{ "status": "ISSUES_FOUND", "findings": [ { "path": "adapter/strings/omp-spt.mjs", "symbol": "runSpt", "lines": "111-128", "severity": "P1", "confidence": 10, "impact": "A normal outbound-command failure can terminate the entire OMP process, bypassing the custody retry/fail-closed path and losing the only local chance to report the accepted message outcome.", "failure_scenario": "Point `OMP_SPT_SPT_BIN` at an executable that starts and immediately exits without reading stdin. `settleItem()` calls `runSpt(..., payload)`, `child.stdin.end(input)` writes to the closed pipe, and the stdin Socket emits `EPIPE`. Only `child.on(\"error\")` is registered; an unhandled `child.stdin` error is an uncaught exception in Node. The tests replace `runSptCommand`, so they never exercise the real pipe.", "fix": "Register stdin error handling before writing, reject the command promise on write/EPIPE errors, and make close/error/stdin completion settle exactly once. Add a real-child regression test whose stdin closes immediately." }, { "path": "adapter/strings/omp-spt.mjs", "symbol": "runSpt / teardownSession / stopResources", "lines": "111-128,272-324,328-340", "severity": "P1", "confidence": 10, "impact": "Bind, delivery, state transitions, fatal shutdown, and normal OMP shutdown can hang forever. A listener that ignores termination can also survive or keep the Node event loop alive.", "failure_scenario": "Use an `spt` process that never exits. A hung bind leaves `session_start` pending; subsequent `session_shutdown` waits forever at `await bindPromise`. A hung `send` leaves custody permanently unsettled. A hung `session-end` prevents teardown completion. `failClosed` waits for that teardown before calling `ctx.shutdown()`. Separately, listener teardown only calls `child.kill()` and neither verifies the return value nor waits for close/escalates.", "fix": "Give every spawned CLI command a command-specific hard deadline, bound captured output, terminate then force-kill on deadline, and track/join children during teardown. `stopResources` must await listener exit with escalation. Teardown must have an overall deadline so `ctx.shutdown()` still executes after reporting failure." }, { "path": "adapter/strings/omp-spt.mjs", "symbol": "drainEvents / listener data queue", "lines": "23-44,154,172,482-497", "severity": "P1", "confidence": 9, "impact": "Any reachable peer, burst, oversized frame, or corrupt listener can grow endpoint memory without limit until OOM. Shutdown then serially processes the entire accepted backlog, multiplying termination latency.", "failure_scenario": "Keep one OMP turn active and stream many valid `` envelopes. Every full envelope is retained in `queue` with no item/byte cap. Alternatively stream an opening ``. `compose_prompt` concatenates the raw transcript tail into the same user-level prompt as the summarization instruction; no higher-priority policy or trust labeling exists. The resulting two-tag stdout is accepted as the durable delta based only on OMP success/nonempty output.", "fix": "Put the summarization contract in a higher-priority OMP system prompt, parse JSONL into explicit role-tagged untrusted data, state that transcript instructions must never be followed, strictly parse and size-bound the exact two-tag result, and ensure resumed consumers treat the delta as quoted data rather than executable instructions. Add adversarial transcript fixtures." }, { "path": "tools/omp-spt/src/launch_omp.rs", "symbol": "propagated_exit_code", "lines": "304-306", "severity": "P2", "confidence": 10, "impact": "Windows OMP crashes are reported as successful endpoint exits, hiding the failure from broker/recovery/operations logic.", "failure_scenario": "An access violation exits with Windows status `0xC0000005`, exposed by Rust as negative `i32` `-1073741819`. `.clamp(0, 255)` converts it to `0`, so `launch-omp` exits success. `STATUS_CONTROL_C_EXIT` and other high-bit exception statuses have the same defect.", "fix": "Return success only for raw code 0. Preserve 1..=255 if desired, but map every other nonzero/absent code to `ExitCode::FAILURE`; add a Windows regression using `ExitStatusExt::from_raw(0xC0000005)`." }, { "path": "tools/omp-spt/src/echo_commune_omp.rs", "symbol": "run history ingestion", "lines": "127-168", "severity": "P2", "confidence": 10, "impact": "The advertised 48 KiB summarizer bound does not bound memory. Large but valid long-lived session transcripts can OOM the helper during shutdown/commune generation.", "failure_scenario": "Create a multi-gigabyte session JSONL. `read_to_string` loads the entire file into `history`; stdin is likewise read to EOF into an unbounded String. Only afterward does `bounded_tail(&history, 48 KiB)` discard the prefix.", "fix": "For files, seek from the end and read only enough bytes to locate a complete-line tail. For stdin, retain a bounded rolling buffer/line ring while reading. Reject or explicitly truncate a single over-limit line without first allocating the whole input." }, { "path": "adapter/strings/omp-spt.mjs", "symbol": "drainEvents / listener malformed-frame handling", "lines": "23-44,482-495", "severity": "P2", "confidence": 9, "impact": "One malformed/truncated frame can silently corrupt one sender's prompt and consume another sender's valid delivery, violating correlated custody while the endpoint remains advertised healthy.", "failure_scenario": "Feed `partial` followed by `ok`. `drainEvents` pairs a's opener with b's closer, emits one corrupted a envelope, and permanently consumes b. A missing `from` frame is only logged and continued, also losing custody without failing the protocol.", "fix": "Use the listener's actual line/framing contract or a validating state machine. Detect nested openers, malformed attributes, missing senders, and frame-size overruns; on protocol corruption fail closed rather than resynchronizing by consuming a later close tag. Add deterministic malformed-frame tests." } ] }