Coverage matrix

September 16, 2026 · View on GitHub

Generated from the capability manifest; last reviewed 2026-07-17. Integration breadth is not uniform verification: a client can have a proven usage lane and an experimental capture lane at the same time. Each lane is rated on its own, and the states are never flattened into one "supported" badge.

States

  • verified — works, with real, dated evidence.
  • verified_partial — works, with real evidence but a narrow, bounded scope.
  • experimental — implemented, but the evidence is synthetic or absent.
  • (unavailable) — not implemented.

Lanes

  • Sessions — session discovery.
  • Usage — usage / cost import.
  • Mechanical — mechanical activity capture (commands, files, tool categories).
  • MCP — MCP tool semantics (sections, checks, files).
  • Model — model attribution.
  • Cache R — cache-read token capture.
  • Cache W — cache-creation token capture.
  • Install — one-command install.
AgentSessionsUsageMechanicalMCPModelCache RCache WInstall
Claude Codeverified_partialverified_partialexperimentalverifiedexperimentalexperimentalexperimentalexperimental
Codexverified_partialverified_partialexperimentalverifiedexperimentalexperimentalexperimentalexperimental
Hermesexperimentalverified_partialexperimentalverified_partialexperimentalexperimentalexperimentalexperimental
OpenCodeexperimentalexperimentalexperimentalverified_partialexperimentalexperimentalexperimentalexperimental
OpenClawexperimentalexperimentalverified_partialexperimentalexperimentalexperimental
DeepSeek Harnessexperimentalexperimentalverified_partialexperimentalexperimentalexperimentalverified_partial
Cursorverified_partialexperimentalexperimentalverified_partial
Gemini CLI
GitHub Copilot CLI
Cline
Windsurf
Aider

Per-agent detail

Claude Code

  • Collection path: projects JSONL
  • Session scope: Root and child transcripts, including sessions with no token row yet.
  • Usage / cost basis: client_reported / unknown
  • Validation scope: single_machine_live_observation (2026-07-17)
  • Exact work attribution still requires client-authored identifiers on the recording call.

Codex

  • Collection path: state_5.sqlite, rollout JSONL
  • Session scope: State database and rollout-only root/child sessions, including zero-token observations.
  • Usage / cost basis: client_reported / unknown
  • Validation scope: single_machine_live_observation (2026-07-17)
  • Usage and semantic work remain separate unless client-log evidence proves the join.

Hermes

  • Collection path: state.db sessions
  • Session scope: Modeled state.db session rows, including rows with no positive usage yet.
  • Usage / cost basis: client_reported / unknown
  • Validation scope: single_machine_live_observation (2026-07-17)
  • Remain provisional until schema-drift recovery and zero-token observation have live evidence.

OpenCode

  • Collection path: native opencode.db SQLite session rollup, captured/exported JSON event stream
  • Session scope: Usage-bearing session rollups (SQLite) or step-finish exports.
  • Usage / cost basis: client_reported / estimated_from_tokens
  • Validation scope: dated_mcp_smoke_only (2026-06-28)
  • Remain experimental until real-client fixtures, namespace hardening, and per-message SQLite granularity exist.

OpenClaw

  • Collection path: assistant-message JSONL
  • Session scope: Usage-bearing assistant rows in session log files only.
  • Usage / cost basis: client_reported / client_reported
  • Validation scope: dated_mcp_smoke_only (2026-06-28)
  • Remain experimental until real fixtures, namespace hardening, routing, and per-model lanes exist.

DeepSeek Harness

  • Collection path: Zstandard-compressed assistant/message JSONL
  • Session scope: Usage-bearing assistant/message events in per-session dsh logs.
  • Usage / cost basis: client_reported / unknown
  • Validation scope: single_machine_live_observation (2026-09-16)
  • Usage-import lanes remain synthetic-fixture experimental until a live dsh usage-log smoke and namespace hardening exist; the MCP self-reporting lanes are verified on one machine/version.

Cursor

  • Collection path: primary User/globalStorage/state.vscdb cursorDiskKV composerData rows, manually wired hook payloads
  • Session scope: Metadata-only root/child composer observations; no prompt, message, title, token, or cost fields.
  • Usage / cost basis: unknown / unknown
  • Validation scope: single_machine_live_observation (2026-07-17)
  • Activity observations never fabricate token usage, cache usage, cost, titles, projects, or semantic work descriptions.

Roadmap only (no implementation or verification yet): Gemini CLI, GitHub Copilot CLI, Cline, Windsurf, Aider.

Refresh semantics (import-once vs --refresh, watcher heartbeat, snapshot TTL) and the full per-path usage/cost truth are in the usage & cost truth table. The dated evidence behind each verified* rating is in adapter capability evidence. Live source detection and importer health are separate from this static manifest — see agentacct capabilities agents for the machine-local reading.