OpenClaw provider: format docs, toolpath-openclaw crate, CLI wiring, real-session validation

[akesling]

Adds OpenClaw (github.com/openclaw/openclaw — multi-channel, local-first assistant) as a full toolpath provider, from format reference to validated round-trips. 22 commits, built spec→plan→TDD.

What's here

Format reference — docs/agents/formats/openclaw/ (12 docs, claude-code style): storage layout, the version-3 JSONL tree/visible-leaf envelope, 10 entry types, messages/content blocks, tools (no stored diffs), two-shape token usage, lineage, the channel/peer actor model, known issues, walkthrough, changelog. Originally compiled from upstream source (openclaw/openclaw @ 68c533cf), since upgraded to observed against real captured sessions.

toolpath-openclaw crate (0.1.0) — forward derive (JSONL → ConversationView → Path) and reverse inception (the projector writes transcript + sessions.json routing entry; a live gateway adopts it without restart — verified with a probing-question test, follow-up turns append to the same transcript). Channel identity is recovered from the routing key into a channel-aware human:<channel>/<peerId> actor. 52 unit + 7 integration tests, incl. a real-fixture suite and a full-chain fidelity contract.

toolpath-convo changes (0.11.0 → 0.13.0, additive):

• DeriveConfig.user_actor — session-level human actor override (the multi-channel case).
• Typed round-trip fidelity fields: Turn.thinking_signature/text_signature/response_model/marker (ConversationMarker::Compaction/BranchSummary), ToolInvocation.thought_signature/execution_mode; wired through derive_path/extract_conversation. No provider-namespaced extras.

CLI wiring — p import/list/show/export openclaw, share aggregation, and inception+ resume: path resume --harness openclaw incepts, then probes OPENCLAW_GATEWAY_ADDR and either reports adoption (gateway running → no exec) or execs openclaw gateway.

Real-session validation — scripts/openclaw-docker.sh runs the official Docker image (v2026.6.11) and captured the fixtures committed at test-fixtures/openclaw/, wired into the cross-harness matrix (all cells pass). Real data drove several fixes:

• sessions_spawn classifies as Delegation; agent:<id>:subagent:<uuid> keys → spawn-child, no fabricated channel actor.
• Token double-count fix: the final assembled reply of a multi-call run (post sessions_yield, no responseId) carries run-cumulative usage — detected via exact field-wise-sum match and never stamped onto a step (real-fixture test pins the single-counted total).
• totalTokens = input+output+cacheRead+cacheWrite (11/11 observed rows); transcript perms are 0644 not the source-implied 0600; docs corrected.
• sessions.json upsert merges raw JSON (was: typed round-trip that could clobber entries it couldn't parse).

Verification

• cargo clippy --workspace --all-targets -- -D warnings: clean.
• cargo test --workspace: all 62 suites green (matrix: all openclaw↔* cells pass on the real fixture).
• Live: real Anthropic turns through the containerized gateway; import/list/show/export verified end-to-end on real transcripts; inception adoption verified against a running gateway.

Known residuals (documented in the format docs as "from source")

• Group/multi-party sessions not yet captured; [from:] group-sender parsing unimplemented (room-level attribution).
• Branch/leaf-move/compaction mechanics not yet observed in real data (real linear sessions contain no leaf rows — recorded).
• Native (non-Docker) openclaw gateway resume exec branch is recipe-tested only.

---

^{Need help on this PR? Tag /codesmith with what you need. Autofix is disabled.}