feat(claude-code): local Claude Code connector with full agent traces

[volod-vana]

Why

A buyer conversation (Semo AI) asked for agent trajectories — the action-observation loop of real runs — not chat transcripts. Claude Code already writes exactly that to ~/.claude/projects/*.jsonl. This publishes the local Claude Code connector and adds the scope that carries it.

claude_code.sessions deliberately carries only the shape of a run (turn counts, models, timing). claude_code.trajectories carries the bodies.

Scopes

Scope	What
claude_code.usage	aggregate tokens, cost, day/hour activity
claude_code.sessions	per-session metadata, no bodies; now also lists subagent runs linked to their parent
claude_code.prompts	typed prompt history
claude_code.trajectories	new — prompts, assistant reasoning and replies, every tool call with its input, every tool result with its output, in order

Tool calls and results are linked by callId, so a consumer can rebuild the loop without reparsing the transcript. Steps carry role, type, at, model, and sidechain.

Why a separate scope: these payloads contain source code, file paths and command output from the machine. Its own scope lets a user grant the cheap metadata without ever granting the transcript, and lets a host gate the two differently.

Safety rails on the trajectory scope

• Redaction before export: Anthropic/OpenAI keys, GitHub tokens and PATs, Slack, AWS, Google, JWTs, bearer tokens, PEM private-key blocks, 0x+64-hex, and SECRET/TOKEN/PASSWORD-style assignments. The count is reported by type; the matched text never is.
• Per-step cap CLAUDE_CODE_MAX_STEP_CHARS (default 8000), each cut body marks itself truncated with its original length.
• Whole-scope budget CLAUDE_CODE_MAX_MB (default 24), newest sessions first.
• Both caps surface as degraded honest-telemetry errors, so a capped export can never be mistaken for a complete one.

Redaction is best-effort pattern matching, not a de-identification pipeline. It is the floor, not the guarantee. Anything sold or shared still needs a real privacy pass.

Two fixes that came out of building it

• Subagent traces were invisible. The old walk only looked at <project>/<sessionId>.jsonl and missed <session>/subagents/agent-*.jsonl. On my machine that was 144 of 310 traces. The walk now recurses and tags each transcript with kind and parentSessionId.
• The result envelope was not conformant. It shipped no requestedScopes, no errors array, and a string exportSummary.details. Now it satisfies validate-honest-telemetry-conformance, and scopes are gated on requestedScopes so a host asking for metadata alone never pays to build traces.
• schemas/scope-catalog.schema.json only accepted *-playwright.json in its provenance list. The manifest schema has always allowed runtime: vanilla, so the filename pattern was the stale half.

Verification (real run, this machine)

node run-connector.cjs connectors/anthropic/claude-code-local.js --runner-dir <desktop runner>

• 310 sessions (166 main + 144 subagent), 26,036 trace steps, 10,354 tool calls
• 1,539 redactions (secret_assignment 1077, hex_secret 429, bearer 16, jwt 10, anthropic_key 3, github_token 2, google_key 2)
• 24 MB at the default budget; 217 older sessions skipped and 358 bodies truncated, both reported as degraded
• classified partial, scopeSummary: {requested: 4, produced: 4, degraded: 1, omitted: 0}
• tool mix: Bash 6674, Edit 970, Read 879, MCP browser 329, Write 193
• all four payloads validate against their published schemas (ajv)

Checks: unit tests 15/15, claude-export-ingest 5/5, validate-manifests 20 manifests 0 errors, check-source-id-stability, scope-catalog:check, check-additive-schemas (51), check-page-api-additive all clean.

Known trade-off

hex_secret redacts any 0x + 64 hex, which also catches transaction hashes and data-point ids. It fires often on Vana work (429 hits here). I kept it fail-safe — a raw private key has the same shape — and the tally makes the cost visible. Easy to narrow to context-matched hits if reviewers prefer.

Not in this PR

• No connector-index artifact or signing; this is the source change only.
• No opencode connector yet. Same shape of work, different store — worth a follow-up if we want coverage across coding agents (that is also where GLM trajectories would come from, since GLM has no account to connect to).

https://claude.ai/code/session_012ZQPAhPHFaF3fmmxUNHma8

[github-actions]

Schema Health Check — All Clear

All 54 scopes have consistent local schema files.