feat: add Grok CLI as a bidirectional harness, plus Claude Code porting fixes

[Solomon-mithra]

This PR adds Grok CLI as a fourth harness, in both directions, plus five Claude Code porting fixes found through hands-on use of openmemory against my real session stores. Every change here came from the same loop: port a session, open the destination harness, notice what's wrong, fix it, verify on the real store.

1. Grok CLI as a bidirectional harness (commits 1 and 2)

Grok stores sessions at ~/.grok/sessions/<percent-encoded-cwd>/<uuid>/ with summary.json (id, cwd, timestamps, generated title) and chat_history.jsonl (one message per line).

Import (src/adapters/grok/import.ts, modeled on the codex adapter): detect via GROK_HOME ?? ~/.grok, discover newest-first from summary timestamps, parse to text turns. Grok wraps every real user turn in a <user_query> envelope; the parser strips exactly that envelope. reasoning, tool_result, and tool-call-only assistant records count into droppedTurns; synthetic injections (synthetic_reason) are surfaced via sourceMetadata.syntheticSkipped; malformed lines land in sourceMetadata.skippedLines (AC7 pattern). cwd is authoritative from summary.json; the fallback percent-decodes the directory name, which is lossless (unlike claude-code's dash munging).

Export (src/adapters/grok/export.ts): synthesizes a native session pair in Grok's own shapes (user text re-wrapped in the <user_query> envelope, assistant content as a plain string with model_id, UUIDv7 session ids via Bun.randomUUIDv7). The summary carries every field present in all native summaries; a store test showed Grok's parser drops sessions whose summary is missing any of them. openmemory_source in the summary is the provenance stamp for teardown() (no adapter ledger, matching the other exports). Runtime state files (events.jsonl, plan.json, ...) are deliberately not fabricated. current_model_id is pinned to a Grok model, never the source harness's model. The envelope strip on import is the exact inverse of the wrap on export, so grok round-trips through the hub with full text and title parity (covered by a round-trip test).

verify gained grok as an isolated destination (setGrokSessionsRoot at the temp root, reimport through the grok import adapter's explicit-path parse).

Verified live: ported a real 5,000-message Grok session into Claude Code and back, and Claude Code and Codex sessions into Grok; the ported sessions appear in each harness's own resume picker.

2. Round-trip echo suppression (commit 3)

Porting A to B and then B back to A duplicated the original: the ported copy has a fresh session id the ledger has never seen. On my store one session tripled through a claude -> grok -> claude -> grok chain.

The data to prevent this already existed, since every export stamps provenance onto its destination. Importers now surface that stamp as sourceMetadata.openmemoryOrigin (grok reads summary.json's openmemory_source; claude-code reads the per-line openmemorySource), and exportSessions() skips any session whose origin harness IS the destination, reporting the original id in the outcome. Exports to a third harness are unaffected; --force deliberately overrides.

3. Claude cwd encoding fix (commit 4)

encodeCwd mapped only / to -, but Claude Code encodes every character outside [A-Za-z0-9]. Any cwd containing _, ., or spaces exported into a sibling project folder Claude never reads, so the session existed on disk and in the ledger but was invisible in claude --resume. Verified against a real store (same path, both encodings):

cwd:              /Users/me/projects/my_app/web
Claude reads:     -Users-me-projects-my-app-web
openmemory wrote: -Users-me-projects-my_app-web

Fix: cwd.replace(/[^A-Za-z0-9]/g, "-"), plus regression cases (underscore and dot-plus-space paths).

4. Session titles preserved in both directions (commits 5 and 6)

Ported sessions degraded to "whatever the first prompt happened to be" in Claude's /resume and in openmemory's own picker.

• Export now appends a summary record ({type, summary, leafUuid}) carrying the IR title, the same mechanism Claude uses for compacted sessions. Verified on Claude Code 2.1.205: the resume picker shows the ported session under its source title.
• Import now honors summary records as the title (last record wins), so sessions ported into Claude, and Claude's own compacted sessions, keep their real titles in the source picker. Exact, no heuristics; round-trip title parity is asserted in tests.

5. Claude discovery hygiene (commits 7 and 8)

Three kinds of junk flooded the Claude source picker on a real store (65 real conversations were buried under roughly 130 garbage rows):

• Summarizer side-sessions: Claude Code runs background sessions to generate the conversation summaries for its own UI, saved as ordinary 3-line .jsonl files with no isMeta/isSidechain marking. 121 on my store, each listed as a bogus "2 messages" chat. Now flagged by the summarizer prompt's fixed preamble (checked only against the first real user turn, so conversations that merely mention the phrase are unaffected) and skipped.
• Subagent transcripts in the newer flat layout (agent-*.jsonl directly in the project dir) and zero-turn/aborted sessions, listed as "0 messages" rows. Discovery now skips zero-turn sessions (matching the codex adapter's no-turn skip, via a shared extractTurns helper) and excludes agent-*.jsonl by name. The find invocation also moved from execSync string interpolation to execFileSync with an argument array.
• Env-context boilerplate titles: sessions whose first "user" turn is an injected <user_info> block (OS Version / Shell / Workspace Path) were titled with that block. Added to the preview boilerplate strip list so titles fall through to the first real human message.

After: discovery on the same store returns 65 sessions, 0 junk.

Tests

• 386 pass, typecheck clean. New suites: grok-import.test.ts (13), grok-export.test.ts (10, including export -> import round-trip parity), echo-suppression cases in orchestrator.test.ts, plus extensions to the detect, registry, preview, and claude-code import/export suites.
• Three failures on my machine (detect real-machine E2E, registry real-capabilities, TUI wizard walk) fail identically on unmodified main here because opencode is not installed; no new failures are introduced.

Known gaps / not in scope

• Codex export still does not serialize session.title (pre-existing, documented in CLAUDE.md).
• Native, never-compacted Claude sessions still title by first user message; Claude Code does not persist its generated titles anywhere a local tool can read (its summarizer side-sessions do not reference the conversation they summarize).

Happy to split any part of this into separate PRs if that's easier to review.

[Solomon-mithra]

@HowieG @taranjeet , do you think this Grok implementation would be helpful