Windows support, never-hang hardening, and a round of native-parity fixes — every behavioral claim verified against recorded native claude -p ground truth or on the windows-latest CI runner.

Windows works now (#12, fixes #1)

clarp -p "test" previously hung forever on Windows: claude.cmd runs through cmd.exe, so node-pty reports the wrapper's pid while the real Claude process is a grandchild — clarp polled a session status file that never exists. clarp now discovers Claude's real session file by working directory + recency, validates process ancestry so concurrent same-cwd runs can't adopt each other's sessions, and pins it once found. findClaude also handles the native installer's claude.exe, and transcript paths handle backslashes. Verified end-to-end on Windows CI via a real-PTY probe (now a permanent regression test).

Never hangs silently

• A startup watchdog now arms unconditionally until Claude is first observed (generous 120s, separate from the 30s in-turn timeout) and emits a terminal result (subtype error) on failure — stream-json consumers always get a parseable terminator instead of a truncated stream. (#12)
• Prompt dispatch has a deadline: if Claude's TUI eats the typed prompt (its status file reports idle a beat before the input box accepts keystrokes), clarp re-sends once — self-healing the race invisibly — then fails with a terminal error result instead of wedging. (#16)

Native parity (#15)

• --max-turns now matches native semantics: it counts assistant API rounds inside the agentic loop, not user prompts; on exhaustion clarp reports result subtype error_max_turns (is_error: true, no result field, num_turns = rounds) and exits 1.
• initialize and set_permission_mode control requests are acknowledged (SDK clients no longer stall on the handshake), using native's nested control_response shape {type, response: {subtype, request_id, response}}. The interrupt and context-usage acks migrated to the same shape — breaking for consumers that parsed clarp's old flat ack shape.
• Turn-level backend API errors (e.g. invalid model) report subtype success with is_error: true and exit 1, matching native — previously subtype error and exit 0 in stream-json mode.
• Thinking blocks now carry their signature (accumulated from signature_delta), needed by consumers that replay assistant content to the API.

Parity infrastructure (#14)

Recorded native ground truth for the 7-case interrupt matrix and a Task-subagent round-trip (captured before claude -p moved to metered API pricing), joining the existing golden capture sets. Replay-based parity testing (npm run parity:replay) diffs clarp against these recordings without ever spawning native. The parity policy (contract / consumed fields / on-demand / won't-fix-by-design) is documented in goldens/PARITY-GAPS.md.

Reliability

• CI retries npm ci to absorb npm's optional-dependency install flake.
• 437 tests (≈90 added since 0.1.11), green on the 3-OS × 3-Node matrix.

Hardening and parity release: observation core, permission flow, and a golden-capture parity harness.

Observation hardening (#9)

• Proxy never appends fabricated JSON once a response has started — it severs the connection instead. Added error/aborted handlers on the upstream stream and client backpressure on the observed SSE path.
• Persistent StringDecoder on the proxy SSE and transcript-tail paths, so a multibyte UTF-8 character split across network/poll chunks is no longer corrupted in clarp's observed output. Bytes forwarded to Claude were always untouched.
• Unknown content-block types (server_tool_use, redacted_thinking, future kinds) pass through raw instead of being dropped, and only delta-accumulated input is JSON-parsed — a raw block carrying a literal string input is preserved verbatim.

Permission-flow parity (#11)

• Intercept --permission-prompt-tool so the interactive TUI never receives a print-mode flag it can't honor.
• Gate can_use_tool emission on stream-json input and output plus the stdio tool; otherwise auto-deny via the pty so the turn completes (native deny-by-default) instead of wedging on an unanswerable dialog.
• Accept SDK-nested control_response shapes alongside the flat legacy shape, preserving updatedInput so the modified-input guard can fire; deny allows whose updatedInput differs from the requested input (clarp can't inject rewritten input into the TUI dialog).
• Resolve permission against a fresh tool_use id with a brief retry — covering the race where waiting is observed before the tool finishes assembling, and preventing a stale tool from being approved against a newly focused dialog. Auto-deny a pending permission on stdin EOF.
• Emit native user tool_result events reshaped from transcript lines.

Parity testing harness (#10)

• Golden native captures plus replay-mode parity testing: record native claude -p ground truth once, then diff clarp against the saved summaries offline (no native invocation), with per-run isolated workspaces.

Fixes

• Narrow the queued readiness timeout so ordinary queued work does not kill a legitimate long-running Claude turn
• Keep startup/no-safe-ready and unresolved permission waits timeout-eligible so queued stdin cannot hang forever
• Add regression tests for long-running turns with queued prompts and queued set_model commands

Fixes

• Restore the 30s readiness timeout for queued prompt/slash operations when Claude never reaches a safe ready state.
• Add regression coverage so queued stdin cannot hang forever if initial Claude readiness is never observed.

Validation

• CI passed on macOS, Ubuntu, and Windows across Node 20/22/24.
• Local validation: npm test, npm run build, git diff --check, npm pack --dry-run.

Fixes

• Fix macOS fresh installs where node-pty's spawn-helper loses its executable bit and Clarp exits with posix_spawnp failed
• Add install-time and launch-time repair for the affected node-pty macOS helper while upstream stable npm releases still ship the bad mode

Fixes

• Improve stream-json interrupt parity with claude -p
• Preserve queued prompts across interrupt, zero-delay prompt, and immediate follow-up races
• Match native interrupted result, SIGINT, duplicate interrupt response, and interrupted-only exit behavior
• Route prompt, interrupt, permission, and slash-command writes through the session operation queue
• Add interrupt parity cases and PTY interrupt probe tooling

Fixes

• Support Windows stdin prompt handling (-p - / bare - sentinel)
• Normalize node-pty kill signals on Windows (fixes Ctrl+C cleanup)
• Surface Claude transcript API errors

• Harden stream-json turn queue under bursty multi-turn input
• Preserve queued prompts while waiting for Claude readiness
• Treat startup/idle interrupt and stop_task controls as no-ops
• Parse trailing stream-json input on EOF without requiring a final newline

• Align stream-json behavior with claude -p defaults
• Make token-level stream_event output explicit via --include-partial-messages
• Add redacted API debug logging and stream parity harness
• Document known sideband event differences