Integration Verification

Use this page when deciding whether an agent integration is truly ready to claim publicly.

Evidence Levels

Evidence	Counts as	Does not count as
lerim connect <agent> --mode mcp --dry-run	Safe preview of the config write.	Installed-client support.
lerim connect <agent> --mode mcp	Config writer can mutate the target file and create a backup.	A live agent tool call.
lerim connect doctor <agent>	Lerim can parse the config and see the MCP entry.	Proof that the client loaded or used the tool.
Local stdio MCP tool-call probe	Lerim's MCP server can execute lerim_context_brief through the MCP protocol.	Proof that an external client selected the tool.
Local stdio trace-submit duplicate probe	Lerim's MCP server can accept, normalize, and idempotently skip a duplicate submitted trace.	Proof that LLM extraction quality passed.
Opt-in local stdio trace-submit extraction probe	When accepted, a synthetic submitted trace ran through the DSPy extraction path and created records.	Proof that an external installed client submitted an organic trace; the current public MCP artifact has 0 extraction acceptances.
Installed client CLI list/get command	The installed client can see the Lerim MCP entry.	Proof that lerim_context_brief was called.
Live tool-call probe	The client called a Lerim MCP context tool.	Native completed-session capture.
Native adapter ingest over a completed session	Native capture for that source shape.	MCP recall support.

Standard Lerim Checks

Run these for every target:

lerim connect <agent> --mode mcp --dry-run
lerim connect <agent> --mode mcp
lerim connect doctor <agent>

doctor should report:

• config exists
• configured is true
• no parse error

If the target config already existed, lerim connect should create a timestamped backup next to it.

Installed Client Checks

The benchmark runner also calls lerim_context_brief directly through Lerim's local stdio MCP server by default. That is a useful server-path check, but keep it separate from installed-client acceptance.

These commands are currently useful where the client exposes an MCP CLI:

Client	Command	Strongest current signal
Codex CLI	codex mcp get lerim	Config visible; no connected marker exposed by this probe.
Claude Code	claude mcp get lerim	Status: Connected.
Gemini CLI	gemini mcp list	lerim listed as connected.
OpenCode	opencode mcp list	lerim connected.

For clients without a CLI list/get command, use config doctor evidence and mark the row as config-writer tested until an installed-client probe exists. Public benchmark artifacts keep aggregate installed-client counts but omit per-machine client inventory from details.jsonl.

Live Tool-Call Probes

Live probes can spend model or subscription budget, so keep them opt-in. The current benchmark runner has an explicit gate:

uv run python benchmarks/lerim_evidence/integration.py \
  --include-real-doctor \
  --include-installed-client-probes \
  --include-tool-call-probes \
  --allow-live-client-tool-calls \
  --tool-call-targets gemini-cli \
  --output-dir <private-output-dir>

The live tool-call acceptance target is simple: the installed client should call lerim_context_brief and return the probe marker. Use --tool-call-targets to rerun only the client whose auth, quota, or config changed.

Current public live tool-call evidence:

Client	Artifact	Result
Gemini CLI	benchmarks/results/raw/mcp-gemini-live-tool-call/report.json	Connected and called lerim_context_brief through the installed client.

See MCP Integration for the current public evidence boundary.

Do not turn a skipped live probe into a support claim.

Native Capture Checks

For Claude Code, Codex CLI, Cursor, OpenCode, and pi, native adapter support should be verified with completed local sessions:

lerim ingest --agent <agent> --no-extract
lerim queue --status pending
lerim ingest --agent <agent>
lerim context records --profile coding

Fixtures and unit tests are useful, but they do not prove that the current local client version still stores sessions in the expected shape.

Troubleshooting

• If doctor reports a parse error, inspect the target config file before running another write.
• If the client cannot start the MCP server, rerun lerim connect <agent> --mode mcp so Lerim writes the absolute Python executable plus -m lerim.mcp_server. For manual configs, set command to an absolute Python path and args to ["-m", "lerim.mcp_server"]; do not point the client command at the lerim executable.
• If a VS Code agent cannot find its config, confirm the active VS Code profile and global storage path.
• If an installed-client probe passes but live tool calls fail, treat support as config-visible only.
• If an MCP-first agent has no stable session export, mark it recall-only unless the agent explicitly calls lerim_trace_submit.