pi-session-mcp 0.8.3 → 0.12.0
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/.pi-session-mcp-cli/client-adapters/command.js +33 -0
- package/.pi-session-mcp-cli/client-adapters/mcp-verifier.js +198 -0
- package/.pi-session-mcp-cli/client-adapters/types.js +1 -0
- package/.pi-session-mcp-cli/config.js +29 -3
- package/.pi-session-mcp-cli/container-runtime.js +18 -0
- package/.pi-session-mcp-cli/doctor.js +370 -4
- package/.pi-session-mcp-cli/execution-modes.js +83 -0
- package/.pi-session-mcp-cli/execution-profile.js +1 -0
- package/.pi-session-mcp-cli/override-lock.js +4 -4
- package/.pi-session-mcp-cli/session-tools.js +45 -0
- package/.pi-session-mcp-cli/setup-platform.js +64 -36
- package/.pi-session-mcp-cli/setup-process.js +455 -0
- package/.pi-session-mcp-cli/skill-library.js +8 -4
- package/.pi-session-mcp-cli/turn-events.js +97 -0
- package/.pi-session-mcp-cli/writer-boundary.js +321 -0
- package/.pi-session-mcp-cli/writer-fence.js +50 -0
- package/.pi-session-mcp-cli/writer-lease.js +140 -0
- package/.pi-session-mcp-cli/writer-lifecycle.js +567 -0
- package/.pi-session-mcp-cli/writer-state.js +199 -0
- package/.pi-session-mcp-cli/writer-wrapper.js +14 -0
- package/CHANGELOG.md +527 -0
- package/README.md +8 -2
- package/dist/capabilities.d.ts +9 -1
- package/dist/capabilities.js +8 -3
- package/dist/child-stdio-transport.d.ts +21 -0
- package/dist/child-stdio-transport.js +88 -0
- package/dist/client-adapters/claude-code.js +3 -2
- package/dist/client-adapters/command.d.ts +8 -0
- package/dist/client-adapters/command.js +9 -1
- package/dist/client-adapters/mcp-verifier.d.ts +7 -0
- package/dist/client-adapters/mcp-verifier.js +21 -4
- package/dist/client-adapters/types.d.ts +5 -1
- package/dist/client-adapters/writer-registration.d.ts +16 -0
- package/dist/client-adapters/writer-registration.js +156 -0
- package/dist/config.d.ts +12 -1
- package/dist/config.js +29 -3
- package/dist/container-runtime.d.ts +2 -0
- package/dist/container-runtime.js +18 -0
- package/dist/doctor.d.ts +36 -2
- package/dist/doctor.js +370 -4
- package/dist/event-spool.d.ts +70 -0
- package/dist/event-spool.js +180 -0
- package/dist/execution-modes.d.ts +60 -0
- package/dist/execution-modes.js +83 -0
- package/dist/execution-profile.d.ts +2 -0
- package/dist/execution-profile.js +1 -0
- package/dist/main.js +63 -8
- package/dist/override-lock.d.ts +2 -0
- package/dist/override-lock.js +4 -4
- package/dist/pi-adapter.d.ts +31 -1
- package/dist/pi-adapter.js +12 -0
- package/dist/process-diagnostics.d.ts +1 -1
- package/dist/process-diagnostics.js +2 -1
- package/dist/refusal-server.d.ts +6 -0
- package/dist/refusal-server.js +46 -0
- package/dist/sdk-pi-adapter.js +16 -9
- package/dist/server.d.ts +3 -2
- package/dist/server.js +27 -53
- package/dist/session-registry.d.ts +32 -2
- package/dist/session-registry.js +61 -11
- package/dist/session-tools.d.ts +14 -5
- package/dist/session-tools.js +13 -9
- package/dist/setup-cli.d.ts +14 -0
- package/dist/setup-cli.js +132 -0
- package/dist/setup-command-guardian.js +9 -3
- package/dist/setup-hooks.d.ts +158 -0
- package/dist/setup-hooks.js +657 -0
- package/dist/setup-main.d.ts +1 -1
- package/dist/setup-main.js +18 -2
- package/dist/setup-modes.d.ts +18 -0
- package/dist/setup-modes.js +131 -0
- package/dist/setup-ownership.d.ts +1 -0
- package/dist/setup-ownership.js +11 -3
- package/dist/setup-platform.d.ts +9 -1
- package/dist/setup-platform.js +64 -36
- package/dist/setup-process.d.ts +23 -1
- package/dist/setup-process.js +88 -10
- package/dist/setup-writers.d.ts +42 -0
- package/dist/setup-writers.js +510 -0
- package/dist/setup.d.ts +4 -1
- package/dist/setup.js +23 -10
- package/dist/skill-library.d.ts +2 -0
- package/dist/skill-library.js +8 -4
- package/dist/tool-schemas.d.ts +339 -0
- package/dist/tool-schemas.js +49 -0
- package/dist/turn-events.d.ts +45 -0
- package/dist/turn-events.js +97 -0
- package/dist/wake-hook.d.ts +180 -0
- package/dist/wake-hook.js +638 -0
- package/dist/wake-main.d.ts +8 -0
- package/dist/wake-main.js +111 -0
- package/dist/writer-acceptance.d.ts +10 -0
- package/dist/writer-acceptance.js +192 -0
- package/dist/writer-boundary.d.ts +52 -0
- package/dist/writer-boundary.js +321 -0
- package/dist/writer-credentials.d.ts +29 -0
- package/dist/writer-credentials.js +85 -0
- package/dist/writer-fence.d.ts +12 -0
- package/dist/writer-fence.js +50 -0
- package/dist/writer-guard.d.ts +13 -0
- package/dist/writer-guard.js +54 -0
- package/dist/writer-launch.d.ts +29 -0
- package/dist/writer-launch.js +232 -0
- package/dist/writer-lease.d.ts +28 -0
- package/dist/writer-lease.js +140 -0
- package/dist/writer-lifecycle.d.ts +96 -0
- package/dist/writer-lifecycle.js +567 -0
- package/dist/writer-relay.d.ts +33 -0
- package/dist/writer-relay.js +359 -0
- package/dist/writer-state.d.ts +111 -0
- package/dist/writer-state.js +199 -0
- package/dist/writer-wrapper.d.ts +3 -0
- package/dist/writer-wrapper.js +14 -0
- package/docs/acceptance.md +1694 -1
- package/docs/adr/0003-post-mvp-maintenance.md +2 -2
- package/docs/adr/0006-pi-extension-allowlist.md +3 -1
- package/docs/adr/0008-coding-profile-isolation.md +295 -0
- package/docs/adr/0009-execution-modes.md +131 -0
- package/docs/adr/0010-read-only-extension-tools.md +68 -0
- package/docs/adr/0011-turn-events-and-controller-wake.md +173 -0
- package/docs/adr/0012-leased-writer-pool.md +357 -0
- package/docs/architecture.md +96 -8
- package/docs/client-setup-design.md +12 -8
- package/docs/client-setup.md +84 -2
- package/docs/doctor-design.md +48 -0
- package/docs/evaluation/2026-10-17-live-screening.md +121 -0
- package/docs/evaluation/2026-10-17-node-runtime-followup.md +33 -0
- package/docs/evaluation/2026-10-17-profile-evaluation-evidence.md +155 -0
- package/docs/evaluation/2026-10-17-test-coverage.md +127 -0
- package/docs/evaluation/README.md +123 -0
- package/docs/evaluation/client-reconnect-2026-10-17/README.md +41 -0
- package/docs/evaluation/client-reconnect-2026-10-17/checks.json +38 -0
- package/docs/evaluation/client-reconnect-2026-10-17/integrity.json +30 -0
- package/docs/evaluation/client-reconnect-2026-10-17/reconnect.json +57 -0
- package/docs/evaluation/client-reconnect-2026-10-17/registration-verification.json +389 -0
- package/docs/evaluation/client-reconnect-2026-10-17/review.json +10 -0
- package/docs/evaluation/controller-ledger-hooks-2026-10-26/README.md +115 -0
- package/docs/evaluation/evidence-2026-10-17/budget.proposed.json +13 -0
- package/docs/evaluation/evidence-2026-10-17/changed-files.txt +82 -0
- package/docs/evaluation/evidence-2026-10-17/checks.json +103 -0
- package/docs/evaluation/evidence-2026-10-17/comparison.json +247 -0
- package/docs/evaluation/evidence-2026-10-17/development-collection-failure.json +15 -0
- package/docs/evaluation/evidence-2026-10-17/full-attempt-default.json +19 -0
- package/docs/evaluation/evidence-2026-10-17/full-tests.json +17 -0
- package/docs/evaluation/evidence-2026-10-17/inventory.json +239 -0
- package/docs/evaluation/evidence-2026-10-17/manifest.json +359 -0
- package/docs/evaluation/evidence-2026-10-17/offline.json +1020 -0
- package/docs/evaluation/evidence-2026-10-17/ui-reference/T3-1280.png +0 -0
- package/docs/evaluation/evidence-2026-10-17/ui-reference/T3-375.png +0 -0
- package/docs/evaluation/evidence-2026-10-17/ui-reference/result.json +141 -0
- package/docs/evaluation/extension-ablation-2026-10-20/README.md +414 -0
- package/docs/evaluation/extension-ablation-2026-10-20/setup.md +140 -0
- package/docs/evaluation/extension-ablation-repeat-2026-10-21/README.md +142 -0
- package/docs/evaluation/extension-ablation-repeat-2026-10-21/setup.md +69 -0
- package/docs/evaluation/extension-audit-2026-10-17/README.md +149 -0
- package/docs/evaluation/extension-audit-2026-10-17/checks.json +38 -0
- package/docs/evaluation/extension-audit-2026-10-17/cursor-static-evidence.json +93 -0
- package/docs/evaluation/extension-audit-2026-10-17/installed-artifact-comparison.json +85 -0
- package/docs/evaluation/extension-audit-2026-10-17/installed-packages.json +100 -0
- package/docs/evaluation/extension-audit-2026-10-17/integrity.json +108 -0
- package/docs/evaluation/extension-audit-2026-10-17/navigation-cursor-registry.json +290 -0
- package/docs/evaluation/extension-audit-2026-10-17/navigation-static-evidence.json +185 -0
- package/docs/evaluation/extension-audit-2026-10-17/openwiki-headless-bindings.json +25 -0
- package/docs/evaluation/extension-audit-2026-10-17/openwiki-headless-cleanup.json +12 -0
- package/docs/evaluation/extension-audit-2026-10-17/openwiki-headless-probe.mjs +63 -0
- package/docs/evaluation/extension-audit-2026-10-17/openwiki-headless-reproduction.json +91 -0
- package/docs/evaluation/extension-audit-2026-10-17/openwiki-headless-result.json +83 -0
- package/docs/evaluation/extension-audit-2026-10-17/permission-cache-artifact-verification.json +24 -0
- package/docs/evaluation/extension-audit-2026-10-17/permission-cache-evidence.json +133 -0
- package/docs/evaluation/extension-audit-2026-10-17/permission-cache-report.md +64 -0
- package/docs/evaluation/extension-audit-2026-10-17/publication-export.json +10 -0
- package/docs/evaluation/extension-audit-2026-10-17/resolver-gates.json +51 -0
- package/docs/evaluation/extension-audit-2026-10-17/review.json +28 -0
- package/docs/evaluation/extension-audit-2026-10-17/run-openwiki-headless-probe.mjs +29 -0
- package/docs/evaluation/extension-audit-2026-10-17/summary.json +123 -0
- package/docs/evaluation/extension-audit-2026-10-17/web-openwiki-artifacts.json +29 -0
- package/docs/evaluation/extension-audit-2026-10-17/web-openwiki-evidence.json +116 -0
- package/docs/evaluation/extension-audit-2026-10-17/web-openwiki-report.md +113 -0
- package/docs/evaluation/extension-functions-2026-10-17/README.md +33 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/README.md +39 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/attempts/extension-functions-bridges-03dc1dad-9959-40df-b51c-7820991bf7e3.json +79 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/attempts/extension-functions-bridges-0c43651e-8c7d-4c7c-aa9b-5b5761141e3b.json +54 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/attempts/extension-functions-bridges-51247b13-87c1-4ec6-886b-82d922b0585a.json +89 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/attempts/extension-functions-bridges-5337f6bf-503f-4c45-aac0-43b142385a87.json +47 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/attempts/extension-functions-bridges-60285194-b1d1-4335-b616-6614941585df.json +46 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/attempts/extension-functions-bridges-6d2fbaa9-a720-4582-9be1-a95f4764786d.json +69 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/attempts/extension-functions-bridges-6fbb576c-d87d-478d-9c05-4b5ef3563598.json +125 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/attempts/extension-functions-bridges-730926d1-ef20-4b2d-ab74-7e542f1065e9.json +48 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/attempts/extension-functions-bridges-a1baa47a-5bbe-4a16-ade8-a350599eb911.json +48 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/attempts/extension-functions-bridges-bc1d69c7-784f-4ddb-9335-0d905330d568.json +91 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/attempts/extension-functions-bridges-d582757d-2859-4683-8310-ec158bf6d551.json +99 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/attempts/extension-functions-bridges-e6533a23-4933-4d6e-a250-b1b27f010d18.json +60 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/attempts.json +138 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/guard.mjs +74 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/integrity.json +86 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/openwiki-final.json +125 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/run.mjs +36 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/web-final.json +99 -0
- package/docs/evaluation/extension-functions-2026-10-17/bridges/worker.mjs +102 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/README.md +21 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/cache-function-evidence-2026-10-17T13-30-31-440Z-b1a47f88-8d86-4238-9ef8-af890d9ba74b.json +129 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/cache-function-evidence-2026-10-17T13-31-04-505Z-a3736add-73b8-4d77-8ea6-084ce8aa4599.json +129 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/cache-function-evidence-2026-10-17T13-31-15-654Z-b0197bff-8363-4d46-b84b-46d9ed67fc39.json +248 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/cache-function-evidence-2026-10-17T13-31-39-137Z-e373f24d-8a8c-46b0-86aa-9f687cfadb9b.json +278 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/cache-function-evidence-2026-10-17T13-32-05-324Z-f63095f1-755e-43fc-9d28-3cd1a9ff4f4d.json +251 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/cache-function-evidence-2026-10-17T13-32-25-191Z-fc9a83af-ceae-40d7-a737-8ef861ba7946.json +282 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/cache-function-evidence-2026-10-17T13-33-14-993Z-5e7eae34-09f0-4c80-9210-b8cc4d1acb5f.json +282 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/cache-function-evidence-2026-10-17T13-38-39-010Z-57cf6a26-f091-45de-8849-b616b303a46c.json +283 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/cache-function-evidence-2026-10-17T13-40-32-118Z-54528e43-ac07-4f9c-8f12-10c041130983.json +283 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/cache-function-evidence-2026-10-17T13-41-49-525Z-48195b6e-f20d-49a5-bd53-a732a0af10c1.json +324 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/cache-function-evidence-2026-10-17T13-42-18-175Z-215330e7-06a6-4666-af9c-2ecb214d8d71.json +324 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/integrity.json +165 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/probe.mjs +161 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipt-archive.json +150 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/commands-headless-3feeffc5-97bf-4079-8505-8119fab7956a.json +54 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/commands-headless-42e5eace-ad87-480a-94d9-448c1dd167b8.json +49 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/commands-headless-50f39f1c-8b48-442b-8d15-1f9c8f63bd81.json +49 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/commands-headless-8482b2b0-f600-488b-9f36-905e0125177a.json +54 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/payload-codex-bypass-128ffdce-6fd5-431b-8019-32b40ba4174d.json +42 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/payload-codex-bypass-3b572182-a737-471f-b9cb-b01d5da1752a.json +42 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/payload-codex-bypass-6c225370-2f47-4a0a-80d5-87d12f6c6fdc.json +37 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/payload-codex-bypass-b48399db-cbed-43de-9071-ac27c34e3757.json +37 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/payload-openai-3a8f52b6-c8d6-4d1c-844c-782a6624a29e.json +45 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/payload-openai-6fcdbc20-b14b-4b80-a4da-bf24078da8d2.json +50 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/payload-openai-a8d6f628-40e1-437e-b192-f7232a051a44.json +50 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/payload-openai-c3c39463-a2d6-4ff8-b985-8f6d1fb1aa02.json +45 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/stats-synthetic-a23c264f-151d-4ff0-aab4-bdbb2f6560d8.json +53 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/stats-synthetic-d8178bd3-6530-4915-9cf1-056a975bfbcd.json +53 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/stats-synthetic-e3985119-e101-459e-91cd-21a0905287ba.json +48 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/stats-synthetic-e7a40382-2d27-4db5-a133-dbe346a5dc60.json +48 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/warming-disabled-unknown-223d7adc-2dbc-4017-b1ec-0c38294320e2.json +41 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/warming-disabled-unknown-347a8548-d2e0-4846-a08e-b0399f8fd3b4.json +41 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/warming-disabled-unknown-49b45e2b-e2d2-456f-b73f-e6bc938b6378.json +46 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/warming-disabled-unknown-aaf432f2-bb91-41a7-b951-2d4316389632.json +46 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/warming-disabled-zero-204ab3e8-f1fe-4894-8fb0-08299deb4c21.json +41 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/warming-disabled-zero-9388d2df-fa15-479c-a4ef-2a774c366c0f.json +46 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/warming-disabled-zero-e0bb4742-214a-4b3f-88d2-44de54394e2f.json +46 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/receipts/warming-disabled-zero-f77e1e2d-742a-490f-9955-1dfb19f59c9c.json +41 -0
- package/docs/evaluation/extension-functions-2026-10-17/cache/runner.mjs +36 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/README.md +20 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-36-10-753Z-2bb4fff5-48cc-4f6b-ae1b-5cc1a07f610b.json +54 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-36-26-537Z-5adc4fb1-0008-4ef9-a106-aa372dabf061.json +56 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-36-32-362Z-30949c31-73a3-4dae-af81-68deded51543.json +56 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-36-38-494Z-8eec81b8-5576-48cb-b2fa-c279269db058.json +56 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-36-45-546Z-991647e5-7ea4-43c4-a90d-868e4ac38d83.json +56 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-37-02-277Z-2296fb35-e5cb-4bf1-9e83-7370bcb46bfe.json +56 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-37-28-390Z-e9c3d106-b4c9-4146-ae83-a2799c907fe5.json +56 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-37-37-371Z-8545c893-a8e3-4dc6-973b-81dfd97d798b.json +56 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-38-41-612Z-c57c4778-261d-4397-8a36-df9cc23ca629.json +58 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-38-51-661Z-0eb898de-7471-4f17-a308-5ef5d86421ff.json +58 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-38-55-412Z-8169d843-a9e7-48f4-88b1-3a90e1592df1.json +58 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-39-15-274Z-a5610333-e428-4cc6-9f5a-9206748ee512.json +58 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-39-39-506Z-d4c80a7b-76b4-4d11-bf0f-552d11b131b2.json +136 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-39-53-515Z-db409f47-8ee5-4354-b7f9-a258eb7e17d0.json +136 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/evidence-2026-10-17T13-42-04-909Z-01d70cb3-6897-46ca-b8aa-1c69d5b5c7fc.json +148 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/integrity.json +205 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/probe.mjs +151 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipt-archive.json +186 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-104e378f-8f82-4ae9-bffc-7f1d52e00e74.json +15 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-182a8009-c52a-4760-9768-9cba15fceb9c.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-1f07360d-046f-4f0c-b40f-6b57a027220c.json +65 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-256cb86e-40ae-4fec-a2d2-9b7f66cc4be8.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-2667d991-3133-44c3-b883-c401b3c14cb1.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-394fc425-ed51-408f-ab09-6d26aee0d48f.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-531d3f1d-c326-4914-b768-3d32103513b9.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-5c7d0d62-8290-494f-8729-5d60bc626ad7.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-64af1d94-ea73-481f-b827-dbf6755d8591.json +66 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-65fce696-864c-4247-960e-59f6d671cbf9.json +15 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-66acde31-19c7-4d3e-9814-ddeb2cffd373.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-79d220d0-fc0b-4cad-85cc-40b9e904fb97.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-a58f0281-71bc-4757-9d6f-e281f813135e.json +65 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-c2b61e4b-fd33-4969-ab29-118957ca5fbb.json +15 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/activation-and-resources-f61b5020-2c05-40ad-9cb8-590b22790186.json +15 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-0ee0df5a-299f-451c-96e3-5abcf8bfa999.json +45 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-169e5146-09a2-4231-8017-ebd8d4b75f0a.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-19d479f2-7162-463e-90de-2d89a13e59a3.json +15 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-1bd882bb-2403-4269-8335-90f9803f57b7.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-2ab9d418-68b2-4679-bf42-f31b6945ea5a.json +15 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-3a6aa7d7-b956-4fe5-9555-a2c63408c067.json +45 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-6333425f-0b46-4ced-a272-3d48895ee853.json +15 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-6eccb170-ff45-40e2-afc5-dfc0f77c229f.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-7f330c57-9c3f-4879-a1d0-67c4c5c36050.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-8d9de272-165c-4bb8-a147-d1b306388472.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-c1ffa69f-25fa-4f45-a691-af3bc422c089.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-c24ad0c8-5e7e-4d71-9b22-3e2708ce1626.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-c5a3140e-a28e-4264-9ee6-51b1310b3123.json +14 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-cdbb3dde-2441-46ae-9111-f2a299d45807.json +46 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/receipts/bridge-disabled-d68f774c-d1f0-4620-b334-301a20edffc3.json +15 -0
- package/docs/evaluation/extension-functions-2026-10-17/cursor/runner.mjs +147 -0
- package/docs/evaluation/extension-functions-2026-10-17/integrity.json +618 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/README.md +22 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/SHA256SUMS.txt +8 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempt.json +13 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-30-37-626Z-7ba0ce60-b805-4bf1-813f-80f0673c519d.json +50 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-31-06-997Z-0df6e689-b7f6-4daf-a730-724638c7e9bb.json +52 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-31-22-489Z-dfa50911-d117-400d-9ea3-76b224693c10.json +59 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-31-39-899Z-6792c0ef-c795-4136-98e2-b9a9b00489bd.json +59 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-31-58-555Z-e23d44fb-d2cb-42b6-b76b-f2ac53645401.json +59 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-32-18-734Z-3708117d-bca6-44a6-9f61-4339a67535ae.json +60 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-32-34-897Z-abbfb675-7120-46fc-8638-2dde943a3c51.json +114 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-33-30-753Z-f1965485-aa41-4a9a-b62e-267fc9d7686c.json +72 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-33-46-589Z-e3b46364-e0dd-4a95-91ce-c11f5192d83f.json +72 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-33-57-569Z-d296148e-2f64-4921-bad7-391d0f3cc61d.json +74 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-34-30-540Z-51791e8c-b2a7-4f93-bff6-9b0467aeba51.json +109 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-34-57-906Z-2c3e469c-6d89-4707-8bc6-e62dcfacdc7c.json +114 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-35-16-679Z-035260fc-f3d1-4c1c-9619-81f2aa2ced55.json +115 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-36-05-641Z-24ceb58d-441b-4ffe-9853-ef4d15eb8c81.json +73 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-37-00-445Z-1ca606b7-306f-41e9-8745-b53879c620bb.json +72 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-37-23-659Z-c958d5fc-5c17-46b9-abc0-87913f14fd4b.json +73 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-38-49-639Z-9e74ee3c-7cea-4d9e-832f-0b5068d161ec.json +112 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-39-53-463Z-5d834e9f-e247-4887-8200-dee41070567f.json +32 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-40-16-811Z-0be9d64c-a0d7-405c-b81a-12f8c6cfebac.json +33 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-40-36-020Z-412777d5-f352-4fa0-9a07-f017b5f04cbb.json +124 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-42-03-506Z-a0cca711-1300-4431-bfff-5814a409d0af.json +120 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-42-28-426Z-fb70da7e-4888-469a-b5cd-b74da52afa9f.json +112 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-42-41-596Z-c8e99863-70e2-40f1-a516-e9f41156ef44.json +71 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-44-06-279Z-39516e45-daf6-409b-b216-64179608cc2e.json +97 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-45-34-538Z-87ddef90-2fda-4211-a513-0516dec235f4.json +112 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-45-48-020Z-06f478f2-43c4-48a7-ac02-7e53d4df8aaf.json +112 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-46-00-807Z-a98cc49d-6cf6-4109-aa95-f2d36e8d35c3.json +97 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-46-34-183Z-77f75abc-9ec3-41b2-896f-b6f4a180dca8.json +160 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/attempts/isolated-extensions-results-2026-10-17T13-46-56-605Z-f46134c2-8523-4892-ab76-e88195d79ed2.json +163 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/child-guard.mjs +30 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/final-result.json +42 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/function-probe.mjs +521 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/function-runner.mjs +201 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/integrity.json +153 -0
- package/docs/evaluation/extension-functions-2026-10-17/navigation/receipt-archive.json +160 -0
- package/docs/evaluation/extension-functions-2026-10-17/review.json +27 -0
- package/docs/evaluation/extension-functions-2026-10-17/summary.json +102 -0
- package/docs/evaluation/extension-functions-2026-10-17/verification.json +103 -0
- package/docs/evaluation/isolation-acceptance-2026-10-18/README.md +106 -0
- package/docs/evaluation/isolation-acceptance-2026-10-18/auth-warmup.mjs +33 -0
- package/docs/evaluation/isolation-acceptance-2026-10-18/contained-agent-ro.json +433 -0
- package/docs/evaluation/isolation-acceptance-2026-10-18/contained.json +439 -0
- package/docs/evaluation/isolation-acceptance-2026-10-18/control.json +403 -0
- package/docs/evaluation/isolation-acceptance-2026-10-18/live.json +75 -0
- package/docs/evaluation/isolation-acceptance-2026-10-18/wrapper.sh +37 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/README.md +360 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/design.md +150 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/inventory.md +113 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/live/README.md +117 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/live/dry-run/dry-01-L1-A.json +202 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/live/dry-run/dry-02-L1-C.json +259 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/live/run/01-L1-A.json +169 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/live/run/02-L1-A.json +202 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/live/run/03-L1-C.json +256 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/live/run/approval.json +31 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/live/run-2/01-L1-C.json +256 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/live/run-2/02-L1-A.json +201 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/live/run-2/approval.json +16 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/live-plan.md +87 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/pre-registration.md +117 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-1/S1.json +361 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-1/S2.json +544 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-1/S3.json +470 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-1/S4.json +271 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-1/S5.json +271 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-1/S6.json +266 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-1/X6.json +29 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-1/summary.json +269 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-2/S1.json +391 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-2/S2.json +600 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-2/S3.json +519 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-2/S4.json +281 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-2/S5.json +294 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-2/S6.json +280 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-2/S8.json +280 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-2/X6.json +32 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-2/summary.json +307 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-3/S1.json +391 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-3/S2.json +600 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-3/S3.json +519 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-3/S4.json +281 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-3/S5.json +294 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-3/S6.json +280 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-3/S8.json +280 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-3/X6.json +32 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-3/summary.json +307 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-4/S1.json +409 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-4/S2.json +634 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-4/S3.json +537 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-4/S4.json +281 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-4/S5.json +294 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-4/S6.json +280 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-4/S8.json +280 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-4/X6.json +32 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-4/summary.json +307 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-5/S1.json +409 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-5/S2.json +634 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-5/S3.json +537 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-5/S4.json +281 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-5/S5.json +294 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-5/S6.json +280 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-5/S8.json +280 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-5/X6.json +32 -0
- package/docs/evaluation/ledger-event-bridge-2026-10-21/probe/attempt-5/summary.json +307 -0
- package/docs/evaluation/live-rescreen-2026-10-17/README.md +114 -0
- package/docs/evaluation/live-rescreen-2026-10-17/approval.json +18 -0
- package/docs/evaluation/live-rescreen-2026-10-17/approved-budget.json +13 -0
- package/docs/evaluation/live-rescreen-2026-10-17/approved-manifest.json +315 -0
- package/docs/evaluation/live-rescreen-2026-10-17/audit-rescreen.mjs +78 -0
- package/docs/evaluation/live-rescreen-2026-10-17/audit.json +152 -0
- package/docs/evaluation/live-rescreen-2026-10-17/backend-independent-checks.json +38 -0
- package/docs/evaluation/live-rescreen-2026-10-17/checks.json +54 -0
- package/docs/evaluation/live-rescreen-2026-10-17/execution.json +19 -0
- package/docs/evaluation/live-rescreen-2026-10-17/integrity.json +138 -0
- package/docs/evaluation/live-rescreen-2026-10-17/planning-ledger.json +26 -0
- package/docs/evaluation/live-rescreen-2026-10-17/review.json +16 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/comparison.json +177 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-1.claim +1 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-1.diagnostic.json +1 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-1.json +97 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-2.claim +1 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-2.json +104 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-2.patch +11 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-2.snapshot/contract.mjs +4 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-2.snapshot/package.json +1 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-3.claim +1 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-3.json +89 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-4.claim +1 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-4.json +104 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-4.patch +11 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-4.snapshot/contract.mjs +4 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-4.snapshot/package.json +1 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-5.claim +1 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-5.diagnostic.json +1 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-5.json +117 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runs/rescreen-6.json +53 -0
- package/docs/evaluation/live-rescreen-2026-10-17/runtime-observation.json +11 -0
- package/docs/evaluation/live-screening-2026-10-17/approval.json +13 -0
- package/docs/evaluation/live-screening-2026-10-17/approved-budget.json +13 -0
- package/docs/evaluation/live-screening-2026-10-17/approved-manifest.json +359 -0
- package/docs/evaluation/live-screening-2026-10-17/audit.json +143 -0
- package/docs/evaluation/live-screening-2026-10-17/integrity.json +148 -0
- package/docs/evaluation/live-screening-2026-10-17/inventory.json +239 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/comparison.json +247 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-1.claim +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-1.diagnostic.json +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-1.json +98 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-2.claim +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-2.json +104 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-2.patch +11 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-2.snapshot/contract.mjs +4 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-2.snapshot/package.json +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-3.claim +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-3.diagnostic.json +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-3.json +97 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-4.claim +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-4.json +104 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-4.patch +11 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-4.snapshot/contract.mjs +4 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-4.snapshot/package.json +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-5.claim +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-5.diagnostic.json +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-5.json +97 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-6.claim +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-6.json +104 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-6.patch +11 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-6.snapshot/contract.mjs +4 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-6.snapshot/package.json +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-7.claim +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-7.json +108 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-8.claim +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-8.diagnostic.json +1 -0
- package/docs/evaluation/live-screening-2026-10-17/runs/screen-8.json +116 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/README.md +124 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/ab-proposal.json +135 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-1/README.md +25 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-1/approved-budget.json +13 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-1/approved-decision.json +79 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-1/approved-manifest.json +271 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-1/token-live-result.json +546 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-2/README.md +22 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-2/approved-budget.json +13 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-2/approved-decision.json +80 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-2/approved-manifest.json +271 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-2/token-live-result.json +1663 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-3/README.md +23 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-3/approved-budget.json +13 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-3/approved-decision.json +71 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-3/approved-manifest.json +271 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-3/token-live-result.json +1748 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-4/README.md +25 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-4/arm-A/approved-budget.json +13 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-4/arm-A/approved-decision.json +71 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-4/arm-A/approved-manifest.json +271 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-4/arm-A/token-live-result.json +3019 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-4/arm-B/approved-budget.json +13 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-4/arm-B/approved-decision.json +71 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-4/arm-B/approved-manifest.json +271 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/probe/attempt-4/arm-B/token-live-result.json +2859 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/README.md +67 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r1-pair1-A/approved-budget.json +13 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r1-pair1-A/approved-decision.json +71 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r1-pair1-A/approved-manifest.json +271 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r1-pair1-A/token-live-result.json +3019 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r2-pair1-B/approved-budget.json +13 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r2-pair1-B/approved-decision.json +71 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r2-pair1-B/approved-manifest.json +271 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r2-pair1-B/token-live-result.json +2981 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r3-pair2-B/approved-budget.json +13 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r3-pair2-B/approved-decision.json +71 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r3-pair2-B/approved-manifest.json +271 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r3-pair2-B/token-live-result.json +2887 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r4-pair2-A/approved-budget.json +13 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r4-pair2-A/approved-decision.json +71 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r4-pair2-A/approved-manifest.json +271 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r4-pair2-A/token-live-result.json +3189 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r5-pair3-B/approved-budget.json +13 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r5-pair3-B/approved-decision.json +71 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r5-pair3-B/approved-manifest.json +271 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r5-pair3-B/token-live-result.json +2929 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r6-pair3-A/approved-budget.json +13 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r6-pair3-A/approved-decision.json +71 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r6-pair3-A/approved-manifest.json +271 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r6-pair3-A/token-live-result.json +3085 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r7-pair4-A/approved-budget.json +13 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r7-pair4-A/approved-decision.json +71 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r7-pair4-A/approved-manifest.json +271 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r7-pair4-A/token-live-result.json +3019 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r8-pair4-B/approved-budget.json +13 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r8-pair4-B/approved-decision.json +71 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r8-pair4-B/approved-manifest.json +271 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/r8-pair4-B/token-live-result.json +2925 -0
- package/docs/evaluation/m1-ab-proposal-2026-10-18/results/summary-rows.json +338 -0
- package/docs/evaluation/node24-followup-2026-10-17/README.md +55 -0
- package/docs/evaluation/node24-followup-2026-10-17/client-migration-review.json +20 -0
- package/docs/evaluation/node24-followup-2026-10-17/client-registration-migration.json +31 -0
- package/docs/evaluation/node24-followup-2026-10-17/client-registration-verification.json +389 -0
- package/docs/evaluation/node24-followup-2026-10-17/closure-checks.json +33 -0
- package/docs/evaluation/node24-followup-2026-10-17/cursor-headless-probe.mjs +259 -0
- package/docs/evaluation/node24-followup-2026-10-17/cursor-headless-report.md +39 -0
- package/docs/evaluation/node24-followup-2026-10-17/cursor-headless-result-attempt-2.json +49 -0
- package/docs/evaluation/node24-followup-2026-10-17/cursor-headless-result-attempt-3.json +76 -0
- package/docs/evaluation/node24-followup-2026-10-17/cursor-headless-result-attempt-4.json +108 -0
- package/docs/evaluation/node24-followup-2026-10-17/cursor-headless-result.json +12 -0
- package/docs/evaluation/node24-followup-2026-10-17/cursor-headless-run.mjs +96 -0
- package/docs/evaluation/node24-followup-2026-10-17/download.json +16 -0
- package/docs/evaluation/node24-followup-2026-10-17/extension-summary.json +97 -0
- package/docs/evaluation/node24-followup-2026-10-17/final-review.json +15 -0
- package/docs/evaluation/node24-followup-2026-10-17/install.json +10 -0
- package/docs/evaluation/node24-followup-2026-10-17/integrity.json +204 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-evidence.json +39 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-probe.mjs +277 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-report.md +35 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-19-48-378Z-b378dc9f-9ab8-4349-8892-78af8184f78f.json +151 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-20-22-821Z-85cfcba4-c05e-4c5a-9658-d8d70f5c8d1a.json +223 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-20-54-825Z-fd5d8373-6e29-4197-84bd-2e65251b4db8.json +207 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-21-44-636Z-488698ad-f8bf-4d7e-b85a-615275c32e5d.json +82 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-22-05-375Z-16f5d77c-32a2-4d0a-905a-e080b392316f.json +94 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-22-07-471Z-fcbe7c98-fed3-4120-a860-b2e3f9572836.json +63 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-22-17-638Z-45ec386d-3429-4a7c-a10e-8673c03861f0.json +63 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-22-22-485Z-1b08e134-667c-4416-8372-cb233939c48b.json +125 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-22-25-490Z-223fd05e-52f9-4996-88d7-3ec43a6ecdd9.json +125 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-23-42-291Z-f4af1e12-844e-4b2d-a090-9068e7626dd5.json +82 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-24-36-866Z-30770d13-c7ae-4309-921b-5c1b43050adb.json +52 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-25-05-378Z-2a5dffc1-3f3a-433a-823f-6e77964dcaee.json +53 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-25-15-842Z-798946f5-b9a5-4810-8a0c-35be6f40abb3.json +83 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-results-2026-10-17T09-25-19-099Z-89bd000d-bb51-4e3a-b713-51ff1ff4e38b.json +87 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-extensions-runner.mjs +172 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-install.json +121 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-package-lock.json +2321 -0
- package/docs/evaluation/node24-followup-2026-10-17/isolated-package.json +12 -0
- package/docs/evaluation/node24-followup-2026-10-17/launcher-review.json +17 -0
- package/docs/evaluation/node24-followup-2026-10-17/launcher-switch.json +28 -0
- package/docs/evaluation/node24-followup-2026-10-17/mcp-launcher-test.json +24 -0
- package/docs/evaluation/node24-followup-2026-10-17/migrate-client-node.mjs +129 -0
- package/docs/evaluation/node24-followup-2026-10-17/node-runtime-report.md +127 -0
- package/docs/evaluation/node24-followup-2026-10-17/node24-checks.json +44 -0
- package/docs/evaluation/node24-followup-2026-10-17/pi-launcher-test.json +13 -0
- package/docs/evaluation/node24-followup-2026-10-17/publication-export.json +11 -0
- package/docs/evaluation/node24-followup-2026-10-17/restore-launchers.mjs +26 -0
- package/docs/evaluation/node24-followup-2026-10-17/verify-client-node.mjs +36 -0
- package/docs/evaluation/permission-gates-2026-10-17/README.md +23 -0
- package/docs/evaluation/permission-gates-2026-10-17/attempts.json +53 -0
- package/docs/evaluation/permission-gates-2026-10-17/integrity.json +46 -0
- package/docs/evaluation/permission-gates-2026-10-17/permission-gate-probe.mjs +132 -0
- package/docs/evaluation/permission-gates-2026-10-17/permission-gate-worker.mjs +209 -0
- package/docs/evaluation/permission-gates-2026-10-17/permission-gates-result-final.json +53 -0
- package/docs/evaluation/permission-gates-2026-10-17/permission-gates-result-verified.json +54 -0
- package/docs/evaluation/permission-gates-2026-10-17/permission-gates-result.json +53 -0
- package/docs/evaluation/permission-gates-2026-10-17/review.json +19 -0
- package/docs/evaluation/permission-gates-2026-10-17/superseded-initial-receipt.json +40 -0
- package/docs/evaluation/pi-lens-registration-2026-10-19/README.md +90 -0
- package/docs/evaluation/pi-lens-registration-2026-10-19/probe.mjs +126 -0
- package/docs/evaluation/pi-lens-registration-2026-10-19/result.json +1 -0
- package/docs/evaluation/plan-continuation-2026-10-17/README.md +107 -0
- package/docs/evaluation/plan-continuation-2026-10-17/budget.proposed.json +13 -0
- package/docs/evaluation/plan-continuation-2026-10-17/checks.json +1048 -0
- package/docs/evaluation/plan-continuation-2026-10-17/integrity.json +58 -0
- package/docs/evaluation/plan-continuation-2026-10-17/inventory.json +239 -0
- package/docs/evaluation/plan-continuation-2026-10-17/manifest.proposed.json +315 -0
- package/docs/evaluation/plan-continuation-2026-10-17/next-stage-ledger.json +19 -0
- package/docs/evaluation/plan-continuation-2026-10-17/prepare-rescreen.mjs +30 -0
- package/docs/evaluation/plan-continuation-2026-10-17/prior-integrity.json +25 -0
- package/docs/evaluation/plan-continuation-2026-10-17/proposal-validation.json +14 -0
- package/docs/evaluation/plan-continuation-2026-10-17/receipt-contract-update.patch +40 -0
- package/docs/evaluation/plan-continuation-2026-10-17/review.json +8 -0
- package/docs/evaluation/plan-continuation-2026-10-17/source-delta.json +14 -0
- package/docs/evaluation/profile-evaluation.md +297 -0
- package/docs/evaluation/publication-2026-10-17/README.md +88 -0
- package/docs/evaluation/publication-2026-10-17/checks.json +117 -0
- package/docs/evaluation/publication-2026-10-17/integrity.json +21 -0
- package/docs/evaluation/publication-2026-10-17/publication-scan.json +63 -0
- package/docs/evaluation/publication-2026-10-17/review.json +21 -0
- package/docs/evaluation/s5-decision-inputs-2026-10-18/README.md +16 -0
- package/docs/evaluation/s5-decision-inputs-2026-10-18/bounded-wait.md +78 -0
- package/docs/evaluation/s5-decision-inputs-2026-10-18/claude-projection/README.md +151 -0
- package/docs/evaluation/s5-decision-inputs-2026-10-18/claude-projection/attempt-1/claude-projection.json +126 -0
- package/docs/evaluation/s5-decision-inputs-2026-10-18/claude-projection/attempt-2/claude-projection.json +1895 -0
- package/docs/evaluation/s5-decision-inputs-2026-10-18/claude-projection/attempt-3/claude-projection.json +1895 -0
- package/docs/evaluation/s5-decision-inputs-2026-10-18/claude-projection/attempt-4/claude-projection.json +1797 -0
- package/docs/evaluation/s5-decision-inputs-2026-10-18/claude-projection/attempt-5/claude-projection.json +1797 -0
- package/docs/evaluation/s5-decision-inputs-2026-10-18/worker-usage.md +63 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/README.md +212 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/binding.json +20 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/config-ev-off.json +43 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/config-ev-spool.json +43 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/config-l3.json +39 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/config-pre.json +31 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e-teardown-setup-remove.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e1-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e1-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e1-digest-ev-e1.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e1-leaks.json +131 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e1-oracle.json +74 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e1-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e1-setup-1-dryrun.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e1-setup-2-apply.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e1-setup-3-verify.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e2-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e2-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e2-digest-ev-e2l.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e2-digest-ev-e2s.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e2-leaks.json +205 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e2-oracle.json +110 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e2-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e3b-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e3b-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e3b-digest-ev-e3.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e3b-leaks.json +215 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e3b-oracle.json +79 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e3b-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e4-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e4-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e4-digest-ev-e4a.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e4-digest-ev-e4b.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e4-late-a.json +16 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e4-late-b.json +16 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e4-leaks.json +167 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e4-oracle.json +70 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e4-run.json +37 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e5-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e5-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e5-digest-ev-e5.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e5-leaks.json +82 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e5-oracle.json +60 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e5-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e5-server-env-names.txt +14 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e6b-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e6b-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e6b-digest-ev-e1.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e6b-leaks.json +156 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e6b-run.json +37 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e6b-setup-apply.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/e6b-teardown-remove.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/final-e-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/kit.sha256 +23 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l2-packed-off.json +1185 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l2-packed-spool.json +1334 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l2-wt-off.json +1185 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l2-wt-spool.json +1334 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv1-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv1-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv1-digest-cc-analysis.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv1-leaks.json +80 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv1-oracle.json +61 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv1-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv2-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv2-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv2-leaks.json +80 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv2-oracle-stage1.json +34 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv2-oracle-stage2.json +35 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv2-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv2-snapshot-stage1.txt +4 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv2-snapshot-stage2.txt +5 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv3-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv3-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv3-digest-cc-abort.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv3-leaks.json +82 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv3-oracle-flow.json +61 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv3-oracle-late.json +16 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cc-inv3-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-controller-skill-remove.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-controller-skill-setup.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv1-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv1-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv1-digest-cx-analysis.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv1-leaks.json +93 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv1-oracle.json +61 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv1-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv2-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv2-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv2-leaks.json +93 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv2-oracle-stage1.json +34 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv2-oracle-stage2.json +35 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv2-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv2-snapshot-stage1.txt +4 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv2-snapshot-stage2.txt +5 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv3-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv3-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv3-digest-cx-abort.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv3-leaks.json +95 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv3-oracle-flow.json +61 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv3-oracle-late.json +16 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/l3-cx-inv3-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0-digest-pre-analysis.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0-digest-pre-code.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0-leaks.json +86 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0-oracle.json +39 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0a-harness-fail-clean-10s.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0a-harness-fail-clean.json +15 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0a-harness-fail-digest-pre-analysis.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0a-harness-fail-digest-pre-code.json +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0a-harness-fail-leaks.json +47 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0a-harness-fail-oracle.json +39 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/p0a-harness-fail-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/regdigest-after-cc.json +6 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/regdigest-after-cx.json +6 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/regdigest-after-e.json +6 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/regdigest-before.json +6 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w1-control-apply.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w1-control-dryrun.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w1-provision-apply-attempt1.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w1-provision-apply-attempt2.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w1-provision-dryrun.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w1-writer-verify.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w2-leaks.json +130 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w2-oracle.txt +14 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w2-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w2b-oracle.txt +16 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w2c-oracle.txt +8 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w2d-oracle.txt +6 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w2r-leaks.json +130 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w2r-oracle.txt +3 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w2r-run.json +33 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w3-collect.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w3-finish.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w3r-collect.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w3r-finish.json +1 -0
- package/docs/evaluation/s7-live-acceptance-2026-10-24/evidence/w4-remove.json +1 -0
- package/docs/evaluation/subagent-probe-2026-10-26/README.md +132 -0
- package/docs/evaluation/subagent-probe-2026-10-26/results/K7.json +463 -0
- package/docs/evaluation/subagent-probe-2026-10-26/results/summary.json +50 -0
- package/docs/evaluation/test-coverage-2026-10-17.json +483 -0
- package/docs/evaluation/test-coverage-review-2026-10-17.json +22 -0
- package/docs/evaluation/test-coverage-verification-2026-10-17.json +31 -0
- package/docs/evaluation/token-footprint-2026-10-17/README.md +297 -0
- package/docs/evaluation/token-footprint-2026-10-17/baseline.json +25871 -0
- package/docs/evaluation/token-footprint-2026-10-17/calibration-2/README.md +24 -0
- package/docs/evaluation/token-footprint-2026-10-17/calibration-2/baseline.json +30346 -0
- package/docs/evaluation/token-footprint-2026-10-17/calibration-2/client-projection.json +2538 -0
- package/docs/evaluation/token-footprint-2026-10-17/calibration-2/live-approval/budget.json +13 -0
- package/docs/evaluation/token-footprint-2026-10-17/calibration-2/live-approval/decision.json +69 -0
- package/docs/evaluation/token-footprint-2026-10-17/calibration-2/live-approval/manifest.json +271 -0
- package/docs/evaluation/token-footprint-2026-10-17/calibration-2/live-package.json +785 -0
- package/docs/evaluation/token-footprint-2026-10-17/calibration-2/measurement.json +37956 -0
- package/docs/evaluation/token-footprint-2026-10-17/calibration-2/summary.json +119 -0
- package/docs/evaluation/token-footprint-2026-10-17/checks-final.json +273 -0
- package/docs/evaluation/token-footprint-2026-10-17/checks.json +190 -0
- package/docs/evaluation/token-footprint-2026-10-17/client-attempts/README.md +25 -0
- package/docs/evaluation/token-footprint-2026-10-17/client-attempts/attempt-4.json +2274 -0
- package/docs/evaluation/token-footprint-2026-10-17/client-attempts/pilot-1.json +77 -0
- package/docs/evaluation/token-footprint-2026-10-17/client-attempts/pilot-2.json +81 -0
- package/docs/evaluation/token-footprint-2026-10-17/client-attempts/pilot-3.json +121 -0
- package/docs/evaluation/token-footprint-2026-10-17/current-calibration/baseline.json +30346 -0
- package/docs/evaluation/token-footprint-2026-10-17/current-calibration/measurement.json +37956 -0
- package/docs/evaluation/token-footprint-2026-10-17/current-calibration/summary.json +119 -0
- package/docs/evaluation/token-footprint-2026-10-17/current-client-projection.json +2538 -0
- package/docs/evaluation/token-footprint-2026-10-17/follow-up.md +76 -0
- package/docs/evaluation/token-footprint-2026-10-17/integrity.json +171 -0
- package/docs/evaluation/token-footprint-2026-10-17/live-approval/README.md +36 -0
- package/docs/evaluation/token-footprint-2026-10-17/live-approval/budget.json +13 -0
- package/docs/evaluation/token-footprint-2026-10-17/live-approval/decision.json +69 -0
- package/docs/evaluation/token-footprint-2026-10-17/live-approval/manifest.json +271 -0
- package/docs/evaluation/token-footprint-2026-10-17/live-package.json +795 -0
- package/docs/evaluation/token-footprint-2026-10-17/live-proposal.json +427 -0
- package/docs/evaluation/token-footprint-2026-10-17/m1-pilot/README.md +75 -0
- package/docs/evaluation/token-footprint-2026-10-17/m1-pilot/attempt-1/approved-budget.json +13 -0
- package/docs/evaluation/token-footprint-2026-10-17/m1-pilot/attempt-1/approved-decision.json +69 -0
- package/docs/evaluation/token-footprint-2026-10-17/m1-pilot/attempt-1/approved-manifest.json +271 -0
- package/docs/evaluation/token-footprint-2026-10-17/m1-pilot/attempt-1/token-live-result.json +837 -0
- package/docs/evaluation/token-footprint-2026-10-17/m1-pilot/attempt-2/approved-budget.json +13 -0
- package/docs/evaluation/token-footprint-2026-10-17/m1-pilot/attempt-2/approved-decision.json +69 -0
- package/docs/evaluation/token-footprint-2026-10-17/m1-pilot/attempt-2/approved-manifest.json +271 -0
- package/docs/evaluation/token-footprint-2026-10-17/m1-pilot/attempt-2/token-live-result.json +985 -0
- package/docs/evaluation/token-footprint-2026-10-17/m1-pilot/attempt-3/approved-budget.json +13 -0
- package/docs/evaluation/token-footprint-2026-10-17/m1-pilot/attempt-3/approved-decision.json +69 -0
- package/docs/evaluation/token-footprint-2026-10-17/m1-pilot/attempt-3/approved-manifest.json +271 -0
- package/docs/evaluation/token-footprint-2026-10-17/m1-pilot/attempt-3/token-live-result.json +3055 -0
- package/docs/evaluation/token-footprint-2026-10-17/measurement.json +32476 -0
- package/docs/evaluation/token-footprint-2026-10-17/prior-calibration/README.md +20 -0
- package/docs/evaluation/token-footprint-2026-10-17/prior-calibration/integrity.json +40 -0
- package/docs/evaluation/token-footprint-2026-10-17/summary.json +119 -0
- package/docs/evaluation/tool-definitions-2026-10-18/README.md +204 -0
- package/docs/evaluation/tool-definitions-2026-10-18/codex-projection/after/client-projection.json +2853 -0
- package/docs/evaluation/tool-definitions-2026-10-18/codex-projection/before/client-projection.json +2853 -0
- package/docs/evaluation/tool-definitions-2026-10-18/compare-validation.mjs +64 -0
- package/docs/evaluation/tool-definitions-2026-10-18/diff-validation.mjs +13 -0
- package/docs/evaluation/tool-definitions-2026-10-18/measure-wire.mjs +31 -0
- package/docs/evaluation/worker-system-prompt-2026-10-26/README.md +95 -0
- package/docs/evaluation/worker-system-prompt-2026-10-26/probe.mjs +155 -0
- package/docs/execution-modes.md +388 -0
- package/docs/installation.md +246 -12
- package/docs/isolation-setup.md +258 -0
- package/docs/macos-setup-design.md +2 -2
- package/docs/pi-extensions-design.md +71 -16
- package/docs/releasing.md +12 -6
- package/docs/reviews/2026-10-18-adr-0008-decision-brief.md +53 -0
- package/docs/roadmap.md +320 -9
- package/docs/session-mcp.md +5 -1
- package/docs/skills.md +23 -0
- package/docs/sprints/2026-10-15-s4-decision.md +73 -4
- package/docs/sprints/2026-10-17-token-footprint.md +130 -0
- package/docs/sprints/2026-10-18-s5-maintenance-orchestration.md +202 -0
- package/docs/sprints/2026-10-21-s6-ledger-event-bridge.md +120 -0
- package/docs/sprints/2026-10-24-s7-stabilization-live-acceptance.md +173 -0
- package/docs/sprints/2026-10-24-s8-writer-migration.md +235 -0
- package/docs/sprints/2026-10-26-s10-orchestration-reliability.md +250 -0
- package/docs/sprints/2026-10-26-s11-orchestration-gaps.md +173 -0
- package/docs/threat-model.md +37 -6
- package/docs/tool-contracts.md +91 -9
- package/package.json +17 -11
- package/scripts/isolation-acceptance.mjs +709 -0
- package/scripts/isolation-probe.cjs +213 -0
- package/skills/pi-session-implementation/SKILL.md +13 -3
- package/skills/pi-session-orchestration/SKILL.md +50 -13
- package/skills/pi-session-orchestration/references/start-errors.md +12 -0
- package/skills/pi-session-review/SKILL.md +13 -5
|
@@ -0,0 +1,2538 @@
|
|
|
1
|
+
{
|
|
2
|
+
"schema_version": "token-client-projection-1",
|
|
3
|
+
"binding": {
|
|
4
|
+
"commit": "d6f360926e26629f7b382583c077016daeb954ca",
|
|
5
|
+
"digest": "sha256:33c450601727e38cf63b1874df0fcadc158c3d06af3b472b3d1603da6383c929"
|
|
6
|
+
},
|
|
7
|
+
"build_digest": "sha256:59e7b855ea0bbb545246fbf28e27692ae3b1f35212a3d5526b3190ac58054065",
|
|
8
|
+
"node": "22.23.2",
|
|
9
|
+
"fixture_digest": "sha256:89d7046116226de6fc138d56660c08c0a1225c1801d8a0fbb7b3b8895fe88314",
|
|
10
|
+
"route": "synthetic-loopback-responses",
|
|
11
|
+
"authentication": "disabled-empty-disposable-client-home",
|
|
12
|
+
"client_version": "0.162.0",
|
|
13
|
+
"started_at": "2026-10-17T21:00:08.245Z",
|
|
14
|
+
"ended_at": "2026-10-17T21:00:31.651Z",
|
|
15
|
+
"source_unchanged": true,
|
|
16
|
+
"build_unchanged": true,
|
|
17
|
+
"rows": [
|
|
18
|
+
{
|
|
19
|
+
"scenario": "normal",
|
|
20
|
+
"repetition": 0,
|
|
21
|
+
"pilot": true,
|
|
22
|
+
"exit_code": 0,
|
|
23
|
+
"timed_out": false,
|
|
24
|
+
"process_cleanup": true,
|
|
25
|
+
"server_cleanup": true,
|
|
26
|
+
"registered_server_cleanup": true,
|
|
27
|
+
"stdout_bytes": 2412,
|
|
28
|
+
"duration_ms": 1484.698617,
|
|
29
|
+
"proxy_integrity": true,
|
|
30
|
+
"requests": [
|
|
31
|
+
{
|
|
32
|
+
"bytes": 43712,
|
|
33
|
+
"input_bytes": 20887,
|
|
34
|
+
"tools_bytes": 21412,
|
|
35
|
+
"outputs": [],
|
|
36
|
+
"newly_added_output_bytes": 0,
|
|
37
|
+
"previous_response_id": false,
|
|
38
|
+
"authorization_present": false
|
|
39
|
+
},
|
|
40
|
+
{
|
|
41
|
+
"bytes": 44509,
|
|
42
|
+
"input_bytes": 21684,
|
|
43
|
+
"tools_bytes": 21412,
|
|
44
|
+
"outputs": [
|
|
45
|
+
{
|
|
46
|
+
"serialized_bytes": 529,
|
|
47
|
+
"marker_hits": 1,
|
|
48
|
+
"error_marker_hits": 0,
|
|
49
|
+
"plaintext_tool_echo_hits": 0,
|
|
50
|
+
"plaintext_thinking_echo_hits": 0,
|
|
51
|
+
"structured_key_present": false,
|
|
52
|
+
"tool": "pi_capabilities_get",
|
|
53
|
+
"call_ordinal": 1,
|
|
54
|
+
"new_in_request": true
|
|
55
|
+
}
|
|
56
|
+
],
|
|
57
|
+
"newly_added_output_bytes": 529,
|
|
58
|
+
"previous_response_id": false,
|
|
59
|
+
"authorization_present": false
|
|
60
|
+
},
|
|
61
|
+
{
|
|
62
|
+
"bytes": 44840,
|
|
63
|
+
"input_bytes": 22015,
|
|
64
|
+
"tools_bytes": 21412,
|
|
65
|
+
"outputs": [
|
|
66
|
+
{
|
|
67
|
+
"serialized_bytes": 529,
|
|
68
|
+
"marker_hits": 1,
|
|
69
|
+
"error_marker_hits": 0,
|
|
70
|
+
"plaintext_tool_echo_hits": 0,
|
|
71
|
+
"plaintext_thinking_echo_hits": 0,
|
|
72
|
+
"structured_key_present": false,
|
|
73
|
+
"tool": "pi_capabilities_get",
|
|
74
|
+
"call_ordinal": 1,
|
|
75
|
+
"new_in_request": false
|
|
76
|
+
},
|
|
77
|
+
{
|
|
78
|
+
"serialized_bytes": 67,
|
|
79
|
+
"marker_hits": 0,
|
|
80
|
+
"error_marker_hits": 0,
|
|
81
|
+
"plaintext_tool_echo_hits": 0,
|
|
82
|
+
"plaintext_thinking_echo_hits": 0,
|
|
83
|
+
"structured_key_present": false,
|
|
84
|
+
"tool": "pi_session_list",
|
|
85
|
+
"call_ordinal": 2,
|
|
86
|
+
"new_in_request": true
|
|
87
|
+
}
|
|
88
|
+
],
|
|
89
|
+
"newly_added_output_bytes": 67,
|
|
90
|
+
"previous_response_id": false,
|
|
91
|
+
"authorization_present": false
|
|
92
|
+
}
|
|
93
|
+
],
|
|
94
|
+
"target_output": {
|
|
95
|
+
"serialized_bytes": 529,
|
|
96
|
+
"marker_hits": 1,
|
|
97
|
+
"error_marker_hits": 0,
|
|
98
|
+
"plaintext_tool_echo_hits": 0,
|
|
99
|
+
"plaintext_thinking_echo_hits": 0,
|
|
100
|
+
"structured_key_present": false,
|
|
101
|
+
"tool": "pi_capabilities_get",
|
|
102
|
+
"call_ordinal": 1,
|
|
103
|
+
"new_in_request": true
|
|
104
|
+
},
|
|
105
|
+
"server_target": {
|
|
106
|
+
"wire_bytes": 1021,
|
|
107
|
+
"content_bytes": 518,
|
|
108
|
+
"structured_bytes": 435,
|
|
109
|
+
"assistant_bytes": 0,
|
|
110
|
+
"state": "none",
|
|
111
|
+
"equal_json_representations": true
|
|
112
|
+
},
|
|
113
|
+
"call_schedule_matched": true,
|
|
114
|
+
"server_representation_counts": [
|
|
115
|
+
2
|
|
116
|
+
],
|
|
117
|
+
"client_completed_truncated_json_decoded": false,
|
|
118
|
+
"session_closed_observed": true,
|
|
119
|
+
"server_completed_capped_turn": false,
|
|
120
|
+
"empty_list_observed": true,
|
|
121
|
+
"invalid": false,
|
|
122
|
+
"passed": true
|
|
123
|
+
},
|
|
124
|
+
{
|
|
125
|
+
"scenario": "normal",
|
|
126
|
+
"repetition": 1,
|
|
127
|
+
"pilot": false,
|
|
128
|
+
"exit_code": 0,
|
|
129
|
+
"timed_out": false,
|
|
130
|
+
"process_cleanup": true,
|
|
131
|
+
"server_cleanup": true,
|
|
132
|
+
"registered_server_cleanup": true,
|
|
133
|
+
"stdout_bytes": 2412,
|
|
134
|
+
"duration_ms": 1367.852488,
|
|
135
|
+
"proxy_integrity": true,
|
|
136
|
+
"requests": [
|
|
137
|
+
{
|
|
138
|
+
"bytes": 43712,
|
|
139
|
+
"input_bytes": 20887,
|
|
140
|
+
"tools_bytes": 21412,
|
|
141
|
+
"outputs": [],
|
|
142
|
+
"newly_added_output_bytes": 0,
|
|
143
|
+
"previous_response_id": false,
|
|
144
|
+
"authorization_present": false
|
|
145
|
+
},
|
|
146
|
+
{
|
|
147
|
+
"bytes": 44509,
|
|
148
|
+
"input_bytes": 21684,
|
|
149
|
+
"tools_bytes": 21412,
|
|
150
|
+
"outputs": [
|
|
151
|
+
{
|
|
152
|
+
"serialized_bytes": 529,
|
|
153
|
+
"marker_hits": 1,
|
|
154
|
+
"error_marker_hits": 0,
|
|
155
|
+
"plaintext_tool_echo_hits": 0,
|
|
156
|
+
"plaintext_thinking_echo_hits": 0,
|
|
157
|
+
"structured_key_present": false,
|
|
158
|
+
"tool": "pi_capabilities_get",
|
|
159
|
+
"call_ordinal": 1,
|
|
160
|
+
"new_in_request": true
|
|
161
|
+
}
|
|
162
|
+
],
|
|
163
|
+
"newly_added_output_bytes": 529,
|
|
164
|
+
"previous_response_id": false,
|
|
165
|
+
"authorization_present": false
|
|
166
|
+
},
|
|
167
|
+
{
|
|
168
|
+
"bytes": 44840,
|
|
169
|
+
"input_bytes": 22015,
|
|
170
|
+
"tools_bytes": 21412,
|
|
171
|
+
"outputs": [
|
|
172
|
+
{
|
|
173
|
+
"serialized_bytes": 529,
|
|
174
|
+
"marker_hits": 1,
|
|
175
|
+
"error_marker_hits": 0,
|
|
176
|
+
"plaintext_tool_echo_hits": 0,
|
|
177
|
+
"plaintext_thinking_echo_hits": 0,
|
|
178
|
+
"structured_key_present": false,
|
|
179
|
+
"tool": "pi_capabilities_get",
|
|
180
|
+
"call_ordinal": 1,
|
|
181
|
+
"new_in_request": false
|
|
182
|
+
},
|
|
183
|
+
{
|
|
184
|
+
"serialized_bytes": 67,
|
|
185
|
+
"marker_hits": 0,
|
|
186
|
+
"error_marker_hits": 0,
|
|
187
|
+
"plaintext_tool_echo_hits": 0,
|
|
188
|
+
"plaintext_thinking_echo_hits": 0,
|
|
189
|
+
"structured_key_present": false,
|
|
190
|
+
"tool": "pi_session_list",
|
|
191
|
+
"call_ordinal": 2,
|
|
192
|
+
"new_in_request": true
|
|
193
|
+
}
|
|
194
|
+
],
|
|
195
|
+
"newly_added_output_bytes": 67,
|
|
196
|
+
"previous_response_id": false,
|
|
197
|
+
"authorization_present": false
|
|
198
|
+
}
|
|
199
|
+
],
|
|
200
|
+
"target_output": {
|
|
201
|
+
"serialized_bytes": 529,
|
|
202
|
+
"marker_hits": 1,
|
|
203
|
+
"error_marker_hits": 0,
|
|
204
|
+
"plaintext_tool_echo_hits": 0,
|
|
205
|
+
"plaintext_thinking_echo_hits": 0,
|
|
206
|
+
"structured_key_present": false,
|
|
207
|
+
"tool": "pi_capabilities_get",
|
|
208
|
+
"call_ordinal": 1,
|
|
209
|
+
"new_in_request": true
|
|
210
|
+
},
|
|
211
|
+
"server_target": {
|
|
212
|
+
"wire_bytes": 1021,
|
|
213
|
+
"content_bytes": 518,
|
|
214
|
+
"structured_bytes": 435,
|
|
215
|
+
"assistant_bytes": 0,
|
|
216
|
+
"state": "none",
|
|
217
|
+
"equal_json_representations": true
|
|
218
|
+
},
|
|
219
|
+
"call_schedule_matched": true,
|
|
220
|
+
"server_representation_counts": [
|
|
221
|
+
2
|
|
222
|
+
],
|
|
223
|
+
"client_completed_truncated_json_decoded": false,
|
|
224
|
+
"session_closed_observed": true,
|
|
225
|
+
"server_completed_capped_turn": false,
|
|
226
|
+
"empty_list_observed": true,
|
|
227
|
+
"invalid": false,
|
|
228
|
+
"passed": true
|
|
229
|
+
},
|
|
230
|
+
{
|
|
231
|
+
"scenario": "normal",
|
|
232
|
+
"repetition": 2,
|
|
233
|
+
"pilot": false,
|
|
234
|
+
"exit_code": 0,
|
|
235
|
+
"timed_out": false,
|
|
236
|
+
"process_cleanup": true,
|
|
237
|
+
"server_cleanup": true,
|
|
238
|
+
"registered_server_cleanup": true,
|
|
239
|
+
"stdout_bytes": 2412,
|
|
240
|
+
"duration_ms": 1273.9514779999995,
|
|
241
|
+
"proxy_integrity": true,
|
|
242
|
+
"requests": [
|
|
243
|
+
{
|
|
244
|
+
"bytes": 43712,
|
|
245
|
+
"input_bytes": 20887,
|
|
246
|
+
"tools_bytes": 21412,
|
|
247
|
+
"outputs": [],
|
|
248
|
+
"newly_added_output_bytes": 0,
|
|
249
|
+
"previous_response_id": false,
|
|
250
|
+
"authorization_present": false
|
|
251
|
+
},
|
|
252
|
+
{
|
|
253
|
+
"bytes": 44509,
|
|
254
|
+
"input_bytes": 21684,
|
|
255
|
+
"tools_bytes": 21412,
|
|
256
|
+
"outputs": [
|
|
257
|
+
{
|
|
258
|
+
"serialized_bytes": 529,
|
|
259
|
+
"marker_hits": 1,
|
|
260
|
+
"error_marker_hits": 0,
|
|
261
|
+
"plaintext_tool_echo_hits": 0,
|
|
262
|
+
"plaintext_thinking_echo_hits": 0,
|
|
263
|
+
"structured_key_present": false,
|
|
264
|
+
"tool": "pi_capabilities_get",
|
|
265
|
+
"call_ordinal": 1,
|
|
266
|
+
"new_in_request": true
|
|
267
|
+
}
|
|
268
|
+
],
|
|
269
|
+
"newly_added_output_bytes": 529,
|
|
270
|
+
"previous_response_id": false,
|
|
271
|
+
"authorization_present": false
|
|
272
|
+
},
|
|
273
|
+
{
|
|
274
|
+
"bytes": 44840,
|
|
275
|
+
"input_bytes": 22015,
|
|
276
|
+
"tools_bytes": 21412,
|
|
277
|
+
"outputs": [
|
|
278
|
+
{
|
|
279
|
+
"serialized_bytes": 529,
|
|
280
|
+
"marker_hits": 1,
|
|
281
|
+
"error_marker_hits": 0,
|
|
282
|
+
"plaintext_tool_echo_hits": 0,
|
|
283
|
+
"plaintext_thinking_echo_hits": 0,
|
|
284
|
+
"structured_key_present": false,
|
|
285
|
+
"tool": "pi_capabilities_get",
|
|
286
|
+
"call_ordinal": 1,
|
|
287
|
+
"new_in_request": false
|
|
288
|
+
},
|
|
289
|
+
{
|
|
290
|
+
"serialized_bytes": 67,
|
|
291
|
+
"marker_hits": 0,
|
|
292
|
+
"error_marker_hits": 0,
|
|
293
|
+
"plaintext_tool_echo_hits": 0,
|
|
294
|
+
"plaintext_thinking_echo_hits": 0,
|
|
295
|
+
"structured_key_present": false,
|
|
296
|
+
"tool": "pi_session_list",
|
|
297
|
+
"call_ordinal": 2,
|
|
298
|
+
"new_in_request": true
|
|
299
|
+
}
|
|
300
|
+
],
|
|
301
|
+
"newly_added_output_bytes": 67,
|
|
302
|
+
"previous_response_id": false,
|
|
303
|
+
"authorization_present": false
|
|
304
|
+
}
|
|
305
|
+
],
|
|
306
|
+
"target_output": {
|
|
307
|
+
"serialized_bytes": 529,
|
|
308
|
+
"marker_hits": 1,
|
|
309
|
+
"error_marker_hits": 0,
|
|
310
|
+
"plaintext_tool_echo_hits": 0,
|
|
311
|
+
"plaintext_thinking_echo_hits": 0,
|
|
312
|
+
"structured_key_present": false,
|
|
313
|
+
"tool": "pi_capabilities_get",
|
|
314
|
+
"call_ordinal": 1,
|
|
315
|
+
"new_in_request": true
|
|
316
|
+
},
|
|
317
|
+
"server_target": {
|
|
318
|
+
"wire_bytes": 1021,
|
|
319
|
+
"content_bytes": 518,
|
|
320
|
+
"structured_bytes": 435,
|
|
321
|
+
"assistant_bytes": 0,
|
|
322
|
+
"state": "none",
|
|
323
|
+
"equal_json_representations": true
|
|
324
|
+
},
|
|
325
|
+
"call_schedule_matched": true,
|
|
326
|
+
"server_representation_counts": [
|
|
327
|
+
2
|
|
328
|
+
],
|
|
329
|
+
"client_completed_truncated_json_decoded": false,
|
|
330
|
+
"session_closed_observed": true,
|
|
331
|
+
"server_completed_capped_turn": false,
|
|
332
|
+
"empty_list_observed": true,
|
|
333
|
+
"invalid": false,
|
|
334
|
+
"passed": true
|
|
335
|
+
},
|
|
336
|
+
{
|
|
337
|
+
"scenario": "normal",
|
|
338
|
+
"repetition": 3,
|
|
339
|
+
"pilot": false,
|
|
340
|
+
"exit_code": 0,
|
|
341
|
+
"timed_out": false,
|
|
342
|
+
"process_cleanup": true,
|
|
343
|
+
"server_cleanup": true,
|
|
344
|
+
"registered_server_cleanup": true,
|
|
345
|
+
"stdout_bytes": 2412,
|
|
346
|
+
"duration_ms": 1400.8376600000001,
|
|
347
|
+
"proxy_integrity": true,
|
|
348
|
+
"requests": [
|
|
349
|
+
{
|
|
350
|
+
"bytes": 43712,
|
|
351
|
+
"input_bytes": 20887,
|
|
352
|
+
"tools_bytes": 21412,
|
|
353
|
+
"outputs": [],
|
|
354
|
+
"newly_added_output_bytes": 0,
|
|
355
|
+
"previous_response_id": false,
|
|
356
|
+
"authorization_present": false
|
|
357
|
+
},
|
|
358
|
+
{
|
|
359
|
+
"bytes": 44509,
|
|
360
|
+
"input_bytes": 21684,
|
|
361
|
+
"tools_bytes": 21412,
|
|
362
|
+
"outputs": [
|
|
363
|
+
{
|
|
364
|
+
"serialized_bytes": 529,
|
|
365
|
+
"marker_hits": 1,
|
|
366
|
+
"error_marker_hits": 0,
|
|
367
|
+
"plaintext_tool_echo_hits": 0,
|
|
368
|
+
"plaintext_thinking_echo_hits": 0,
|
|
369
|
+
"structured_key_present": false,
|
|
370
|
+
"tool": "pi_capabilities_get",
|
|
371
|
+
"call_ordinal": 1,
|
|
372
|
+
"new_in_request": true
|
|
373
|
+
}
|
|
374
|
+
],
|
|
375
|
+
"newly_added_output_bytes": 529,
|
|
376
|
+
"previous_response_id": false,
|
|
377
|
+
"authorization_present": false
|
|
378
|
+
},
|
|
379
|
+
{
|
|
380
|
+
"bytes": 44840,
|
|
381
|
+
"input_bytes": 22015,
|
|
382
|
+
"tools_bytes": 21412,
|
|
383
|
+
"outputs": [
|
|
384
|
+
{
|
|
385
|
+
"serialized_bytes": 529,
|
|
386
|
+
"marker_hits": 1,
|
|
387
|
+
"error_marker_hits": 0,
|
|
388
|
+
"plaintext_tool_echo_hits": 0,
|
|
389
|
+
"plaintext_thinking_echo_hits": 0,
|
|
390
|
+
"structured_key_present": false,
|
|
391
|
+
"tool": "pi_capabilities_get",
|
|
392
|
+
"call_ordinal": 1,
|
|
393
|
+
"new_in_request": false
|
|
394
|
+
},
|
|
395
|
+
{
|
|
396
|
+
"serialized_bytes": 67,
|
|
397
|
+
"marker_hits": 0,
|
|
398
|
+
"error_marker_hits": 0,
|
|
399
|
+
"plaintext_tool_echo_hits": 0,
|
|
400
|
+
"plaintext_thinking_echo_hits": 0,
|
|
401
|
+
"structured_key_present": false,
|
|
402
|
+
"tool": "pi_session_list",
|
|
403
|
+
"call_ordinal": 2,
|
|
404
|
+
"new_in_request": true
|
|
405
|
+
}
|
|
406
|
+
],
|
|
407
|
+
"newly_added_output_bytes": 67,
|
|
408
|
+
"previous_response_id": false,
|
|
409
|
+
"authorization_present": false
|
|
410
|
+
}
|
|
411
|
+
],
|
|
412
|
+
"target_output": {
|
|
413
|
+
"serialized_bytes": 529,
|
|
414
|
+
"marker_hits": 1,
|
|
415
|
+
"error_marker_hits": 0,
|
|
416
|
+
"plaintext_tool_echo_hits": 0,
|
|
417
|
+
"plaintext_thinking_echo_hits": 0,
|
|
418
|
+
"structured_key_present": false,
|
|
419
|
+
"tool": "pi_capabilities_get",
|
|
420
|
+
"call_ordinal": 1,
|
|
421
|
+
"new_in_request": true
|
|
422
|
+
},
|
|
423
|
+
"server_target": {
|
|
424
|
+
"wire_bytes": 1021,
|
|
425
|
+
"content_bytes": 518,
|
|
426
|
+
"structured_bytes": 435,
|
|
427
|
+
"assistant_bytes": 0,
|
|
428
|
+
"state": "none",
|
|
429
|
+
"equal_json_representations": true
|
|
430
|
+
},
|
|
431
|
+
"call_schedule_matched": true,
|
|
432
|
+
"server_representation_counts": [
|
|
433
|
+
2
|
|
434
|
+
],
|
|
435
|
+
"client_completed_truncated_json_decoded": false,
|
|
436
|
+
"session_closed_observed": true,
|
|
437
|
+
"server_completed_capped_turn": false,
|
|
438
|
+
"empty_list_observed": true,
|
|
439
|
+
"invalid": false,
|
|
440
|
+
"passed": true
|
|
441
|
+
},
|
|
442
|
+
{
|
|
443
|
+
"scenario": "normal",
|
|
444
|
+
"repetition": 4,
|
|
445
|
+
"pilot": false,
|
|
446
|
+
"exit_code": 0,
|
|
447
|
+
"timed_out": false,
|
|
448
|
+
"process_cleanup": true,
|
|
449
|
+
"server_cleanup": true,
|
|
450
|
+
"registered_server_cleanup": true,
|
|
451
|
+
"stdout_bytes": 2412,
|
|
452
|
+
"duration_ms": 1354.91478,
|
|
453
|
+
"proxy_integrity": true,
|
|
454
|
+
"requests": [
|
|
455
|
+
{
|
|
456
|
+
"bytes": 43712,
|
|
457
|
+
"input_bytes": 20887,
|
|
458
|
+
"tools_bytes": 21412,
|
|
459
|
+
"outputs": [],
|
|
460
|
+
"newly_added_output_bytes": 0,
|
|
461
|
+
"previous_response_id": false,
|
|
462
|
+
"authorization_present": false
|
|
463
|
+
},
|
|
464
|
+
{
|
|
465
|
+
"bytes": 44509,
|
|
466
|
+
"input_bytes": 21684,
|
|
467
|
+
"tools_bytes": 21412,
|
|
468
|
+
"outputs": [
|
|
469
|
+
{
|
|
470
|
+
"serialized_bytes": 529,
|
|
471
|
+
"marker_hits": 1,
|
|
472
|
+
"error_marker_hits": 0,
|
|
473
|
+
"plaintext_tool_echo_hits": 0,
|
|
474
|
+
"plaintext_thinking_echo_hits": 0,
|
|
475
|
+
"structured_key_present": false,
|
|
476
|
+
"tool": "pi_capabilities_get",
|
|
477
|
+
"call_ordinal": 1,
|
|
478
|
+
"new_in_request": true
|
|
479
|
+
}
|
|
480
|
+
],
|
|
481
|
+
"newly_added_output_bytes": 529,
|
|
482
|
+
"previous_response_id": false,
|
|
483
|
+
"authorization_present": false
|
|
484
|
+
},
|
|
485
|
+
{
|
|
486
|
+
"bytes": 44840,
|
|
487
|
+
"input_bytes": 22015,
|
|
488
|
+
"tools_bytes": 21412,
|
|
489
|
+
"outputs": [
|
|
490
|
+
{
|
|
491
|
+
"serialized_bytes": 529,
|
|
492
|
+
"marker_hits": 1,
|
|
493
|
+
"error_marker_hits": 0,
|
|
494
|
+
"plaintext_tool_echo_hits": 0,
|
|
495
|
+
"plaintext_thinking_echo_hits": 0,
|
|
496
|
+
"structured_key_present": false,
|
|
497
|
+
"tool": "pi_capabilities_get",
|
|
498
|
+
"call_ordinal": 1,
|
|
499
|
+
"new_in_request": false
|
|
500
|
+
},
|
|
501
|
+
{
|
|
502
|
+
"serialized_bytes": 67,
|
|
503
|
+
"marker_hits": 0,
|
|
504
|
+
"error_marker_hits": 0,
|
|
505
|
+
"plaintext_tool_echo_hits": 0,
|
|
506
|
+
"plaintext_thinking_echo_hits": 0,
|
|
507
|
+
"structured_key_present": false,
|
|
508
|
+
"tool": "pi_session_list",
|
|
509
|
+
"call_ordinal": 2,
|
|
510
|
+
"new_in_request": true
|
|
511
|
+
}
|
|
512
|
+
],
|
|
513
|
+
"newly_added_output_bytes": 67,
|
|
514
|
+
"previous_response_id": false,
|
|
515
|
+
"authorization_present": false
|
|
516
|
+
}
|
|
517
|
+
],
|
|
518
|
+
"target_output": {
|
|
519
|
+
"serialized_bytes": 529,
|
|
520
|
+
"marker_hits": 1,
|
|
521
|
+
"error_marker_hits": 0,
|
|
522
|
+
"plaintext_tool_echo_hits": 0,
|
|
523
|
+
"plaintext_thinking_echo_hits": 0,
|
|
524
|
+
"structured_key_present": false,
|
|
525
|
+
"tool": "pi_capabilities_get",
|
|
526
|
+
"call_ordinal": 1,
|
|
527
|
+
"new_in_request": true
|
|
528
|
+
},
|
|
529
|
+
"server_target": {
|
|
530
|
+
"wire_bytes": 1021,
|
|
531
|
+
"content_bytes": 518,
|
|
532
|
+
"structured_bytes": 435,
|
|
533
|
+
"assistant_bytes": 0,
|
|
534
|
+
"state": "none",
|
|
535
|
+
"equal_json_representations": true
|
|
536
|
+
},
|
|
537
|
+
"call_schedule_matched": true,
|
|
538
|
+
"server_representation_counts": [
|
|
539
|
+
2
|
|
540
|
+
],
|
|
541
|
+
"client_completed_truncated_json_decoded": false,
|
|
542
|
+
"session_closed_observed": true,
|
|
543
|
+
"server_completed_capped_turn": false,
|
|
544
|
+
"empty_list_observed": true,
|
|
545
|
+
"invalid": false,
|
|
546
|
+
"passed": true
|
|
547
|
+
},
|
|
548
|
+
{
|
|
549
|
+
"scenario": "normal",
|
|
550
|
+
"repetition": 5,
|
|
551
|
+
"pilot": false,
|
|
552
|
+
"exit_code": 0,
|
|
553
|
+
"timed_out": false,
|
|
554
|
+
"process_cleanup": true,
|
|
555
|
+
"server_cleanup": true,
|
|
556
|
+
"registered_server_cleanup": true,
|
|
557
|
+
"stdout_bytes": 2412,
|
|
558
|
+
"duration_ms": 1320.5823859999991,
|
|
559
|
+
"proxy_integrity": true,
|
|
560
|
+
"requests": [
|
|
561
|
+
{
|
|
562
|
+
"bytes": 43712,
|
|
563
|
+
"input_bytes": 20887,
|
|
564
|
+
"tools_bytes": 21412,
|
|
565
|
+
"outputs": [],
|
|
566
|
+
"newly_added_output_bytes": 0,
|
|
567
|
+
"previous_response_id": false,
|
|
568
|
+
"authorization_present": false
|
|
569
|
+
},
|
|
570
|
+
{
|
|
571
|
+
"bytes": 44509,
|
|
572
|
+
"input_bytes": 21684,
|
|
573
|
+
"tools_bytes": 21412,
|
|
574
|
+
"outputs": [
|
|
575
|
+
{
|
|
576
|
+
"serialized_bytes": 529,
|
|
577
|
+
"marker_hits": 1,
|
|
578
|
+
"error_marker_hits": 0,
|
|
579
|
+
"plaintext_tool_echo_hits": 0,
|
|
580
|
+
"plaintext_thinking_echo_hits": 0,
|
|
581
|
+
"structured_key_present": false,
|
|
582
|
+
"tool": "pi_capabilities_get",
|
|
583
|
+
"call_ordinal": 1,
|
|
584
|
+
"new_in_request": true
|
|
585
|
+
}
|
|
586
|
+
],
|
|
587
|
+
"newly_added_output_bytes": 529,
|
|
588
|
+
"previous_response_id": false,
|
|
589
|
+
"authorization_present": false
|
|
590
|
+
},
|
|
591
|
+
{
|
|
592
|
+
"bytes": 44840,
|
|
593
|
+
"input_bytes": 22015,
|
|
594
|
+
"tools_bytes": 21412,
|
|
595
|
+
"outputs": [
|
|
596
|
+
{
|
|
597
|
+
"serialized_bytes": 529,
|
|
598
|
+
"marker_hits": 1,
|
|
599
|
+
"error_marker_hits": 0,
|
|
600
|
+
"plaintext_tool_echo_hits": 0,
|
|
601
|
+
"plaintext_thinking_echo_hits": 0,
|
|
602
|
+
"structured_key_present": false,
|
|
603
|
+
"tool": "pi_capabilities_get",
|
|
604
|
+
"call_ordinal": 1,
|
|
605
|
+
"new_in_request": false
|
|
606
|
+
},
|
|
607
|
+
{
|
|
608
|
+
"serialized_bytes": 67,
|
|
609
|
+
"marker_hits": 0,
|
|
610
|
+
"error_marker_hits": 0,
|
|
611
|
+
"plaintext_tool_echo_hits": 0,
|
|
612
|
+
"plaintext_thinking_echo_hits": 0,
|
|
613
|
+
"structured_key_present": false,
|
|
614
|
+
"tool": "pi_session_list",
|
|
615
|
+
"call_ordinal": 2,
|
|
616
|
+
"new_in_request": true
|
|
617
|
+
}
|
|
618
|
+
],
|
|
619
|
+
"newly_added_output_bytes": 67,
|
|
620
|
+
"previous_response_id": false,
|
|
621
|
+
"authorization_present": false
|
|
622
|
+
}
|
|
623
|
+
],
|
|
624
|
+
"target_output": {
|
|
625
|
+
"serialized_bytes": 529,
|
|
626
|
+
"marker_hits": 1,
|
|
627
|
+
"error_marker_hits": 0,
|
|
628
|
+
"plaintext_tool_echo_hits": 0,
|
|
629
|
+
"plaintext_thinking_echo_hits": 0,
|
|
630
|
+
"structured_key_present": false,
|
|
631
|
+
"tool": "pi_capabilities_get",
|
|
632
|
+
"call_ordinal": 1,
|
|
633
|
+
"new_in_request": true
|
|
634
|
+
},
|
|
635
|
+
"server_target": {
|
|
636
|
+
"wire_bytes": 1021,
|
|
637
|
+
"content_bytes": 518,
|
|
638
|
+
"structured_bytes": 435,
|
|
639
|
+
"assistant_bytes": 0,
|
|
640
|
+
"state": "none",
|
|
641
|
+
"equal_json_representations": true
|
|
642
|
+
},
|
|
643
|
+
"call_schedule_matched": true,
|
|
644
|
+
"server_representation_counts": [
|
|
645
|
+
2
|
|
646
|
+
],
|
|
647
|
+
"client_completed_truncated_json_decoded": false,
|
|
648
|
+
"session_closed_observed": true,
|
|
649
|
+
"server_completed_capped_turn": false,
|
|
650
|
+
"empty_list_observed": true,
|
|
651
|
+
"invalid": false,
|
|
652
|
+
"passed": true
|
|
653
|
+
},
|
|
654
|
+
{
|
|
655
|
+
"scenario": "domain-error",
|
|
656
|
+
"repetition": 1,
|
|
657
|
+
"pilot": false,
|
|
658
|
+
"exit_code": 0,
|
|
659
|
+
"timed_out": false,
|
|
660
|
+
"process_cleanup": true,
|
|
661
|
+
"server_cleanup": true,
|
|
662
|
+
"registered_server_cleanup": true,
|
|
663
|
+
"stdout_bytes": 1759,
|
|
664
|
+
"duration_ms": 1309.1705340000008,
|
|
665
|
+
"proxy_integrity": true,
|
|
666
|
+
"requests": [
|
|
667
|
+
{
|
|
668
|
+
"bytes": 43712,
|
|
669
|
+
"input_bytes": 20887,
|
|
670
|
+
"tools_bytes": 21412,
|
|
671
|
+
"outputs": [],
|
|
672
|
+
"newly_added_output_bytes": 0,
|
|
673
|
+
"previous_response_id": false,
|
|
674
|
+
"authorization_present": false
|
|
675
|
+
},
|
|
676
|
+
{
|
|
677
|
+
"bytes": 44170,
|
|
678
|
+
"input_bytes": 21345,
|
|
679
|
+
"tools_bytes": 21412,
|
|
680
|
+
"outputs": [
|
|
681
|
+
{
|
|
682
|
+
"serialized_bytes": 129,
|
|
683
|
+
"marker_hits": 0,
|
|
684
|
+
"error_marker_hits": 1,
|
|
685
|
+
"plaintext_tool_echo_hits": 0,
|
|
686
|
+
"plaintext_thinking_echo_hits": 0,
|
|
687
|
+
"structured_key_present": false,
|
|
688
|
+
"tool": "pi_session_start",
|
|
689
|
+
"call_ordinal": 1,
|
|
690
|
+
"new_in_request": true
|
|
691
|
+
}
|
|
692
|
+
],
|
|
693
|
+
"newly_added_output_bytes": 129,
|
|
694
|
+
"previous_response_id": false,
|
|
695
|
+
"authorization_present": false
|
|
696
|
+
},
|
|
697
|
+
{
|
|
698
|
+
"bytes": 44501,
|
|
699
|
+
"input_bytes": 21676,
|
|
700
|
+
"tools_bytes": 21412,
|
|
701
|
+
"outputs": [
|
|
702
|
+
{
|
|
703
|
+
"serialized_bytes": 129,
|
|
704
|
+
"marker_hits": 0,
|
|
705
|
+
"error_marker_hits": 1,
|
|
706
|
+
"plaintext_tool_echo_hits": 0,
|
|
707
|
+
"plaintext_thinking_echo_hits": 0,
|
|
708
|
+
"structured_key_present": false,
|
|
709
|
+
"tool": "pi_session_start",
|
|
710
|
+
"call_ordinal": 1,
|
|
711
|
+
"new_in_request": false
|
|
712
|
+
},
|
|
713
|
+
{
|
|
714
|
+
"serialized_bytes": 67,
|
|
715
|
+
"marker_hits": 0,
|
|
716
|
+
"error_marker_hits": 0,
|
|
717
|
+
"plaintext_tool_echo_hits": 0,
|
|
718
|
+
"plaintext_thinking_echo_hits": 0,
|
|
719
|
+
"structured_key_present": false,
|
|
720
|
+
"tool": "pi_session_list",
|
|
721
|
+
"call_ordinal": 2,
|
|
722
|
+
"new_in_request": true
|
|
723
|
+
}
|
|
724
|
+
],
|
|
725
|
+
"newly_added_output_bytes": 67,
|
|
726
|
+
"previous_response_id": false,
|
|
727
|
+
"authorization_present": false
|
|
728
|
+
}
|
|
729
|
+
],
|
|
730
|
+
"target_output": {
|
|
731
|
+
"serialized_bytes": 129,
|
|
732
|
+
"marker_hits": 0,
|
|
733
|
+
"error_marker_hits": 1,
|
|
734
|
+
"plaintext_tool_echo_hits": 0,
|
|
735
|
+
"plaintext_thinking_echo_hits": 0,
|
|
736
|
+
"structured_key_present": false,
|
|
737
|
+
"tool": "pi_session_start",
|
|
738
|
+
"call_ordinal": 1,
|
|
739
|
+
"new_in_request": true
|
|
740
|
+
},
|
|
741
|
+
"server_target": {
|
|
742
|
+
"wire_bytes": 280,
|
|
743
|
+
"content_bytes": 118,
|
|
744
|
+
"structured_bytes": 79,
|
|
745
|
+
"assistant_bytes": 0,
|
|
746
|
+
"state": "domain-error",
|
|
747
|
+
"equal_json_representations": true
|
|
748
|
+
},
|
|
749
|
+
"call_schedule_matched": true,
|
|
750
|
+
"server_representation_counts": [
|
|
751
|
+
2
|
|
752
|
+
],
|
|
753
|
+
"client_completed_truncated_json_decoded": false,
|
|
754
|
+
"session_closed_observed": true,
|
|
755
|
+
"server_completed_capped_turn": false,
|
|
756
|
+
"empty_list_observed": true,
|
|
757
|
+
"invalid": false,
|
|
758
|
+
"passed": true
|
|
759
|
+
},
|
|
760
|
+
{
|
|
761
|
+
"scenario": "domain-error",
|
|
762
|
+
"repetition": 2,
|
|
763
|
+
"pilot": false,
|
|
764
|
+
"exit_code": 0,
|
|
765
|
+
"timed_out": false,
|
|
766
|
+
"process_cleanup": true,
|
|
767
|
+
"server_cleanup": true,
|
|
768
|
+
"registered_server_cleanup": true,
|
|
769
|
+
"stdout_bytes": 1759,
|
|
770
|
+
"duration_ms": 1232.8938200000011,
|
|
771
|
+
"proxy_integrity": true,
|
|
772
|
+
"requests": [
|
|
773
|
+
{
|
|
774
|
+
"bytes": 43712,
|
|
775
|
+
"input_bytes": 20887,
|
|
776
|
+
"tools_bytes": 21412,
|
|
777
|
+
"outputs": [],
|
|
778
|
+
"newly_added_output_bytes": 0,
|
|
779
|
+
"previous_response_id": false,
|
|
780
|
+
"authorization_present": false
|
|
781
|
+
},
|
|
782
|
+
{
|
|
783
|
+
"bytes": 44170,
|
|
784
|
+
"input_bytes": 21345,
|
|
785
|
+
"tools_bytes": 21412,
|
|
786
|
+
"outputs": [
|
|
787
|
+
{
|
|
788
|
+
"serialized_bytes": 129,
|
|
789
|
+
"marker_hits": 0,
|
|
790
|
+
"error_marker_hits": 1,
|
|
791
|
+
"plaintext_tool_echo_hits": 0,
|
|
792
|
+
"plaintext_thinking_echo_hits": 0,
|
|
793
|
+
"structured_key_present": false,
|
|
794
|
+
"tool": "pi_session_start",
|
|
795
|
+
"call_ordinal": 1,
|
|
796
|
+
"new_in_request": true
|
|
797
|
+
}
|
|
798
|
+
],
|
|
799
|
+
"newly_added_output_bytes": 129,
|
|
800
|
+
"previous_response_id": false,
|
|
801
|
+
"authorization_present": false
|
|
802
|
+
},
|
|
803
|
+
{
|
|
804
|
+
"bytes": 44501,
|
|
805
|
+
"input_bytes": 21676,
|
|
806
|
+
"tools_bytes": 21412,
|
|
807
|
+
"outputs": [
|
|
808
|
+
{
|
|
809
|
+
"serialized_bytes": 129,
|
|
810
|
+
"marker_hits": 0,
|
|
811
|
+
"error_marker_hits": 1,
|
|
812
|
+
"plaintext_tool_echo_hits": 0,
|
|
813
|
+
"plaintext_thinking_echo_hits": 0,
|
|
814
|
+
"structured_key_present": false,
|
|
815
|
+
"tool": "pi_session_start",
|
|
816
|
+
"call_ordinal": 1,
|
|
817
|
+
"new_in_request": false
|
|
818
|
+
},
|
|
819
|
+
{
|
|
820
|
+
"serialized_bytes": 67,
|
|
821
|
+
"marker_hits": 0,
|
|
822
|
+
"error_marker_hits": 0,
|
|
823
|
+
"plaintext_tool_echo_hits": 0,
|
|
824
|
+
"plaintext_thinking_echo_hits": 0,
|
|
825
|
+
"structured_key_present": false,
|
|
826
|
+
"tool": "pi_session_list",
|
|
827
|
+
"call_ordinal": 2,
|
|
828
|
+
"new_in_request": true
|
|
829
|
+
}
|
|
830
|
+
],
|
|
831
|
+
"newly_added_output_bytes": 67,
|
|
832
|
+
"previous_response_id": false,
|
|
833
|
+
"authorization_present": false
|
|
834
|
+
}
|
|
835
|
+
],
|
|
836
|
+
"target_output": {
|
|
837
|
+
"serialized_bytes": 129,
|
|
838
|
+
"marker_hits": 0,
|
|
839
|
+
"error_marker_hits": 1,
|
|
840
|
+
"plaintext_tool_echo_hits": 0,
|
|
841
|
+
"plaintext_thinking_echo_hits": 0,
|
|
842
|
+
"structured_key_present": false,
|
|
843
|
+
"tool": "pi_session_start",
|
|
844
|
+
"call_ordinal": 1,
|
|
845
|
+
"new_in_request": true
|
|
846
|
+
},
|
|
847
|
+
"server_target": {
|
|
848
|
+
"wire_bytes": 280,
|
|
849
|
+
"content_bytes": 118,
|
|
850
|
+
"structured_bytes": 79,
|
|
851
|
+
"assistant_bytes": 0,
|
|
852
|
+
"state": "domain-error",
|
|
853
|
+
"equal_json_representations": true
|
|
854
|
+
},
|
|
855
|
+
"call_schedule_matched": true,
|
|
856
|
+
"server_representation_counts": [
|
|
857
|
+
2
|
|
858
|
+
],
|
|
859
|
+
"client_completed_truncated_json_decoded": false,
|
|
860
|
+
"session_closed_observed": true,
|
|
861
|
+
"server_completed_capped_turn": false,
|
|
862
|
+
"empty_list_observed": true,
|
|
863
|
+
"invalid": false,
|
|
864
|
+
"passed": true
|
|
865
|
+
},
|
|
866
|
+
{
|
|
867
|
+
"scenario": "domain-error",
|
|
868
|
+
"repetition": 3,
|
|
869
|
+
"pilot": false,
|
|
870
|
+
"exit_code": 0,
|
|
871
|
+
"timed_out": false,
|
|
872
|
+
"process_cleanup": true,
|
|
873
|
+
"server_cleanup": true,
|
|
874
|
+
"registered_server_cleanup": true,
|
|
875
|
+
"stdout_bytes": 1759,
|
|
876
|
+
"duration_ms": 1285.728631,
|
|
877
|
+
"proxy_integrity": true,
|
|
878
|
+
"requests": [
|
|
879
|
+
{
|
|
880
|
+
"bytes": 43712,
|
|
881
|
+
"input_bytes": 20887,
|
|
882
|
+
"tools_bytes": 21412,
|
|
883
|
+
"outputs": [],
|
|
884
|
+
"newly_added_output_bytes": 0,
|
|
885
|
+
"previous_response_id": false,
|
|
886
|
+
"authorization_present": false
|
|
887
|
+
},
|
|
888
|
+
{
|
|
889
|
+
"bytes": 44170,
|
|
890
|
+
"input_bytes": 21345,
|
|
891
|
+
"tools_bytes": 21412,
|
|
892
|
+
"outputs": [
|
|
893
|
+
{
|
|
894
|
+
"serialized_bytes": 129,
|
|
895
|
+
"marker_hits": 0,
|
|
896
|
+
"error_marker_hits": 1,
|
|
897
|
+
"plaintext_tool_echo_hits": 0,
|
|
898
|
+
"plaintext_thinking_echo_hits": 0,
|
|
899
|
+
"structured_key_present": false,
|
|
900
|
+
"tool": "pi_session_start",
|
|
901
|
+
"call_ordinal": 1,
|
|
902
|
+
"new_in_request": true
|
|
903
|
+
}
|
|
904
|
+
],
|
|
905
|
+
"newly_added_output_bytes": 129,
|
|
906
|
+
"previous_response_id": false,
|
|
907
|
+
"authorization_present": false
|
|
908
|
+
},
|
|
909
|
+
{
|
|
910
|
+
"bytes": 44501,
|
|
911
|
+
"input_bytes": 21676,
|
|
912
|
+
"tools_bytes": 21412,
|
|
913
|
+
"outputs": [
|
|
914
|
+
{
|
|
915
|
+
"serialized_bytes": 129,
|
|
916
|
+
"marker_hits": 0,
|
|
917
|
+
"error_marker_hits": 1,
|
|
918
|
+
"plaintext_tool_echo_hits": 0,
|
|
919
|
+
"plaintext_thinking_echo_hits": 0,
|
|
920
|
+
"structured_key_present": false,
|
|
921
|
+
"tool": "pi_session_start",
|
|
922
|
+
"call_ordinal": 1,
|
|
923
|
+
"new_in_request": false
|
|
924
|
+
},
|
|
925
|
+
{
|
|
926
|
+
"serialized_bytes": 67,
|
|
927
|
+
"marker_hits": 0,
|
|
928
|
+
"error_marker_hits": 0,
|
|
929
|
+
"plaintext_tool_echo_hits": 0,
|
|
930
|
+
"plaintext_thinking_echo_hits": 0,
|
|
931
|
+
"structured_key_present": false,
|
|
932
|
+
"tool": "pi_session_list",
|
|
933
|
+
"call_ordinal": 2,
|
|
934
|
+
"new_in_request": true
|
|
935
|
+
}
|
|
936
|
+
],
|
|
937
|
+
"newly_added_output_bytes": 67,
|
|
938
|
+
"previous_response_id": false,
|
|
939
|
+
"authorization_present": false
|
|
940
|
+
}
|
|
941
|
+
],
|
|
942
|
+
"target_output": {
|
|
943
|
+
"serialized_bytes": 129,
|
|
944
|
+
"marker_hits": 0,
|
|
945
|
+
"error_marker_hits": 1,
|
|
946
|
+
"plaintext_tool_echo_hits": 0,
|
|
947
|
+
"plaintext_thinking_echo_hits": 0,
|
|
948
|
+
"structured_key_present": false,
|
|
949
|
+
"tool": "pi_session_start",
|
|
950
|
+
"call_ordinal": 1,
|
|
951
|
+
"new_in_request": true
|
|
952
|
+
},
|
|
953
|
+
"server_target": {
|
|
954
|
+
"wire_bytes": 280,
|
|
955
|
+
"content_bytes": 118,
|
|
956
|
+
"structured_bytes": 79,
|
|
957
|
+
"assistant_bytes": 0,
|
|
958
|
+
"state": "domain-error",
|
|
959
|
+
"equal_json_representations": true
|
|
960
|
+
},
|
|
961
|
+
"call_schedule_matched": true,
|
|
962
|
+
"server_representation_counts": [
|
|
963
|
+
2
|
|
964
|
+
],
|
|
965
|
+
"client_completed_truncated_json_decoded": false,
|
|
966
|
+
"session_closed_observed": true,
|
|
967
|
+
"server_completed_capped_turn": false,
|
|
968
|
+
"empty_list_observed": true,
|
|
969
|
+
"invalid": false,
|
|
970
|
+
"passed": true
|
|
971
|
+
},
|
|
972
|
+
{
|
|
973
|
+
"scenario": "domain-error",
|
|
974
|
+
"repetition": 4,
|
|
975
|
+
"pilot": false,
|
|
976
|
+
"exit_code": 0,
|
|
977
|
+
"timed_out": false,
|
|
978
|
+
"process_cleanup": true,
|
|
979
|
+
"server_cleanup": true,
|
|
980
|
+
"registered_server_cleanup": true,
|
|
981
|
+
"stdout_bytes": 1759,
|
|
982
|
+
"duration_ms": 1370.5734510000002,
|
|
983
|
+
"proxy_integrity": true,
|
|
984
|
+
"requests": [
|
|
985
|
+
{
|
|
986
|
+
"bytes": 43712,
|
|
987
|
+
"input_bytes": 20887,
|
|
988
|
+
"tools_bytes": 21412,
|
|
989
|
+
"outputs": [],
|
|
990
|
+
"newly_added_output_bytes": 0,
|
|
991
|
+
"previous_response_id": false,
|
|
992
|
+
"authorization_present": false
|
|
993
|
+
},
|
|
994
|
+
{
|
|
995
|
+
"bytes": 44170,
|
|
996
|
+
"input_bytes": 21345,
|
|
997
|
+
"tools_bytes": 21412,
|
|
998
|
+
"outputs": [
|
|
999
|
+
{
|
|
1000
|
+
"serialized_bytes": 129,
|
|
1001
|
+
"marker_hits": 0,
|
|
1002
|
+
"error_marker_hits": 1,
|
|
1003
|
+
"plaintext_tool_echo_hits": 0,
|
|
1004
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1005
|
+
"structured_key_present": false,
|
|
1006
|
+
"tool": "pi_session_start",
|
|
1007
|
+
"call_ordinal": 1,
|
|
1008
|
+
"new_in_request": true
|
|
1009
|
+
}
|
|
1010
|
+
],
|
|
1011
|
+
"newly_added_output_bytes": 129,
|
|
1012
|
+
"previous_response_id": false,
|
|
1013
|
+
"authorization_present": false
|
|
1014
|
+
},
|
|
1015
|
+
{
|
|
1016
|
+
"bytes": 44501,
|
|
1017
|
+
"input_bytes": 21676,
|
|
1018
|
+
"tools_bytes": 21412,
|
|
1019
|
+
"outputs": [
|
|
1020
|
+
{
|
|
1021
|
+
"serialized_bytes": 129,
|
|
1022
|
+
"marker_hits": 0,
|
|
1023
|
+
"error_marker_hits": 1,
|
|
1024
|
+
"plaintext_tool_echo_hits": 0,
|
|
1025
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1026
|
+
"structured_key_present": false,
|
|
1027
|
+
"tool": "pi_session_start",
|
|
1028
|
+
"call_ordinal": 1,
|
|
1029
|
+
"new_in_request": false
|
|
1030
|
+
},
|
|
1031
|
+
{
|
|
1032
|
+
"serialized_bytes": 67,
|
|
1033
|
+
"marker_hits": 0,
|
|
1034
|
+
"error_marker_hits": 0,
|
|
1035
|
+
"plaintext_tool_echo_hits": 0,
|
|
1036
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1037
|
+
"structured_key_present": false,
|
|
1038
|
+
"tool": "pi_session_list",
|
|
1039
|
+
"call_ordinal": 2,
|
|
1040
|
+
"new_in_request": true
|
|
1041
|
+
}
|
|
1042
|
+
],
|
|
1043
|
+
"newly_added_output_bytes": 67,
|
|
1044
|
+
"previous_response_id": false,
|
|
1045
|
+
"authorization_present": false
|
|
1046
|
+
}
|
|
1047
|
+
],
|
|
1048
|
+
"target_output": {
|
|
1049
|
+
"serialized_bytes": 129,
|
|
1050
|
+
"marker_hits": 0,
|
|
1051
|
+
"error_marker_hits": 1,
|
|
1052
|
+
"plaintext_tool_echo_hits": 0,
|
|
1053
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1054
|
+
"structured_key_present": false,
|
|
1055
|
+
"tool": "pi_session_start",
|
|
1056
|
+
"call_ordinal": 1,
|
|
1057
|
+
"new_in_request": true
|
|
1058
|
+
},
|
|
1059
|
+
"server_target": {
|
|
1060
|
+
"wire_bytes": 280,
|
|
1061
|
+
"content_bytes": 118,
|
|
1062
|
+
"structured_bytes": 79,
|
|
1063
|
+
"assistant_bytes": 0,
|
|
1064
|
+
"state": "domain-error",
|
|
1065
|
+
"equal_json_representations": true
|
|
1066
|
+
},
|
|
1067
|
+
"call_schedule_matched": true,
|
|
1068
|
+
"server_representation_counts": [
|
|
1069
|
+
2
|
|
1070
|
+
],
|
|
1071
|
+
"client_completed_truncated_json_decoded": false,
|
|
1072
|
+
"session_closed_observed": true,
|
|
1073
|
+
"server_completed_capped_turn": false,
|
|
1074
|
+
"empty_list_observed": true,
|
|
1075
|
+
"invalid": false,
|
|
1076
|
+
"passed": true
|
|
1077
|
+
},
|
|
1078
|
+
{
|
|
1079
|
+
"scenario": "domain-error",
|
|
1080
|
+
"repetition": 5,
|
|
1081
|
+
"pilot": false,
|
|
1082
|
+
"exit_code": 0,
|
|
1083
|
+
"timed_out": false,
|
|
1084
|
+
"process_cleanup": true,
|
|
1085
|
+
"server_cleanup": true,
|
|
1086
|
+
"registered_server_cleanup": true,
|
|
1087
|
+
"stdout_bytes": 1759,
|
|
1088
|
+
"duration_ms": 1251.6357939999998,
|
|
1089
|
+
"proxy_integrity": true,
|
|
1090
|
+
"requests": [
|
|
1091
|
+
{
|
|
1092
|
+
"bytes": 43712,
|
|
1093
|
+
"input_bytes": 20887,
|
|
1094
|
+
"tools_bytes": 21412,
|
|
1095
|
+
"outputs": [],
|
|
1096
|
+
"newly_added_output_bytes": 0,
|
|
1097
|
+
"previous_response_id": false,
|
|
1098
|
+
"authorization_present": false
|
|
1099
|
+
},
|
|
1100
|
+
{
|
|
1101
|
+
"bytes": 44170,
|
|
1102
|
+
"input_bytes": 21345,
|
|
1103
|
+
"tools_bytes": 21412,
|
|
1104
|
+
"outputs": [
|
|
1105
|
+
{
|
|
1106
|
+
"serialized_bytes": 129,
|
|
1107
|
+
"marker_hits": 0,
|
|
1108
|
+
"error_marker_hits": 1,
|
|
1109
|
+
"plaintext_tool_echo_hits": 0,
|
|
1110
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1111
|
+
"structured_key_present": false,
|
|
1112
|
+
"tool": "pi_session_start",
|
|
1113
|
+
"call_ordinal": 1,
|
|
1114
|
+
"new_in_request": true
|
|
1115
|
+
}
|
|
1116
|
+
],
|
|
1117
|
+
"newly_added_output_bytes": 129,
|
|
1118
|
+
"previous_response_id": false,
|
|
1119
|
+
"authorization_present": false
|
|
1120
|
+
},
|
|
1121
|
+
{
|
|
1122
|
+
"bytes": 44501,
|
|
1123
|
+
"input_bytes": 21676,
|
|
1124
|
+
"tools_bytes": 21412,
|
|
1125
|
+
"outputs": [
|
|
1126
|
+
{
|
|
1127
|
+
"serialized_bytes": 129,
|
|
1128
|
+
"marker_hits": 0,
|
|
1129
|
+
"error_marker_hits": 1,
|
|
1130
|
+
"plaintext_tool_echo_hits": 0,
|
|
1131
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1132
|
+
"structured_key_present": false,
|
|
1133
|
+
"tool": "pi_session_start",
|
|
1134
|
+
"call_ordinal": 1,
|
|
1135
|
+
"new_in_request": false
|
|
1136
|
+
},
|
|
1137
|
+
{
|
|
1138
|
+
"serialized_bytes": 67,
|
|
1139
|
+
"marker_hits": 0,
|
|
1140
|
+
"error_marker_hits": 0,
|
|
1141
|
+
"plaintext_tool_echo_hits": 0,
|
|
1142
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1143
|
+
"structured_key_present": false,
|
|
1144
|
+
"tool": "pi_session_list",
|
|
1145
|
+
"call_ordinal": 2,
|
|
1146
|
+
"new_in_request": true
|
|
1147
|
+
}
|
|
1148
|
+
],
|
|
1149
|
+
"newly_added_output_bytes": 67,
|
|
1150
|
+
"previous_response_id": false,
|
|
1151
|
+
"authorization_present": false
|
|
1152
|
+
}
|
|
1153
|
+
],
|
|
1154
|
+
"target_output": {
|
|
1155
|
+
"serialized_bytes": 129,
|
|
1156
|
+
"marker_hits": 0,
|
|
1157
|
+
"error_marker_hits": 1,
|
|
1158
|
+
"plaintext_tool_echo_hits": 0,
|
|
1159
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1160
|
+
"structured_key_present": false,
|
|
1161
|
+
"tool": "pi_session_start",
|
|
1162
|
+
"call_ordinal": 1,
|
|
1163
|
+
"new_in_request": true
|
|
1164
|
+
},
|
|
1165
|
+
"server_target": {
|
|
1166
|
+
"wire_bytes": 280,
|
|
1167
|
+
"content_bytes": 118,
|
|
1168
|
+
"structured_bytes": 79,
|
|
1169
|
+
"assistant_bytes": 0,
|
|
1170
|
+
"state": "domain-error",
|
|
1171
|
+
"equal_json_representations": true
|
|
1172
|
+
},
|
|
1173
|
+
"call_schedule_matched": true,
|
|
1174
|
+
"server_representation_counts": [
|
|
1175
|
+
2
|
|
1176
|
+
],
|
|
1177
|
+
"client_completed_truncated_json_decoded": false,
|
|
1178
|
+
"session_closed_observed": true,
|
|
1179
|
+
"server_completed_capped_turn": false,
|
|
1180
|
+
"empty_list_observed": true,
|
|
1181
|
+
"invalid": false,
|
|
1182
|
+
"passed": true
|
|
1183
|
+
},
|
|
1184
|
+
{
|
|
1185
|
+
"scenario": "truncated",
|
|
1186
|
+
"repetition": 1,
|
|
1187
|
+
"pilot": false,
|
|
1188
|
+
"exit_code": 0,
|
|
1189
|
+
"timed_out": false,
|
|
1190
|
+
"process_cleanup": true,
|
|
1191
|
+
"server_cleanup": true,
|
|
1192
|
+
"registered_server_cleanup": true,
|
|
1193
|
+
"stdout_bytes": 137521,
|
|
1194
|
+
"duration_ms": 1843.724521,
|
|
1195
|
+
"proxy_integrity": true,
|
|
1196
|
+
"requests": [
|
|
1197
|
+
{
|
|
1198
|
+
"bytes": 43712,
|
|
1199
|
+
"input_bytes": 20887,
|
|
1200
|
+
"tools_bytes": 21412,
|
|
1201
|
+
"outputs": [],
|
|
1202
|
+
"newly_added_output_bytes": 0,
|
|
1203
|
+
"previous_response_id": false,
|
|
1204
|
+
"authorization_present": false
|
|
1205
|
+
},
|
|
1206
|
+
{
|
|
1207
|
+
"bytes": 44445,
|
|
1208
|
+
"input_bytes": 21620,
|
|
1209
|
+
"tools_bytes": 21412,
|
|
1210
|
+
"outputs": [
|
|
1211
|
+
{
|
|
1212
|
+
"serialized_bytes": 394,
|
|
1213
|
+
"marker_hits": 1,
|
|
1214
|
+
"error_marker_hits": 0,
|
|
1215
|
+
"plaintext_tool_echo_hits": 0,
|
|
1216
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1217
|
+
"structured_key_present": false,
|
|
1218
|
+
"tool": "pi_session_start",
|
|
1219
|
+
"call_ordinal": 1,
|
|
1220
|
+
"new_in_request": true
|
|
1221
|
+
}
|
|
1222
|
+
],
|
|
1223
|
+
"newly_added_output_bytes": 394,
|
|
1224
|
+
"previous_response_id": false,
|
|
1225
|
+
"authorization_present": false
|
|
1226
|
+
},
|
|
1227
|
+
{
|
|
1228
|
+
"bytes": 45440,
|
|
1229
|
+
"input_bytes": 22615,
|
|
1230
|
+
"tools_bytes": 21412,
|
|
1231
|
+
"outputs": [
|
|
1232
|
+
{
|
|
1233
|
+
"serialized_bytes": 394,
|
|
1234
|
+
"marker_hits": 1,
|
|
1235
|
+
"error_marker_hits": 0,
|
|
1236
|
+
"plaintext_tool_echo_hits": 0,
|
|
1237
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1238
|
+
"structured_key_present": false,
|
|
1239
|
+
"tool": "pi_session_start",
|
|
1240
|
+
"call_ordinal": 1,
|
|
1241
|
+
"new_in_request": false
|
|
1242
|
+
},
|
|
1243
|
+
{
|
|
1244
|
+
"serialized_bytes": 623,
|
|
1245
|
+
"marker_hits": 1,
|
|
1246
|
+
"error_marker_hits": 0,
|
|
1247
|
+
"plaintext_tool_echo_hits": 0,
|
|
1248
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1249
|
+
"structured_key_present": false,
|
|
1250
|
+
"tool": "pi_session_prompt",
|
|
1251
|
+
"call_ordinal": 2,
|
|
1252
|
+
"new_in_request": true
|
|
1253
|
+
}
|
|
1254
|
+
],
|
|
1255
|
+
"newly_added_output_bytes": 623,
|
|
1256
|
+
"previous_response_id": false,
|
|
1257
|
+
"authorization_present": false
|
|
1258
|
+
},
|
|
1259
|
+
{
|
|
1260
|
+
"bytes": 57868,
|
|
1261
|
+
"input_bytes": 35043,
|
|
1262
|
+
"tools_bytes": 21412,
|
|
1263
|
+
"outputs": [
|
|
1264
|
+
{
|
|
1265
|
+
"serialized_bytes": 394,
|
|
1266
|
+
"marker_hits": 1,
|
|
1267
|
+
"error_marker_hits": 0,
|
|
1268
|
+
"plaintext_tool_echo_hits": 0,
|
|
1269
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1270
|
+
"structured_key_present": false,
|
|
1271
|
+
"tool": "pi_session_start",
|
|
1272
|
+
"call_ordinal": 1,
|
|
1273
|
+
"new_in_request": false
|
|
1274
|
+
},
|
|
1275
|
+
{
|
|
1276
|
+
"serialized_bytes": 623,
|
|
1277
|
+
"marker_hits": 1,
|
|
1278
|
+
"error_marker_hits": 0,
|
|
1279
|
+
"plaintext_tool_echo_hits": 0,
|
|
1280
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1281
|
+
"structured_key_present": false,
|
|
1282
|
+
"tool": "pi_session_prompt",
|
|
1283
|
+
"call_ordinal": 2,
|
|
1284
|
+
"new_in_request": false
|
|
1285
|
+
},
|
|
1286
|
+
{
|
|
1287
|
+
"serialized_bytes": 12062,
|
|
1288
|
+
"marker_hits": 1,
|
|
1289
|
+
"error_marker_hits": 0,
|
|
1290
|
+
"plaintext_tool_echo_hits": 1,
|
|
1291
|
+
"plaintext_thinking_echo_hits": 1,
|
|
1292
|
+
"structured_key_present": false,
|
|
1293
|
+
"tool": "pi_turn_get",
|
|
1294
|
+
"call_ordinal": 3,
|
|
1295
|
+
"new_in_request": true
|
|
1296
|
+
}
|
|
1297
|
+
],
|
|
1298
|
+
"newly_added_output_bytes": 12062,
|
|
1299
|
+
"previous_response_id": false,
|
|
1300
|
+
"authorization_present": false
|
|
1301
|
+
},
|
|
1302
|
+
{
|
|
1303
|
+
"bytes": 58583,
|
|
1304
|
+
"input_bytes": 35758,
|
|
1305
|
+
"tools_bytes": 21412,
|
|
1306
|
+
"outputs": [
|
|
1307
|
+
{
|
|
1308
|
+
"serialized_bytes": 394,
|
|
1309
|
+
"marker_hits": 1,
|
|
1310
|
+
"error_marker_hits": 0,
|
|
1311
|
+
"plaintext_tool_echo_hits": 0,
|
|
1312
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1313
|
+
"structured_key_present": false,
|
|
1314
|
+
"tool": "pi_session_start",
|
|
1315
|
+
"call_ordinal": 1,
|
|
1316
|
+
"new_in_request": false
|
|
1317
|
+
},
|
|
1318
|
+
{
|
|
1319
|
+
"serialized_bytes": 623,
|
|
1320
|
+
"marker_hits": 1,
|
|
1321
|
+
"error_marker_hits": 0,
|
|
1322
|
+
"plaintext_tool_echo_hits": 0,
|
|
1323
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1324
|
+
"structured_key_present": false,
|
|
1325
|
+
"tool": "pi_session_prompt",
|
|
1326
|
+
"call_ordinal": 2,
|
|
1327
|
+
"new_in_request": false
|
|
1328
|
+
},
|
|
1329
|
+
{
|
|
1330
|
+
"serialized_bytes": 12062,
|
|
1331
|
+
"marker_hits": 1,
|
|
1332
|
+
"error_marker_hits": 0,
|
|
1333
|
+
"plaintext_tool_echo_hits": 1,
|
|
1334
|
+
"plaintext_thinking_echo_hits": 1,
|
|
1335
|
+
"structured_key_present": false,
|
|
1336
|
+
"tool": "pi_turn_get",
|
|
1337
|
+
"call_ordinal": 3,
|
|
1338
|
+
"new_in_request": false
|
|
1339
|
+
},
|
|
1340
|
+
{
|
|
1341
|
+
"serialized_bytes": 396,
|
|
1342
|
+
"marker_hits": 1,
|
|
1343
|
+
"error_marker_hits": 0,
|
|
1344
|
+
"plaintext_tool_echo_hits": 0,
|
|
1345
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1346
|
+
"structured_key_present": false,
|
|
1347
|
+
"tool": "pi_session_close",
|
|
1348
|
+
"call_ordinal": 4,
|
|
1349
|
+
"new_in_request": true
|
|
1350
|
+
}
|
|
1351
|
+
],
|
|
1352
|
+
"newly_added_output_bytes": 396,
|
|
1353
|
+
"previous_response_id": false,
|
|
1354
|
+
"authorization_present": false
|
|
1355
|
+
},
|
|
1356
|
+
{
|
|
1357
|
+
"bytes": 58914,
|
|
1358
|
+
"input_bytes": 36089,
|
|
1359
|
+
"tools_bytes": 21412,
|
|
1360
|
+
"outputs": [
|
|
1361
|
+
{
|
|
1362
|
+
"serialized_bytes": 394,
|
|
1363
|
+
"marker_hits": 1,
|
|
1364
|
+
"error_marker_hits": 0,
|
|
1365
|
+
"plaintext_tool_echo_hits": 0,
|
|
1366
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1367
|
+
"structured_key_present": false,
|
|
1368
|
+
"tool": "pi_session_start",
|
|
1369
|
+
"call_ordinal": 1,
|
|
1370
|
+
"new_in_request": false
|
|
1371
|
+
},
|
|
1372
|
+
{
|
|
1373
|
+
"serialized_bytes": 623,
|
|
1374
|
+
"marker_hits": 1,
|
|
1375
|
+
"error_marker_hits": 0,
|
|
1376
|
+
"plaintext_tool_echo_hits": 0,
|
|
1377
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1378
|
+
"structured_key_present": false,
|
|
1379
|
+
"tool": "pi_session_prompt",
|
|
1380
|
+
"call_ordinal": 2,
|
|
1381
|
+
"new_in_request": false
|
|
1382
|
+
},
|
|
1383
|
+
{
|
|
1384
|
+
"serialized_bytes": 12062,
|
|
1385
|
+
"marker_hits": 1,
|
|
1386
|
+
"error_marker_hits": 0,
|
|
1387
|
+
"plaintext_tool_echo_hits": 1,
|
|
1388
|
+
"plaintext_thinking_echo_hits": 1,
|
|
1389
|
+
"structured_key_present": false,
|
|
1390
|
+
"tool": "pi_turn_get",
|
|
1391
|
+
"call_ordinal": 3,
|
|
1392
|
+
"new_in_request": false
|
|
1393
|
+
},
|
|
1394
|
+
{
|
|
1395
|
+
"serialized_bytes": 396,
|
|
1396
|
+
"marker_hits": 1,
|
|
1397
|
+
"error_marker_hits": 0,
|
|
1398
|
+
"plaintext_tool_echo_hits": 0,
|
|
1399
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1400
|
+
"structured_key_present": false,
|
|
1401
|
+
"tool": "pi_session_close",
|
|
1402
|
+
"call_ordinal": 4,
|
|
1403
|
+
"new_in_request": false
|
|
1404
|
+
},
|
|
1405
|
+
{
|
|
1406
|
+
"serialized_bytes": 67,
|
|
1407
|
+
"marker_hits": 0,
|
|
1408
|
+
"error_marker_hits": 0,
|
|
1409
|
+
"plaintext_tool_echo_hits": 0,
|
|
1410
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1411
|
+
"structured_key_present": false,
|
|
1412
|
+
"tool": "pi_session_list",
|
|
1413
|
+
"call_ordinal": 5,
|
|
1414
|
+
"new_in_request": true
|
|
1415
|
+
}
|
|
1416
|
+
],
|
|
1417
|
+
"newly_added_output_bytes": 67,
|
|
1418
|
+
"previous_response_id": false,
|
|
1419
|
+
"authorization_present": false
|
|
1420
|
+
}
|
|
1421
|
+
],
|
|
1422
|
+
"target_output": {
|
|
1423
|
+
"serialized_bytes": 12062,
|
|
1424
|
+
"marker_hits": 1,
|
|
1425
|
+
"error_marker_hits": 0,
|
|
1426
|
+
"plaintext_tool_echo_hits": 1,
|
|
1427
|
+
"plaintext_thinking_echo_hits": 1,
|
|
1428
|
+
"structured_key_present": false,
|
|
1429
|
+
"tool": "pi_turn_get",
|
|
1430
|
+
"call_ordinal": 3,
|
|
1431
|
+
"new_in_request": true
|
|
1432
|
+
},
|
|
1433
|
+
"server_target": {
|
|
1434
|
+
"wire_bytes": 131789,
|
|
1435
|
+
"content_bytes": 65891,
|
|
1436
|
+
"structured_bytes": 65830,
|
|
1437
|
+
"assistant_bytes": 65536,
|
|
1438
|
+
"state": "completed",
|
|
1439
|
+
"equal_json_representations": true
|
|
1440
|
+
},
|
|
1441
|
+
"call_schedule_matched": true,
|
|
1442
|
+
"server_representation_counts": [
|
|
1443
|
+
2
|
|
1444
|
+
],
|
|
1445
|
+
"client_completed_truncated_json_decoded": true,
|
|
1446
|
+
"session_closed_observed": true,
|
|
1447
|
+
"server_completed_capped_turn": true,
|
|
1448
|
+
"empty_list_observed": true,
|
|
1449
|
+
"invalid": false,
|
|
1450
|
+
"passed": true
|
|
1451
|
+
},
|
|
1452
|
+
{
|
|
1453
|
+
"scenario": "truncated",
|
|
1454
|
+
"repetition": 2,
|
|
1455
|
+
"pilot": false,
|
|
1456
|
+
"exit_code": 0,
|
|
1457
|
+
"timed_out": false,
|
|
1458
|
+
"process_cleanup": true,
|
|
1459
|
+
"server_cleanup": true,
|
|
1460
|
+
"registered_server_cleanup": true,
|
|
1461
|
+
"stdout_bytes": 137521,
|
|
1462
|
+
"duration_ms": 2010.7479209999983,
|
|
1463
|
+
"proxy_integrity": true,
|
|
1464
|
+
"requests": [
|
|
1465
|
+
{
|
|
1466
|
+
"bytes": 43712,
|
|
1467
|
+
"input_bytes": 20887,
|
|
1468
|
+
"tools_bytes": 21412,
|
|
1469
|
+
"outputs": [],
|
|
1470
|
+
"newly_added_output_bytes": 0,
|
|
1471
|
+
"previous_response_id": false,
|
|
1472
|
+
"authorization_present": false
|
|
1473
|
+
},
|
|
1474
|
+
{
|
|
1475
|
+
"bytes": 44445,
|
|
1476
|
+
"input_bytes": 21620,
|
|
1477
|
+
"tools_bytes": 21412,
|
|
1478
|
+
"outputs": [
|
|
1479
|
+
{
|
|
1480
|
+
"serialized_bytes": 394,
|
|
1481
|
+
"marker_hits": 1,
|
|
1482
|
+
"error_marker_hits": 0,
|
|
1483
|
+
"plaintext_tool_echo_hits": 0,
|
|
1484
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1485
|
+
"structured_key_present": false,
|
|
1486
|
+
"tool": "pi_session_start",
|
|
1487
|
+
"call_ordinal": 1,
|
|
1488
|
+
"new_in_request": true
|
|
1489
|
+
}
|
|
1490
|
+
],
|
|
1491
|
+
"newly_added_output_bytes": 394,
|
|
1492
|
+
"previous_response_id": false,
|
|
1493
|
+
"authorization_present": false
|
|
1494
|
+
},
|
|
1495
|
+
{
|
|
1496
|
+
"bytes": 45440,
|
|
1497
|
+
"input_bytes": 22615,
|
|
1498
|
+
"tools_bytes": 21412,
|
|
1499
|
+
"outputs": [
|
|
1500
|
+
{
|
|
1501
|
+
"serialized_bytes": 394,
|
|
1502
|
+
"marker_hits": 1,
|
|
1503
|
+
"error_marker_hits": 0,
|
|
1504
|
+
"plaintext_tool_echo_hits": 0,
|
|
1505
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1506
|
+
"structured_key_present": false,
|
|
1507
|
+
"tool": "pi_session_start",
|
|
1508
|
+
"call_ordinal": 1,
|
|
1509
|
+
"new_in_request": false
|
|
1510
|
+
},
|
|
1511
|
+
{
|
|
1512
|
+
"serialized_bytes": 623,
|
|
1513
|
+
"marker_hits": 1,
|
|
1514
|
+
"error_marker_hits": 0,
|
|
1515
|
+
"plaintext_tool_echo_hits": 0,
|
|
1516
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1517
|
+
"structured_key_present": false,
|
|
1518
|
+
"tool": "pi_session_prompt",
|
|
1519
|
+
"call_ordinal": 2,
|
|
1520
|
+
"new_in_request": true
|
|
1521
|
+
}
|
|
1522
|
+
],
|
|
1523
|
+
"newly_added_output_bytes": 623,
|
|
1524
|
+
"previous_response_id": false,
|
|
1525
|
+
"authorization_present": false
|
|
1526
|
+
},
|
|
1527
|
+
{
|
|
1528
|
+
"bytes": 57868,
|
|
1529
|
+
"input_bytes": 35043,
|
|
1530
|
+
"tools_bytes": 21412,
|
|
1531
|
+
"outputs": [
|
|
1532
|
+
{
|
|
1533
|
+
"serialized_bytes": 394,
|
|
1534
|
+
"marker_hits": 1,
|
|
1535
|
+
"error_marker_hits": 0,
|
|
1536
|
+
"plaintext_tool_echo_hits": 0,
|
|
1537
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1538
|
+
"structured_key_present": false,
|
|
1539
|
+
"tool": "pi_session_start",
|
|
1540
|
+
"call_ordinal": 1,
|
|
1541
|
+
"new_in_request": false
|
|
1542
|
+
},
|
|
1543
|
+
{
|
|
1544
|
+
"serialized_bytes": 623,
|
|
1545
|
+
"marker_hits": 1,
|
|
1546
|
+
"error_marker_hits": 0,
|
|
1547
|
+
"plaintext_tool_echo_hits": 0,
|
|
1548
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1549
|
+
"structured_key_present": false,
|
|
1550
|
+
"tool": "pi_session_prompt",
|
|
1551
|
+
"call_ordinal": 2,
|
|
1552
|
+
"new_in_request": false
|
|
1553
|
+
},
|
|
1554
|
+
{
|
|
1555
|
+
"serialized_bytes": 12062,
|
|
1556
|
+
"marker_hits": 1,
|
|
1557
|
+
"error_marker_hits": 0,
|
|
1558
|
+
"plaintext_tool_echo_hits": 1,
|
|
1559
|
+
"plaintext_thinking_echo_hits": 1,
|
|
1560
|
+
"structured_key_present": false,
|
|
1561
|
+
"tool": "pi_turn_get",
|
|
1562
|
+
"call_ordinal": 3,
|
|
1563
|
+
"new_in_request": true
|
|
1564
|
+
}
|
|
1565
|
+
],
|
|
1566
|
+
"newly_added_output_bytes": 12062,
|
|
1567
|
+
"previous_response_id": false,
|
|
1568
|
+
"authorization_present": false
|
|
1569
|
+
},
|
|
1570
|
+
{
|
|
1571
|
+
"bytes": 58583,
|
|
1572
|
+
"input_bytes": 35758,
|
|
1573
|
+
"tools_bytes": 21412,
|
|
1574
|
+
"outputs": [
|
|
1575
|
+
{
|
|
1576
|
+
"serialized_bytes": 394,
|
|
1577
|
+
"marker_hits": 1,
|
|
1578
|
+
"error_marker_hits": 0,
|
|
1579
|
+
"plaintext_tool_echo_hits": 0,
|
|
1580
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1581
|
+
"structured_key_present": false,
|
|
1582
|
+
"tool": "pi_session_start",
|
|
1583
|
+
"call_ordinal": 1,
|
|
1584
|
+
"new_in_request": false
|
|
1585
|
+
},
|
|
1586
|
+
{
|
|
1587
|
+
"serialized_bytes": 623,
|
|
1588
|
+
"marker_hits": 1,
|
|
1589
|
+
"error_marker_hits": 0,
|
|
1590
|
+
"plaintext_tool_echo_hits": 0,
|
|
1591
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1592
|
+
"structured_key_present": false,
|
|
1593
|
+
"tool": "pi_session_prompt",
|
|
1594
|
+
"call_ordinal": 2,
|
|
1595
|
+
"new_in_request": false
|
|
1596
|
+
},
|
|
1597
|
+
{
|
|
1598
|
+
"serialized_bytes": 12062,
|
|
1599
|
+
"marker_hits": 1,
|
|
1600
|
+
"error_marker_hits": 0,
|
|
1601
|
+
"plaintext_tool_echo_hits": 1,
|
|
1602
|
+
"plaintext_thinking_echo_hits": 1,
|
|
1603
|
+
"structured_key_present": false,
|
|
1604
|
+
"tool": "pi_turn_get",
|
|
1605
|
+
"call_ordinal": 3,
|
|
1606
|
+
"new_in_request": false
|
|
1607
|
+
},
|
|
1608
|
+
{
|
|
1609
|
+
"serialized_bytes": 396,
|
|
1610
|
+
"marker_hits": 1,
|
|
1611
|
+
"error_marker_hits": 0,
|
|
1612
|
+
"plaintext_tool_echo_hits": 0,
|
|
1613
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1614
|
+
"structured_key_present": false,
|
|
1615
|
+
"tool": "pi_session_close",
|
|
1616
|
+
"call_ordinal": 4,
|
|
1617
|
+
"new_in_request": true
|
|
1618
|
+
}
|
|
1619
|
+
],
|
|
1620
|
+
"newly_added_output_bytes": 396,
|
|
1621
|
+
"previous_response_id": false,
|
|
1622
|
+
"authorization_present": false
|
|
1623
|
+
},
|
|
1624
|
+
{
|
|
1625
|
+
"bytes": 58914,
|
|
1626
|
+
"input_bytes": 36089,
|
|
1627
|
+
"tools_bytes": 21412,
|
|
1628
|
+
"outputs": [
|
|
1629
|
+
{
|
|
1630
|
+
"serialized_bytes": 394,
|
|
1631
|
+
"marker_hits": 1,
|
|
1632
|
+
"error_marker_hits": 0,
|
|
1633
|
+
"plaintext_tool_echo_hits": 0,
|
|
1634
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1635
|
+
"structured_key_present": false,
|
|
1636
|
+
"tool": "pi_session_start",
|
|
1637
|
+
"call_ordinal": 1,
|
|
1638
|
+
"new_in_request": false
|
|
1639
|
+
},
|
|
1640
|
+
{
|
|
1641
|
+
"serialized_bytes": 623,
|
|
1642
|
+
"marker_hits": 1,
|
|
1643
|
+
"error_marker_hits": 0,
|
|
1644
|
+
"plaintext_tool_echo_hits": 0,
|
|
1645
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1646
|
+
"structured_key_present": false,
|
|
1647
|
+
"tool": "pi_session_prompt",
|
|
1648
|
+
"call_ordinal": 2,
|
|
1649
|
+
"new_in_request": false
|
|
1650
|
+
},
|
|
1651
|
+
{
|
|
1652
|
+
"serialized_bytes": 12062,
|
|
1653
|
+
"marker_hits": 1,
|
|
1654
|
+
"error_marker_hits": 0,
|
|
1655
|
+
"plaintext_tool_echo_hits": 1,
|
|
1656
|
+
"plaintext_thinking_echo_hits": 1,
|
|
1657
|
+
"structured_key_present": false,
|
|
1658
|
+
"tool": "pi_turn_get",
|
|
1659
|
+
"call_ordinal": 3,
|
|
1660
|
+
"new_in_request": false
|
|
1661
|
+
},
|
|
1662
|
+
{
|
|
1663
|
+
"serialized_bytes": 396,
|
|
1664
|
+
"marker_hits": 1,
|
|
1665
|
+
"error_marker_hits": 0,
|
|
1666
|
+
"plaintext_tool_echo_hits": 0,
|
|
1667
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1668
|
+
"structured_key_present": false,
|
|
1669
|
+
"tool": "pi_session_close",
|
|
1670
|
+
"call_ordinal": 4,
|
|
1671
|
+
"new_in_request": false
|
|
1672
|
+
},
|
|
1673
|
+
{
|
|
1674
|
+
"serialized_bytes": 67,
|
|
1675
|
+
"marker_hits": 0,
|
|
1676
|
+
"error_marker_hits": 0,
|
|
1677
|
+
"plaintext_tool_echo_hits": 0,
|
|
1678
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1679
|
+
"structured_key_present": false,
|
|
1680
|
+
"tool": "pi_session_list",
|
|
1681
|
+
"call_ordinal": 5,
|
|
1682
|
+
"new_in_request": true
|
|
1683
|
+
}
|
|
1684
|
+
],
|
|
1685
|
+
"newly_added_output_bytes": 67,
|
|
1686
|
+
"previous_response_id": false,
|
|
1687
|
+
"authorization_present": false
|
|
1688
|
+
}
|
|
1689
|
+
],
|
|
1690
|
+
"target_output": {
|
|
1691
|
+
"serialized_bytes": 12062,
|
|
1692
|
+
"marker_hits": 1,
|
|
1693
|
+
"error_marker_hits": 0,
|
|
1694
|
+
"plaintext_tool_echo_hits": 1,
|
|
1695
|
+
"plaintext_thinking_echo_hits": 1,
|
|
1696
|
+
"structured_key_present": false,
|
|
1697
|
+
"tool": "pi_turn_get",
|
|
1698
|
+
"call_ordinal": 3,
|
|
1699
|
+
"new_in_request": true
|
|
1700
|
+
},
|
|
1701
|
+
"server_target": {
|
|
1702
|
+
"wire_bytes": 131789,
|
|
1703
|
+
"content_bytes": 65891,
|
|
1704
|
+
"structured_bytes": 65830,
|
|
1705
|
+
"assistant_bytes": 65536,
|
|
1706
|
+
"state": "completed",
|
|
1707
|
+
"equal_json_representations": true
|
|
1708
|
+
},
|
|
1709
|
+
"call_schedule_matched": true,
|
|
1710
|
+
"server_representation_counts": [
|
|
1711
|
+
2
|
|
1712
|
+
],
|
|
1713
|
+
"client_completed_truncated_json_decoded": true,
|
|
1714
|
+
"session_closed_observed": true,
|
|
1715
|
+
"server_completed_capped_turn": true,
|
|
1716
|
+
"empty_list_observed": true,
|
|
1717
|
+
"invalid": false,
|
|
1718
|
+
"passed": true
|
|
1719
|
+
},
|
|
1720
|
+
{
|
|
1721
|
+
"scenario": "truncated",
|
|
1722
|
+
"repetition": 3,
|
|
1723
|
+
"pilot": false,
|
|
1724
|
+
"exit_code": 0,
|
|
1725
|
+
"timed_out": false,
|
|
1726
|
+
"process_cleanup": true,
|
|
1727
|
+
"server_cleanup": true,
|
|
1728
|
+
"registered_server_cleanup": true,
|
|
1729
|
+
"stdout_bytes": 137521,
|
|
1730
|
+
"duration_ms": 1742.1262869999991,
|
|
1731
|
+
"proxy_integrity": true,
|
|
1732
|
+
"requests": [
|
|
1733
|
+
{
|
|
1734
|
+
"bytes": 43712,
|
|
1735
|
+
"input_bytes": 20887,
|
|
1736
|
+
"tools_bytes": 21412,
|
|
1737
|
+
"outputs": [],
|
|
1738
|
+
"newly_added_output_bytes": 0,
|
|
1739
|
+
"previous_response_id": false,
|
|
1740
|
+
"authorization_present": false
|
|
1741
|
+
},
|
|
1742
|
+
{
|
|
1743
|
+
"bytes": 44445,
|
|
1744
|
+
"input_bytes": 21620,
|
|
1745
|
+
"tools_bytes": 21412,
|
|
1746
|
+
"outputs": [
|
|
1747
|
+
{
|
|
1748
|
+
"serialized_bytes": 394,
|
|
1749
|
+
"marker_hits": 1,
|
|
1750
|
+
"error_marker_hits": 0,
|
|
1751
|
+
"plaintext_tool_echo_hits": 0,
|
|
1752
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1753
|
+
"structured_key_present": false,
|
|
1754
|
+
"tool": "pi_session_start",
|
|
1755
|
+
"call_ordinal": 1,
|
|
1756
|
+
"new_in_request": true
|
|
1757
|
+
}
|
|
1758
|
+
],
|
|
1759
|
+
"newly_added_output_bytes": 394,
|
|
1760
|
+
"previous_response_id": false,
|
|
1761
|
+
"authorization_present": false
|
|
1762
|
+
},
|
|
1763
|
+
{
|
|
1764
|
+
"bytes": 45440,
|
|
1765
|
+
"input_bytes": 22615,
|
|
1766
|
+
"tools_bytes": 21412,
|
|
1767
|
+
"outputs": [
|
|
1768
|
+
{
|
|
1769
|
+
"serialized_bytes": 394,
|
|
1770
|
+
"marker_hits": 1,
|
|
1771
|
+
"error_marker_hits": 0,
|
|
1772
|
+
"plaintext_tool_echo_hits": 0,
|
|
1773
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1774
|
+
"structured_key_present": false,
|
|
1775
|
+
"tool": "pi_session_start",
|
|
1776
|
+
"call_ordinal": 1,
|
|
1777
|
+
"new_in_request": false
|
|
1778
|
+
},
|
|
1779
|
+
{
|
|
1780
|
+
"serialized_bytes": 623,
|
|
1781
|
+
"marker_hits": 1,
|
|
1782
|
+
"error_marker_hits": 0,
|
|
1783
|
+
"plaintext_tool_echo_hits": 0,
|
|
1784
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1785
|
+
"structured_key_present": false,
|
|
1786
|
+
"tool": "pi_session_prompt",
|
|
1787
|
+
"call_ordinal": 2,
|
|
1788
|
+
"new_in_request": true
|
|
1789
|
+
}
|
|
1790
|
+
],
|
|
1791
|
+
"newly_added_output_bytes": 623,
|
|
1792
|
+
"previous_response_id": false,
|
|
1793
|
+
"authorization_present": false
|
|
1794
|
+
},
|
|
1795
|
+
{
|
|
1796
|
+
"bytes": 57868,
|
|
1797
|
+
"input_bytes": 35043,
|
|
1798
|
+
"tools_bytes": 21412,
|
|
1799
|
+
"outputs": [
|
|
1800
|
+
{
|
|
1801
|
+
"serialized_bytes": 394,
|
|
1802
|
+
"marker_hits": 1,
|
|
1803
|
+
"error_marker_hits": 0,
|
|
1804
|
+
"plaintext_tool_echo_hits": 0,
|
|
1805
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1806
|
+
"structured_key_present": false,
|
|
1807
|
+
"tool": "pi_session_start",
|
|
1808
|
+
"call_ordinal": 1,
|
|
1809
|
+
"new_in_request": false
|
|
1810
|
+
},
|
|
1811
|
+
{
|
|
1812
|
+
"serialized_bytes": 623,
|
|
1813
|
+
"marker_hits": 1,
|
|
1814
|
+
"error_marker_hits": 0,
|
|
1815
|
+
"plaintext_tool_echo_hits": 0,
|
|
1816
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1817
|
+
"structured_key_present": false,
|
|
1818
|
+
"tool": "pi_session_prompt",
|
|
1819
|
+
"call_ordinal": 2,
|
|
1820
|
+
"new_in_request": false
|
|
1821
|
+
},
|
|
1822
|
+
{
|
|
1823
|
+
"serialized_bytes": 12062,
|
|
1824
|
+
"marker_hits": 1,
|
|
1825
|
+
"error_marker_hits": 0,
|
|
1826
|
+
"plaintext_tool_echo_hits": 1,
|
|
1827
|
+
"plaintext_thinking_echo_hits": 1,
|
|
1828
|
+
"structured_key_present": false,
|
|
1829
|
+
"tool": "pi_turn_get",
|
|
1830
|
+
"call_ordinal": 3,
|
|
1831
|
+
"new_in_request": true
|
|
1832
|
+
}
|
|
1833
|
+
],
|
|
1834
|
+
"newly_added_output_bytes": 12062,
|
|
1835
|
+
"previous_response_id": false,
|
|
1836
|
+
"authorization_present": false
|
|
1837
|
+
},
|
|
1838
|
+
{
|
|
1839
|
+
"bytes": 58583,
|
|
1840
|
+
"input_bytes": 35758,
|
|
1841
|
+
"tools_bytes": 21412,
|
|
1842
|
+
"outputs": [
|
|
1843
|
+
{
|
|
1844
|
+
"serialized_bytes": 394,
|
|
1845
|
+
"marker_hits": 1,
|
|
1846
|
+
"error_marker_hits": 0,
|
|
1847
|
+
"plaintext_tool_echo_hits": 0,
|
|
1848
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1849
|
+
"structured_key_present": false,
|
|
1850
|
+
"tool": "pi_session_start",
|
|
1851
|
+
"call_ordinal": 1,
|
|
1852
|
+
"new_in_request": false
|
|
1853
|
+
},
|
|
1854
|
+
{
|
|
1855
|
+
"serialized_bytes": 623,
|
|
1856
|
+
"marker_hits": 1,
|
|
1857
|
+
"error_marker_hits": 0,
|
|
1858
|
+
"plaintext_tool_echo_hits": 0,
|
|
1859
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1860
|
+
"structured_key_present": false,
|
|
1861
|
+
"tool": "pi_session_prompt",
|
|
1862
|
+
"call_ordinal": 2,
|
|
1863
|
+
"new_in_request": false
|
|
1864
|
+
},
|
|
1865
|
+
{
|
|
1866
|
+
"serialized_bytes": 12062,
|
|
1867
|
+
"marker_hits": 1,
|
|
1868
|
+
"error_marker_hits": 0,
|
|
1869
|
+
"plaintext_tool_echo_hits": 1,
|
|
1870
|
+
"plaintext_thinking_echo_hits": 1,
|
|
1871
|
+
"structured_key_present": false,
|
|
1872
|
+
"tool": "pi_turn_get",
|
|
1873
|
+
"call_ordinal": 3,
|
|
1874
|
+
"new_in_request": false
|
|
1875
|
+
},
|
|
1876
|
+
{
|
|
1877
|
+
"serialized_bytes": 396,
|
|
1878
|
+
"marker_hits": 1,
|
|
1879
|
+
"error_marker_hits": 0,
|
|
1880
|
+
"plaintext_tool_echo_hits": 0,
|
|
1881
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1882
|
+
"structured_key_present": false,
|
|
1883
|
+
"tool": "pi_session_close",
|
|
1884
|
+
"call_ordinal": 4,
|
|
1885
|
+
"new_in_request": true
|
|
1886
|
+
}
|
|
1887
|
+
],
|
|
1888
|
+
"newly_added_output_bytes": 396,
|
|
1889
|
+
"previous_response_id": false,
|
|
1890
|
+
"authorization_present": false
|
|
1891
|
+
},
|
|
1892
|
+
{
|
|
1893
|
+
"bytes": 58914,
|
|
1894
|
+
"input_bytes": 36089,
|
|
1895
|
+
"tools_bytes": 21412,
|
|
1896
|
+
"outputs": [
|
|
1897
|
+
{
|
|
1898
|
+
"serialized_bytes": 394,
|
|
1899
|
+
"marker_hits": 1,
|
|
1900
|
+
"error_marker_hits": 0,
|
|
1901
|
+
"plaintext_tool_echo_hits": 0,
|
|
1902
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1903
|
+
"structured_key_present": false,
|
|
1904
|
+
"tool": "pi_session_start",
|
|
1905
|
+
"call_ordinal": 1,
|
|
1906
|
+
"new_in_request": false
|
|
1907
|
+
},
|
|
1908
|
+
{
|
|
1909
|
+
"serialized_bytes": 623,
|
|
1910
|
+
"marker_hits": 1,
|
|
1911
|
+
"error_marker_hits": 0,
|
|
1912
|
+
"plaintext_tool_echo_hits": 0,
|
|
1913
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1914
|
+
"structured_key_present": false,
|
|
1915
|
+
"tool": "pi_session_prompt",
|
|
1916
|
+
"call_ordinal": 2,
|
|
1917
|
+
"new_in_request": false
|
|
1918
|
+
},
|
|
1919
|
+
{
|
|
1920
|
+
"serialized_bytes": 12062,
|
|
1921
|
+
"marker_hits": 1,
|
|
1922
|
+
"error_marker_hits": 0,
|
|
1923
|
+
"plaintext_tool_echo_hits": 1,
|
|
1924
|
+
"plaintext_thinking_echo_hits": 1,
|
|
1925
|
+
"structured_key_present": false,
|
|
1926
|
+
"tool": "pi_turn_get",
|
|
1927
|
+
"call_ordinal": 3,
|
|
1928
|
+
"new_in_request": false
|
|
1929
|
+
},
|
|
1930
|
+
{
|
|
1931
|
+
"serialized_bytes": 396,
|
|
1932
|
+
"marker_hits": 1,
|
|
1933
|
+
"error_marker_hits": 0,
|
|
1934
|
+
"plaintext_tool_echo_hits": 0,
|
|
1935
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1936
|
+
"structured_key_present": false,
|
|
1937
|
+
"tool": "pi_session_close",
|
|
1938
|
+
"call_ordinal": 4,
|
|
1939
|
+
"new_in_request": false
|
|
1940
|
+
},
|
|
1941
|
+
{
|
|
1942
|
+
"serialized_bytes": 67,
|
|
1943
|
+
"marker_hits": 0,
|
|
1944
|
+
"error_marker_hits": 0,
|
|
1945
|
+
"plaintext_tool_echo_hits": 0,
|
|
1946
|
+
"plaintext_thinking_echo_hits": 0,
|
|
1947
|
+
"structured_key_present": false,
|
|
1948
|
+
"tool": "pi_session_list",
|
|
1949
|
+
"call_ordinal": 5,
|
|
1950
|
+
"new_in_request": true
|
|
1951
|
+
}
|
|
1952
|
+
],
|
|
1953
|
+
"newly_added_output_bytes": 67,
|
|
1954
|
+
"previous_response_id": false,
|
|
1955
|
+
"authorization_present": false
|
|
1956
|
+
}
|
|
1957
|
+
],
|
|
1958
|
+
"target_output": {
|
|
1959
|
+
"serialized_bytes": 12062,
|
|
1960
|
+
"marker_hits": 1,
|
|
1961
|
+
"error_marker_hits": 0,
|
|
1962
|
+
"plaintext_tool_echo_hits": 1,
|
|
1963
|
+
"plaintext_thinking_echo_hits": 1,
|
|
1964
|
+
"structured_key_present": false,
|
|
1965
|
+
"tool": "pi_turn_get",
|
|
1966
|
+
"call_ordinal": 3,
|
|
1967
|
+
"new_in_request": true
|
|
1968
|
+
},
|
|
1969
|
+
"server_target": {
|
|
1970
|
+
"wire_bytes": 131789,
|
|
1971
|
+
"content_bytes": 65891,
|
|
1972
|
+
"structured_bytes": 65830,
|
|
1973
|
+
"assistant_bytes": 65536,
|
|
1974
|
+
"state": "completed",
|
|
1975
|
+
"equal_json_representations": true
|
|
1976
|
+
},
|
|
1977
|
+
"call_schedule_matched": true,
|
|
1978
|
+
"server_representation_counts": [
|
|
1979
|
+
2
|
|
1980
|
+
],
|
|
1981
|
+
"client_completed_truncated_json_decoded": true,
|
|
1982
|
+
"session_closed_observed": true,
|
|
1983
|
+
"server_completed_capped_turn": true,
|
|
1984
|
+
"empty_list_observed": true,
|
|
1985
|
+
"invalid": false,
|
|
1986
|
+
"passed": true
|
|
1987
|
+
},
|
|
1988
|
+
{
|
|
1989
|
+
"scenario": "truncated",
|
|
1990
|
+
"repetition": 4,
|
|
1991
|
+
"pilot": false,
|
|
1992
|
+
"exit_code": 0,
|
|
1993
|
+
"timed_out": false,
|
|
1994
|
+
"process_cleanup": true,
|
|
1995
|
+
"server_cleanup": true,
|
|
1996
|
+
"registered_server_cleanup": true,
|
|
1997
|
+
"stdout_bytes": 137521,
|
|
1998
|
+
"duration_ms": 1474.219184999998,
|
|
1999
|
+
"proxy_integrity": true,
|
|
2000
|
+
"requests": [
|
|
2001
|
+
{
|
|
2002
|
+
"bytes": 43712,
|
|
2003
|
+
"input_bytes": 20887,
|
|
2004
|
+
"tools_bytes": 21412,
|
|
2005
|
+
"outputs": [],
|
|
2006
|
+
"newly_added_output_bytes": 0,
|
|
2007
|
+
"previous_response_id": false,
|
|
2008
|
+
"authorization_present": false
|
|
2009
|
+
},
|
|
2010
|
+
{
|
|
2011
|
+
"bytes": 44445,
|
|
2012
|
+
"input_bytes": 21620,
|
|
2013
|
+
"tools_bytes": 21412,
|
|
2014
|
+
"outputs": [
|
|
2015
|
+
{
|
|
2016
|
+
"serialized_bytes": 394,
|
|
2017
|
+
"marker_hits": 1,
|
|
2018
|
+
"error_marker_hits": 0,
|
|
2019
|
+
"plaintext_tool_echo_hits": 0,
|
|
2020
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2021
|
+
"structured_key_present": false,
|
|
2022
|
+
"tool": "pi_session_start",
|
|
2023
|
+
"call_ordinal": 1,
|
|
2024
|
+
"new_in_request": true
|
|
2025
|
+
}
|
|
2026
|
+
],
|
|
2027
|
+
"newly_added_output_bytes": 394,
|
|
2028
|
+
"previous_response_id": false,
|
|
2029
|
+
"authorization_present": false
|
|
2030
|
+
},
|
|
2031
|
+
{
|
|
2032
|
+
"bytes": 45440,
|
|
2033
|
+
"input_bytes": 22615,
|
|
2034
|
+
"tools_bytes": 21412,
|
|
2035
|
+
"outputs": [
|
|
2036
|
+
{
|
|
2037
|
+
"serialized_bytes": 394,
|
|
2038
|
+
"marker_hits": 1,
|
|
2039
|
+
"error_marker_hits": 0,
|
|
2040
|
+
"plaintext_tool_echo_hits": 0,
|
|
2041
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2042
|
+
"structured_key_present": false,
|
|
2043
|
+
"tool": "pi_session_start",
|
|
2044
|
+
"call_ordinal": 1,
|
|
2045
|
+
"new_in_request": false
|
|
2046
|
+
},
|
|
2047
|
+
{
|
|
2048
|
+
"serialized_bytes": 623,
|
|
2049
|
+
"marker_hits": 1,
|
|
2050
|
+
"error_marker_hits": 0,
|
|
2051
|
+
"plaintext_tool_echo_hits": 0,
|
|
2052
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2053
|
+
"structured_key_present": false,
|
|
2054
|
+
"tool": "pi_session_prompt",
|
|
2055
|
+
"call_ordinal": 2,
|
|
2056
|
+
"new_in_request": true
|
|
2057
|
+
}
|
|
2058
|
+
],
|
|
2059
|
+
"newly_added_output_bytes": 623,
|
|
2060
|
+
"previous_response_id": false,
|
|
2061
|
+
"authorization_present": false
|
|
2062
|
+
},
|
|
2063
|
+
{
|
|
2064
|
+
"bytes": 57868,
|
|
2065
|
+
"input_bytes": 35043,
|
|
2066
|
+
"tools_bytes": 21412,
|
|
2067
|
+
"outputs": [
|
|
2068
|
+
{
|
|
2069
|
+
"serialized_bytes": 394,
|
|
2070
|
+
"marker_hits": 1,
|
|
2071
|
+
"error_marker_hits": 0,
|
|
2072
|
+
"plaintext_tool_echo_hits": 0,
|
|
2073
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2074
|
+
"structured_key_present": false,
|
|
2075
|
+
"tool": "pi_session_start",
|
|
2076
|
+
"call_ordinal": 1,
|
|
2077
|
+
"new_in_request": false
|
|
2078
|
+
},
|
|
2079
|
+
{
|
|
2080
|
+
"serialized_bytes": 623,
|
|
2081
|
+
"marker_hits": 1,
|
|
2082
|
+
"error_marker_hits": 0,
|
|
2083
|
+
"plaintext_tool_echo_hits": 0,
|
|
2084
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2085
|
+
"structured_key_present": false,
|
|
2086
|
+
"tool": "pi_session_prompt",
|
|
2087
|
+
"call_ordinal": 2,
|
|
2088
|
+
"new_in_request": false
|
|
2089
|
+
},
|
|
2090
|
+
{
|
|
2091
|
+
"serialized_bytes": 12062,
|
|
2092
|
+
"marker_hits": 1,
|
|
2093
|
+
"error_marker_hits": 0,
|
|
2094
|
+
"plaintext_tool_echo_hits": 1,
|
|
2095
|
+
"plaintext_thinking_echo_hits": 1,
|
|
2096
|
+
"structured_key_present": false,
|
|
2097
|
+
"tool": "pi_turn_get",
|
|
2098
|
+
"call_ordinal": 3,
|
|
2099
|
+
"new_in_request": true
|
|
2100
|
+
}
|
|
2101
|
+
],
|
|
2102
|
+
"newly_added_output_bytes": 12062,
|
|
2103
|
+
"previous_response_id": false,
|
|
2104
|
+
"authorization_present": false
|
|
2105
|
+
},
|
|
2106
|
+
{
|
|
2107
|
+
"bytes": 58583,
|
|
2108
|
+
"input_bytes": 35758,
|
|
2109
|
+
"tools_bytes": 21412,
|
|
2110
|
+
"outputs": [
|
|
2111
|
+
{
|
|
2112
|
+
"serialized_bytes": 394,
|
|
2113
|
+
"marker_hits": 1,
|
|
2114
|
+
"error_marker_hits": 0,
|
|
2115
|
+
"plaintext_tool_echo_hits": 0,
|
|
2116
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2117
|
+
"structured_key_present": false,
|
|
2118
|
+
"tool": "pi_session_start",
|
|
2119
|
+
"call_ordinal": 1,
|
|
2120
|
+
"new_in_request": false
|
|
2121
|
+
},
|
|
2122
|
+
{
|
|
2123
|
+
"serialized_bytes": 623,
|
|
2124
|
+
"marker_hits": 1,
|
|
2125
|
+
"error_marker_hits": 0,
|
|
2126
|
+
"plaintext_tool_echo_hits": 0,
|
|
2127
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2128
|
+
"structured_key_present": false,
|
|
2129
|
+
"tool": "pi_session_prompt",
|
|
2130
|
+
"call_ordinal": 2,
|
|
2131
|
+
"new_in_request": false
|
|
2132
|
+
},
|
|
2133
|
+
{
|
|
2134
|
+
"serialized_bytes": 12062,
|
|
2135
|
+
"marker_hits": 1,
|
|
2136
|
+
"error_marker_hits": 0,
|
|
2137
|
+
"plaintext_tool_echo_hits": 1,
|
|
2138
|
+
"plaintext_thinking_echo_hits": 1,
|
|
2139
|
+
"structured_key_present": false,
|
|
2140
|
+
"tool": "pi_turn_get",
|
|
2141
|
+
"call_ordinal": 3,
|
|
2142
|
+
"new_in_request": false
|
|
2143
|
+
},
|
|
2144
|
+
{
|
|
2145
|
+
"serialized_bytes": 396,
|
|
2146
|
+
"marker_hits": 1,
|
|
2147
|
+
"error_marker_hits": 0,
|
|
2148
|
+
"plaintext_tool_echo_hits": 0,
|
|
2149
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2150
|
+
"structured_key_present": false,
|
|
2151
|
+
"tool": "pi_session_close",
|
|
2152
|
+
"call_ordinal": 4,
|
|
2153
|
+
"new_in_request": true
|
|
2154
|
+
}
|
|
2155
|
+
],
|
|
2156
|
+
"newly_added_output_bytes": 396,
|
|
2157
|
+
"previous_response_id": false,
|
|
2158
|
+
"authorization_present": false
|
|
2159
|
+
},
|
|
2160
|
+
{
|
|
2161
|
+
"bytes": 58914,
|
|
2162
|
+
"input_bytes": 36089,
|
|
2163
|
+
"tools_bytes": 21412,
|
|
2164
|
+
"outputs": [
|
|
2165
|
+
{
|
|
2166
|
+
"serialized_bytes": 394,
|
|
2167
|
+
"marker_hits": 1,
|
|
2168
|
+
"error_marker_hits": 0,
|
|
2169
|
+
"plaintext_tool_echo_hits": 0,
|
|
2170
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2171
|
+
"structured_key_present": false,
|
|
2172
|
+
"tool": "pi_session_start",
|
|
2173
|
+
"call_ordinal": 1,
|
|
2174
|
+
"new_in_request": false
|
|
2175
|
+
},
|
|
2176
|
+
{
|
|
2177
|
+
"serialized_bytes": 623,
|
|
2178
|
+
"marker_hits": 1,
|
|
2179
|
+
"error_marker_hits": 0,
|
|
2180
|
+
"plaintext_tool_echo_hits": 0,
|
|
2181
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2182
|
+
"structured_key_present": false,
|
|
2183
|
+
"tool": "pi_session_prompt",
|
|
2184
|
+
"call_ordinal": 2,
|
|
2185
|
+
"new_in_request": false
|
|
2186
|
+
},
|
|
2187
|
+
{
|
|
2188
|
+
"serialized_bytes": 12062,
|
|
2189
|
+
"marker_hits": 1,
|
|
2190
|
+
"error_marker_hits": 0,
|
|
2191
|
+
"plaintext_tool_echo_hits": 1,
|
|
2192
|
+
"plaintext_thinking_echo_hits": 1,
|
|
2193
|
+
"structured_key_present": false,
|
|
2194
|
+
"tool": "pi_turn_get",
|
|
2195
|
+
"call_ordinal": 3,
|
|
2196
|
+
"new_in_request": false
|
|
2197
|
+
},
|
|
2198
|
+
{
|
|
2199
|
+
"serialized_bytes": 396,
|
|
2200
|
+
"marker_hits": 1,
|
|
2201
|
+
"error_marker_hits": 0,
|
|
2202
|
+
"plaintext_tool_echo_hits": 0,
|
|
2203
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2204
|
+
"structured_key_present": false,
|
|
2205
|
+
"tool": "pi_session_close",
|
|
2206
|
+
"call_ordinal": 4,
|
|
2207
|
+
"new_in_request": false
|
|
2208
|
+
},
|
|
2209
|
+
{
|
|
2210
|
+
"serialized_bytes": 67,
|
|
2211
|
+
"marker_hits": 0,
|
|
2212
|
+
"error_marker_hits": 0,
|
|
2213
|
+
"plaintext_tool_echo_hits": 0,
|
|
2214
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2215
|
+
"structured_key_present": false,
|
|
2216
|
+
"tool": "pi_session_list",
|
|
2217
|
+
"call_ordinal": 5,
|
|
2218
|
+
"new_in_request": true
|
|
2219
|
+
}
|
|
2220
|
+
],
|
|
2221
|
+
"newly_added_output_bytes": 67,
|
|
2222
|
+
"previous_response_id": false,
|
|
2223
|
+
"authorization_present": false
|
|
2224
|
+
}
|
|
2225
|
+
],
|
|
2226
|
+
"target_output": {
|
|
2227
|
+
"serialized_bytes": 12062,
|
|
2228
|
+
"marker_hits": 1,
|
|
2229
|
+
"error_marker_hits": 0,
|
|
2230
|
+
"plaintext_tool_echo_hits": 1,
|
|
2231
|
+
"plaintext_thinking_echo_hits": 1,
|
|
2232
|
+
"structured_key_present": false,
|
|
2233
|
+
"tool": "pi_turn_get",
|
|
2234
|
+
"call_ordinal": 3,
|
|
2235
|
+
"new_in_request": true
|
|
2236
|
+
},
|
|
2237
|
+
"server_target": {
|
|
2238
|
+
"wire_bytes": 131789,
|
|
2239
|
+
"content_bytes": 65891,
|
|
2240
|
+
"structured_bytes": 65830,
|
|
2241
|
+
"assistant_bytes": 65536,
|
|
2242
|
+
"state": "completed",
|
|
2243
|
+
"equal_json_representations": true
|
|
2244
|
+
},
|
|
2245
|
+
"call_schedule_matched": true,
|
|
2246
|
+
"server_representation_counts": [
|
|
2247
|
+
2
|
|
2248
|
+
],
|
|
2249
|
+
"client_completed_truncated_json_decoded": true,
|
|
2250
|
+
"session_closed_observed": true,
|
|
2251
|
+
"server_completed_capped_turn": true,
|
|
2252
|
+
"empty_list_observed": true,
|
|
2253
|
+
"invalid": false,
|
|
2254
|
+
"passed": true
|
|
2255
|
+
},
|
|
2256
|
+
{
|
|
2257
|
+
"scenario": "truncated",
|
|
2258
|
+
"repetition": 5,
|
|
2259
|
+
"pilot": false,
|
|
2260
|
+
"exit_code": 0,
|
|
2261
|
+
"timed_out": false,
|
|
2262
|
+
"process_cleanup": true,
|
|
2263
|
+
"server_cleanup": true,
|
|
2264
|
+
"registered_server_cleanup": true,
|
|
2265
|
+
"stdout_bytes": 137521,
|
|
2266
|
+
"duration_ms": 1535.5659919999998,
|
|
2267
|
+
"proxy_integrity": true,
|
|
2268
|
+
"requests": [
|
|
2269
|
+
{
|
|
2270
|
+
"bytes": 43712,
|
|
2271
|
+
"input_bytes": 20887,
|
|
2272
|
+
"tools_bytes": 21412,
|
|
2273
|
+
"outputs": [],
|
|
2274
|
+
"newly_added_output_bytes": 0,
|
|
2275
|
+
"previous_response_id": false,
|
|
2276
|
+
"authorization_present": false
|
|
2277
|
+
},
|
|
2278
|
+
{
|
|
2279
|
+
"bytes": 44445,
|
|
2280
|
+
"input_bytes": 21620,
|
|
2281
|
+
"tools_bytes": 21412,
|
|
2282
|
+
"outputs": [
|
|
2283
|
+
{
|
|
2284
|
+
"serialized_bytes": 394,
|
|
2285
|
+
"marker_hits": 1,
|
|
2286
|
+
"error_marker_hits": 0,
|
|
2287
|
+
"plaintext_tool_echo_hits": 0,
|
|
2288
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2289
|
+
"structured_key_present": false,
|
|
2290
|
+
"tool": "pi_session_start",
|
|
2291
|
+
"call_ordinal": 1,
|
|
2292
|
+
"new_in_request": true
|
|
2293
|
+
}
|
|
2294
|
+
],
|
|
2295
|
+
"newly_added_output_bytes": 394,
|
|
2296
|
+
"previous_response_id": false,
|
|
2297
|
+
"authorization_present": false
|
|
2298
|
+
},
|
|
2299
|
+
{
|
|
2300
|
+
"bytes": 45440,
|
|
2301
|
+
"input_bytes": 22615,
|
|
2302
|
+
"tools_bytes": 21412,
|
|
2303
|
+
"outputs": [
|
|
2304
|
+
{
|
|
2305
|
+
"serialized_bytes": 394,
|
|
2306
|
+
"marker_hits": 1,
|
|
2307
|
+
"error_marker_hits": 0,
|
|
2308
|
+
"plaintext_tool_echo_hits": 0,
|
|
2309
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2310
|
+
"structured_key_present": false,
|
|
2311
|
+
"tool": "pi_session_start",
|
|
2312
|
+
"call_ordinal": 1,
|
|
2313
|
+
"new_in_request": false
|
|
2314
|
+
},
|
|
2315
|
+
{
|
|
2316
|
+
"serialized_bytes": 623,
|
|
2317
|
+
"marker_hits": 1,
|
|
2318
|
+
"error_marker_hits": 0,
|
|
2319
|
+
"plaintext_tool_echo_hits": 0,
|
|
2320
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2321
|
+
"structured_key_present": false,
|
|
2322
|
+
"tool": "pi_session_prompt",
|
|
2323
|
+
"call_ordinal": 2,
|
|
2324
|
+
"new_in_request": true
|
|
2325
|
+
}
|
|
2326
|
+
],
|
|
2327
|
+
"newly_added_output_bytes": 623,
|
|
2328
|
+
"previous_response_id": false,
|
|
2329
|
+
"authorization_present": false
|
|
2330
|
+
},
|
|
2331
|
+
{
|
|
2332
|
+
"bytes": 57868,
|
|
2333
|
+
"input_bytes": 35043,
|
|
2334
|
+
"tools_bytes": 21412,
|
|
2335
|
+
"outputs": [
|
|
2336
|
+
{
|
|
2337
|
+
"serialized_bytes": 394,
|
|
2338
|
+
"marker_hits": 1,
|
|
2339
|
+
"error_marker_hits": 0,
|
|
2340
|
+
"plaintext_tool_echo_hits": 0,
|
|
2341
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2342
|
+
"structured_key_present": false,
|
|
2343
|
+
"tool": "pi_session_start",
|
|
2344
|
+
"call_ordinal": 1,
|
|
2345
|
+
"new_in_request": false
|
|
2346
|
+
},
|
|
2347
|
+
{
|
|
2348
|
+
"serialized_bytes": 623,
|
|
2349
|
+
"marker_hits": 1,
|
|
2350
|
+
"error_marker_hits": 0,
|
|
2351
|
+
"plaintext_tool_echo_hits": 0,
|
|
2352
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2353
|
+
"structured_key_present": false,
|
|
2354
|
+
"tool": "pi_session_prompt",
|
|
2355
|
+
"call_ordinal": 2,
|
|
2356
|
+
"new_in_request": false
|
|
2357
|
+
},
|
|
2358
|
+
{
|
|
2359
|
+
"serialized_bytes": 12062,
|
|
2360
|
+
"marker_hits": 1,
|
|
2361
|
+
"error_marker_hits": 0,
|
|
2362
|
+
"plaintext_tool_echo_hits": 1,
|
|
2363
|
+
"plaintext_thinking_echo_hits": 1,
|
|
2364
|
+
"structured_key_present": false,
|
|
2365
|
+
"tool": "pi_turn_get",
|
|
2366
|
+
"call_ordinal": 3,
|
|
2367
|
+
"new_in_request": true
|
|
2368
|
+
}
|
|
2369
|
+
],
|
|
2370
|
+
"newly_added_output_bytes": 12062,
|
|
2371
|
+
"previous_response_id": false,
|
|
2372
|
+
"authorization_present": false
|
|
2373
|
+
},
|
|
2374
|
+
{
|
|
2375
|
+
"bytes": 58583,
|
|
2376
|
+
"input_bytes": 35758,
|
|
2377
|
+
"tools_bytes": 21412,
|
|
2378
|
+
"outputs": [
|
|
2379
|
+
{
|
|
2380
|
+
"serialized_bytes": 394,
|
|
2381
|
+
"marker_hits": 1,
|
|
2382
|
+
"error_marker_hits": 0,
|
|
2383
|
+
"plaintext_tool_echo_hits": 0,
|
|
2384
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2385
|
+
"structured_key_present": false,
|
|
2386
|
+
"tool": "pi_session_start",
|
|
2387
|
+
"call_ordinal": 1,
|
|
2388
|
+
"new_in_request": false
|
|
2389
|
+
},
|
|
2390
|
+
{
|
|
2391
|
+
"serialized_bytes": 623,
|
|
2392
|
+
"marker_hits": 1,
|
|
2393
|
+
"error_marker_hits": 0,
|
|
2394
|
+
"plaintext_tool_echo_hits": 0,
|
|
2395
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2396
|
+
"structured_key_present": false,
|
|
2397
|
+
"tool": "pi_session_prompt",
|
|
2398
|
+
"call_ordinal": 2,
|
|
2399
|
+
"new_in_request": false
|
|
2400
|
+
},
|
|
2401
|
+
{
|
|
2402
|
+
"serialized_bytes": 12062,
|
|
2403
|
+
"marker_hits": 1,
|
|
2404
|
+
"error_marker_hits": 0,
|
|
2405
|
+
"plaintext_tool_echo_hits": 1,
|
|
2406
|
+
"plaintext_thinking_echo_hits": 1,
|
|
2407
|
+
"structured_key_present": false,
|
|
2408
|
+
"tool": "pi_turn_get",
|
|
2409
|
+
"call_ordinal": 3,
|
|
2410
|
+
"new_in_request": false
|
|
2411
|
+
},
|
|
2412
|
+
{
|
|
2413
|
+
"serialized_bytes": 396,
|
|
2414
|
+
"marker_hits": 1,
|
|
2415
|
+
"error_marker_hits": 0,
|
|
2416
|
+
"plaintext_tool_echo_hits": 0,
|
|
2417
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2418
|
+
"structured_key_present": false,
|
|
2419
|
+
"tool": "pi_session_close",
|
|
2420
|
+
"call_ordinal": 4,
|
|
2421
|
+
"new_in_request": true
|
|
2422
|
+
}
|
|
2423
|
+
],
|
|
2424
|
+
"newly_added_output_bytes": 396,
|
|
2425
|
+
"previous_response_id": false,
|
|
2426
|
+
"authorization_present": false
|
|
2427
|
+
},
|
|
2428
|
+
{
|
|
2429
|
+
"bytes": 58914,
|
|
2430
|
+
"input_bytes": 36089,
|
|
2431
|
+
"tools_bytes": 21412,
|
|
2432
|
+
"outputs": [
|
|
2433
|
+
{
|
|
2434
|
+
"serialized_bytes": 394,
|
|
2435
|
+
"marker_hits": 1,
|
|
2436
|
+
"error_marker_hits": 0,
|
|
2437
|
+
"plaintext_tool_echo_hits": 0,
|
|
2438
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2439
|
+
"structured_key_present": false,
|
|
2440
|
+
"tool": "pi_session_start",
|
|
2441
|
+
"call_ordinal": 1,
|
|
2442
|
+
"new_in_request": false
|
|
2443
|
+
},
|
|
2444
|
+
{
|
|
2445
|
+
"serialized_bytes": 623,
|
|
2446
|
+
"marker_hits": 1,
|
|
2447
|
+
"error_marker_hits": 0,
|
|
2448
|
+
"plaintext_tool_echo_hits": 0,
|
|
2449
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2450
|
+
"structured_key_present": false,
|
|
2451
|
+
"tool": "pi_session_prompt",
|
|
2452
|
+
"call_ordinal": 2,
|
|
2453
|
+
"new_in_request": false
|
|
2454
|
+
},
|
|
2455
|
+
{
|
|
2456
|
+
"serialized_bytes": 12062,
|
|
2457
|
+
"marker_hits": 1,
|
|
2458
|
+
"error_marker_hits": 0,
|
|
2459
|
+
"plaintext_tool_echo_hits": 1,
|
|
2460
|
+
"plaintext_thinking_echo_hits": 1,
|
|
2461
|
+
"structured_key_present": false,
|
|
2462
|
+
"tool": "pi_turn_get",
|
|
2463
|
+
"call_ordinal": 3,
|
|
2464
|
+
"new_in_request": false
|
|
2465
|
+
},
|
|
2466
|
+
{
|
|
2467
|
+
"serialized_bytes": 396,
|
|
2468
|
+
"marker_hits": 1,
|
|
2469
|
+
"error_marker_hits": 0,
|
|
2470
|
+
"plaintext_tool_echo_hits": 0,
|
|
2471
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2472
|
+
"structured_key_present": false,
|
|
2473
|
+
"tool": "pi_session_close",
|
|
2474
|
+
"call_ordinal": 4,
|
|
2475
|
+
"new_in_request": false
|
|
2476
|
+
},
|
|
2477
|
+
{
|
|
2478
|
+
"serialized_bytes": 67,
|
|
2479
|
+
"marker_hits": 0,
|
|
2480
|
+
"error_marker_hits": 0,
|
|
2481
|
+
"plaintext_tool_echo_hits": 0,
|
|
2482
|
+
"plaintext_thinking_echo_hits": 0,
|
|
2483
|
+
"structured_key_present": false,
|
|
2484
|
+
"tool": "pi_session_list",
|
|
2485
|
+
"call_ordinal": 5,
|
|
2486
|
+
"new_in_request": true
|
|
2487
|
+
}
|
|
2488
|
+
],
|
|
2489
|
+
"newly_added_output_bytes": 67,
|
|
2490
|
+
"previous_response_id": false,
|
|
2491
|
+
"authorization_present": false
|
|
2492
|
+
}
|
|
2493
|
+
],
|
|
2494
|
+
"target_output": {
|
|
2495
|
+
"serialized_bytes": 12062,
|
|
2496
|
+
"marker_hits": 1,
|
|
2497
|
+
"error_marker_hits": 0,
|
|
2498
|
+
"plaintext_tool_echo_hits": 1,
|
|
2499
|
+
"plaintext_thinking_echo_hits": 1,
|
|
2500
|
+
"structured_key_present": false,
|
|
2501
|
+
"tool": "pi_turn_get",
|
|
2502
|
+
"call_ordinal": 3,
|
|
2503
|
+
"new_in_request": true
|
|
2504
|
+
},
|
|
2505
|
+
"server_target": {
|
|
2506
|
+
"wire_bytes": 131789,
|
|
2507
|
+
"content_bytes": 65891,
|
|
2508
|
+
"structured_bytes": 65830,
|
|
2509
|
+
"assistant_bytes": 65536,
|
|
2510
|
+
"state": "completed",
|
|
2511
|
+
"equal_json_representations": true
|
|
2512
|
+
},
|
|
2513
|
+
"call_schedule_matched": true,
|
|
2514
|
+
"server_representation_counts": [
|
|
2515
|
+
2
|
|
2516
|
+
],
|
|
2517
|
+
"client_completed_truncated_json_decoded": true,
|
|
2518
|
+
"session_closed_observed": true,
|
|
2519
|
+
"server_completed_capped_turn": true,
|
|
2520
|
+
"empty_list_observed": true,
|
|
2521
|
+
"invalid": false,
|
|
2522
|
+
"passed": true
|
|
2523
|
+
}
|
|
2524
|
+
],
|
|
2525
|
+
"model_usage": null,
|
|
2526
|
+
"money": null,
|
|
2527
|
+
"real_cloud_projection": null,
|
|
2528
|
+
"os_egress_containment": null,
|
|
2529
|
+
"metric_semantics": {
|
|
2530
|
+
"bytes": "direct-observation-of-serialized-utf8",
|
|
2531
|
+
"marker_hits": "occurrences-in-selected-function-output",
|
|
2532
|
+
"newly_added_output_bytes": "first-observed-call-id-output-serialization",
|
|
2533
|
+
"request_bytes": "whole-http-json-body-not-token-usage",
|
|
2534
|
+
"model_usage": "unknown-no-model-executed",
|
|
2535
|
+
"money": "unknown-no-meter-or-tariff"
|
|
2536
|
+
},
|
|
2537
|
+
"harness_status": "PASS"
|
|
2538
|
+
}
|