },
|
|
301
|
-
"scenarioId": { "type": "string", "minLength": 1 },
|
|
302
|
-
"declaredStepIds": {
|
|
303
|
-
"type": "array",
|
|
304
|
-
"items": { "type": "string", "minLength": 1 }
|
|
305
|
-
},
|
|
306
|
-
"targetMutation": { "enum": ["not-attempted", "possibly-mutated"] },
|
|
307
|
-
"startedAt": { "type": "string", "format": "date-time" },
|
|
308
|
-
"durationMs": { "type": "number", "minimum": 0 },
|
|
309
|
-
"containment": {
|
|
310
|
-
"type": "object",
|
|
311
|
-
"additionalProperties": false,
|
|
312
|
-
"required": ["phase", "trigger"],
|
|
313
|
-
"properties": {
|
|
314
|
-
"phase": { "enum": ["start", "run", "cleanup", "terminal"] },
|
|
315
|
-
"trigger": {
|
|
316
|
-
"enum": [
|
|
317
|
-
"start-frame-too-large",
|
|
318
|
-
"run-deadline-expired",
|
|
319
|
-
"cleanup-deadline-expired",
|
|
320
|
-
"worker-result-unavailable"
|
|
321
|
-
]
|
|
322
|
-
}
|
|
323
|
-
}
|
|
324
|
-
},
|
|
325
|
-
"diagnostics": {
|
|
326
|
-
"type": "array",
|
|
327
|
-
"minItems": 1,
|
|
328
|
-
"items": { "$ref": "#/$defs/diagnostic" }
|
|
329
|
-
}
|
|
330
|
-
}
|
|
331
|
-
}
|
|
332
|
-
}
|
|
333
|
-
}
|
|
@@ -1,86 +0,0 @@
|
|
|
1
|
-
{
|
|
2
|
-
"$schema": "https://json-schema.org/draft/2020-12/schema",
|
|
3
|
-
"$id": "https://checkmate.dev/schemas/validation-result.v1.json",
|
|
4
|
-
"title": "Checkmate validation result version 1",
|
|
5
|
-
"oneOf": [{ "$ref": "#/$defs/valid" }, { "$ref": "#/$defs/notValid" }],
|
|
6
|
-
"$defs": {
|
|
7
|
-
"diagnostic": {
|
|
8
|
-
"type": "object",
|
|
9
|
-
"additionalProperties": false,
|
|
10
|
-
"required": ["code", "path", "message"],
|
|
11
|
-
"properties": {
|
|
12
|
-
"code": { "type": "string", "minLength": 1 },
|
|
13
|
-
"path": { "type": "string", "pattern": "^(?:/(?:[^~/]|~[01])*)*$" },
|
|
14
|
-
"message": { "type": "string", "minLength": 1 }
|
|
15
|
-
}
|
|
16
|
-
},
|
|
17
|
-
"effectiveLimits": {
|
|
18
|
-
"type": "object",
|
|
19
|
-
"additionalProperties": false,
|
|
20
|
-
"required": [
|
|
21
|
-
"scenarioTimeoutMs",
|
|
22
|
-
"stepTimeoutMs",
|
|
23
|
-
"turnsPerStep",
|
|
24
|
-
"requestTimeoutMs",
|
|
25
|
-
"maxRetries",
|
|
26
|
-
"loopMaxRepetitions",
|
|
27
|
-
"cleanupTimeoutMs"
|
|
28
|
-
],
|
|
29
|
-
"properties": {
|
|
30
|
-
"scenarioTimeoutMs": { "type": "integer", "minimum": 1 },
|
|
31
|
-
"stepTimeoutMs": { "type": "integer", "minimum": 1 },
|
|
32
|
-
"turnsPerStep": { "type": "integer", "minimum": 1 },
|
|
33
|
-
"requestTimeoutMs": { "type": "integer", "minimum": 1 },
|
|
34
|
-
"maxRetries": { "type": "integer", "minimum": 0 },
|
|
35
|
-
"loopMaxRepetitions": { "type": "integer", "minimum": 1 },
|
|
36
|
-
"cleanupTimeoutMs": { "type": "integer", "minimum": 1 },
|
|
37
|
-
"budgetTokens": { "type": "integer", "minimum": 1 }
|
|
38
|
-
}
|
|
39
|
-
},
|
|
40
|
-
"valid": {
|
|
41
|
-
"type": "object",
|
|
42
|
-
"additionalProperties": false,
|
|
43
|
-
"required": ["kind", "schemaVersion", "status", "scenarioId", "policy", "driver", "diagnostics"],
|
|
44
|
-
"properties": {
|
|
45
|
-
"kind": { "const": "validation-result" },
|
|
46
|
-
"schemaVersion": { "const": 1 },
|
|
47
|
-
"status": { "const": "valid" },
|
|
48
|
-
"scenarioId": { "type": "string", "minLength": 1 },
|
|
49
|
-
"policy": {
|
|
50
|
-
"type": "object",
|
|
51
|
-
"additionalProperties": false,
|
|
52
|
-
"required": ["id", "effectiveLimits"],
|
|
53
|
-
"properties": {
|
|
54
|
-
"id": { "type": "string", "minLength": 1 },
|
|
55
|
-
"effectiveLimits": { "$ref": "#/$defs/effectiveLimits" }
|
|
56
|
-
}
|
|
57
|
-
},
|
|
58
|
-
"driver": {
|
|
59
|
-
"type": "object",
|
|
60
|
-
"additionalProperties": false,
|
|
61
|
-
"required": ["id", "contractVersion"],
|
|
62
|
-
"properties": {
|
|
63
|
-
"id": { "type": "string", "minLength": 1 },
|
|
64
|
-
"contractVersion": { "const": 1 }
|
|
65
|
-
}
|
|
66
|
-
},
|
|
67
|
-
"diagnostics": { "type": "array", "maxItems": 0 }
|
|
68
|
-
}
|
|
69
|
-
},
|
|
70
|
-
"notValid": {
|
|
71
|
-
"type": "object",
|
|
72
|
-
"additionalProperties": false,
|
|
73
|
-
"required": ["kind", "schemaVersion", "status", "diagnostics"],
|
|
74
|
-
"properties": {
|
|
75
|
-
"kind": { "const": "validation-result" },
|
|
76
|
-
"schemaVersion": { "const": 1 },
|
|
77
|
-
"status": { "enum": ["invalid", "error"] },
|
|
78
|
-
"diagnostics": {
|
|
79
|
-
"type": "array",
|
|
80
|
-
"minItems": 1,
|
|
81
|
-
"items": { "$ref": "#/$defs/diagnostic" }
|
|
82
|
-
}
|
|
83
|
-
}
|
|
84
|
-
}
|
|
85
|
-
}
|
|
86
|
-
}
|