Adapter capability evidence

This page records bounded evidence referenced by the machine-readable agent capability manifest. A result verifies only the lane named here; it does not turn an agent into a binary “supported” client.

2026-07-17 local import and Dashboard observation

The post-P5 private-build checkpoint ran a real local import against the maintainer's existing client stores, then restarted the ownership-safe local Dashboard and inspected the resulting API and UI.

Observed import result:

• 1,414 client sessions observed.
• 1,409 sessions had client-reported usage.
• 5 real sessions had no usage row and remained visible as observation-only sessions.
• Claude Code: 699 observed sessions, 695 usage sessions, 4 without usage, no current ingestion errors.
• Codex: 682 observed sessions, 681 usage sessions, 1 without usage, no current ingestion errors.
• Hermes: 33 usage-bearing sessions remained present after refresh.

Observed product result:

• The Dashboard health endpoint returned HTTP 200 after an ownership-safe restart.
• The ingestion health endpoint reported healthy with no current issues.
• An observation-only Task rendered token, cache, and cost fields as Unavailable; it did not manufacture zero usage or zero cost.
• The usage summary kept 651 non-additive Codex replay-like rows out of additive totals while preserving them for inspection.

This is a live observation of the declared local import paths on one machine, not a cross-version stability guarantee. The manifest therefore uses verified_partial, not a whole-client stable or supported claim. Hermes remains explicitly limited because only usage-bearing rows have live-store evidence; zero-token observation and multi-home fail-closed behavior are deterministic-fixture verified, while schema-drift recovery remains provisional.

Evidence boundaries

• Hermes, OpenCode, and OpenClaw MCP verification is recorded separately in Coding agent integrations.
• Synthetic parser fixtures prove deterministic field handling only. They do not prove current client schema compatibility, installation, or runtime health.
• Runtime presence belongs to /usage/sources; latest importer health belongs to /ingestion/health; historical evidence belongs to /evidence/product.

2026-07-17 Cursor 3.9.16 primary-state observation

A read-only dry run exercised the primary Cursor 3.9.16 store on the maintainer machine while Cursor had an empty regular WAL sidecar and a regular 32 KB SHM sidecar.

Observed result:

• 35 composerData:* sessions discovered.
• 20 recent root groups selected and 20 session observations returned.
• 0 usage events and 20 sessions with usage unavailable.
• 5 explicit model labels retained; missing/default labels remained unattributed.
• 0 titles and 0 working directories retained.
• 0 scan errors.
• An isolated-store write persisted 20 session_observed events carrying no usage, token, cache, or cost fields.
• Repeating the import wrote 0 new events and preserved the same 20 observations, proving the bounded path is idempotent.

This verifies the bounded primary-state session-discovery and explicit-model lanes on one Cursor version. It does not verify token, cache, or cost import, and it is not a multi-version stability claim.

2026-07-17 Cursor observation-only fixture boundary

Deterministic fixtures exercise the primary cursorDiskKV composer path: metadata-only root/child observation, root-group limiting, source namespaces, id/schema/graph validation, symlink rejection, active-WAL rejection, corrupt/replaced database handling, and source-discovery/import parity. The fixtures also prove that usage, cache, cost, prompt/message/title, cwd, and project data do not enter the produced observation.

Fixtures remain the evidence for malformed/schema/path/lineage failure handling. The separate Cursor 3.9.16 smoke above verifies the positive session-discovery and explicit-model lanes on one real version, so the manifest marks only those bounded capabilities verified_partial with single_machine_live_observation stability. Usage, cache, cost, automatic installation, and multi-version stability remain unavailable or unverified as declared in the manifest.

2026-09-16 DeepSeek Harness (dsh) 0.1.5-rc.1 MCP self-reporting smoke

A real dsh 0.1.5-rc.1 headless session, configured only by agentacct onboard --agent dsh (the home patch $DSH_HOME/cordis.patch.yml plus $DSH_HOME/AGENTS.md), loaded the bundled @deepseek-ai/dsh-mcp-client plugin in-box — no dsh plugin --profile <name> add step was needed — launched agentacct's MCP server, and called agentacct_record_section (started, then completed). The isolated agentacct store recorded both events with source: dsh and the requested section id.

This verifies the positive MCP self-reporting lane end to end — home-patch plugin resolution, server launch, tool availability, and the record pipeline — on one real version and one machine, with the recording task explicitly requested. It is not a cross-version stability guarantee and does not verify spontaneous (uninstructed) recording. The manifest therefore marks the dsh mcp_semantics and automatic_install lanes verified_partial with single_machine_live_observation stability. The usage-import lanes (session-log parsing) remain synthetic-fixture experimental: dsh writes Zstandard-compressed JSONL session logs whose parsing is verified against source and third-party parsers only, with no live-client log smoke yet.