Adversarial review: area commands

Scope: cli/src/framework/**, cli/src/cli/{main,output,formatters,region,repl,scaffold,_rich*,_orjson}.ts, cli/src/bin/abx.ts, cli/src/cli/commands/{config_cmd,sandbox,sandbox_files,sandbox_process,sandbox_system}.ts against Python agent_sandbox/cli/**. Reviewer: independent (did not write the port). cli/src/** was not edited.

Placeholders: $ORACLE_VENV = virtualenv with the original Python abx (real Click 8.5, Rich, httpx), $ORIG_REPO = original checkout. All harness runs used an isolated HOME, ABX_NO_KEYCHAIN=1, no real credentials and a private stub HTTP server per run.

Deliverables

Run: ORACLE_VENV=... node parity/diff/adv_commands_<x>.mjs (prints DIFF ...; prints SKIP and exits 0 without the oracle). Env knob: ADV_DROP_HTTP_INFO=1 hides the INFO: HTTP Request: lines (see C-05) so other diffs are visible. Tests: npx vitest run tests/test_cli/adv_commands.test.ts (moved from tests/adversarial/commands/commands.test.ts; about 40 s; ADV-12/13 need python3, ADV-07/10 need /dev/full).

Counts

• BUG: 13 open-or-fixed (1 HIGH C-01, 1 MEDIUM C-03, 1 MEDIUM-LOW C-02, 10 LOW); 4 of the LOW/MEDIUM ones (C-10a..d) were fixed in cli/src while I worked.
• GAP-UNJUSTIFIED: 2 (C-04, C-13) - ACCEPTED-OK: see the last section.
• Vitest at the end of the review: 15 cases, 9 FAIL (open: C-01..C-09), 6 PASS (C-10a..d: fixed meanwhile, now regression guards).

Test map: ADV-01 = C-10a, 02 = C-10b, 03 = C-10c, 04 = C-10d, 05 = C-04, 06 = C-05, 07 = C-03, 08 = C-06, 09 = C-07, 10 = C-08, 11 = C-09, 12 = C-01, 13 = C-02, 14 = C-10b (signal), 15 = C-10c (request timeout).

Fix status

All BUG / GAP-UNJUSTIFIED findings were addressed in cli/src; the ADV-commands-NN tests moved to tests/test_cli/adv_commands.test.ts (18 tests, all active, no test.skip; it.skipIf only when python3 or /dev/full is missing on the machine).

Findings

Severity is the user impact on a real run. "expected" is always the Python oracle.

C-01 REPL drops type-ahead / pasted lines and then hangs (BUG, HIGH) - open, ADV-commands-12

• Repro: abx connect SBX on a TTY; paste echo a<CR>echo b<CR>echo c<CR>exit<CR> in one write.
• Expected: a, b, c run, then exit 0. Actual: only echo a runs, then the prompt hangs forever (the rest of the paste is gone).
• Root cause: src/cli/repl.ts:177 _editedInput creates a NEW readline.Interface per prompt and rl.close()s it after the first line (finish, ~line 188); bytes of the same chunk that belong to later lines are consumed and discarded with it.
• Fix: keep one persistent Interface for the whole REPL and feed lines through a queue (rl.on("line") + a pending-resolver list), or rl.pause() instead of close() and re-use it; keep history in the same instance.

C-02 REPL stops reacting after a very long line (BUG, MEDIUM-LOW) - open, ADV-commands-13

• Repro: in a pty type echo + 5000 x, Enter, then echo hello. Threshold is about 4090 bytes (4080 works).
• Expected: both lines run (Python handles 5000+ chars). Actual: first result and a new prompt are printed, later input is ignored (hang). Node also redraws the whole line for every keystroke (112 KB of output for a 4000 char line vs 8 KB in Python).
• Root cause: not fully pinned down; same area as C-01 (repl.ts _editedInput, one throw-away readline per line on a raw-mode TTY, pty canonical buffer limit 4096 is hit while raw mode is toggled between lines). A persistent Interface (C-01 fix) most likely fixes it.

C-03 failed stdout/stderr write crashes with an unhandled stream error (BUG, MEDIUM) - open, ADV-commands-07

• Repro: abx config list > /dev/full (also list --quiet).
• Expected: OS error: [Errno 28] No space left on device + Suggestion: Check file permissions and disk space., exit 1. Actual: Node prints a stack trace (node:events: throw er; // Unhandled 'error' event ... ENOSPC) and exits 1. The same code path swallowed the failure silently earlier in the review (exit 0), so the behaviour is also unstable between revisions.
• Root cause: src/framework/io.ts:40-47 ProcessOutput.write wraps stream.write in try/catch, but write errors on a SyncWriteStream are emitted asynchronously as 'error' events, not thrown, and there is no stream.on("error") handler. Pipes: | head -c1 is fine (EPIPE ends quietly, like Python).
• Fix: attach an error listener on process.stdout/stderr in ProcessOutput, record the first error and rethrow it (as toOSError) from the next write/flush or at exit, so handle_errors prints Python's OS error: ... message and exits 1.

C-04 --log-level info|debug and --verbose omit httpx's INFO: HTTP Request: line (GAP-UNJUSTIFIED, LOW-MEDIUM) - open, ADV-commands-05

• Repro: abx --log-level info list. Expected: INFO: HTTP Request: GET http://HOST/sandboxes?limit=20&offset=0 "HTTP/1.1 200 OK" on stderr (one per request, also at DEBUG and -v). Actual: silent. KNOWN-GAPS.md only mentions losing SDK log lines.
• Root cause: the transport (src/transport/_httpx.ts) never logs through the httpx logger; the CLI log handler never sees it.
• Fix: log "HTTP Request: %s %s \"%s %d %s\"" at INFO from the same place the response status is known (also redirects and streamed responses). The httpcore/asyncio DEBUG lines (connect_tcp.started, Using selector: EpollSelector) stay ACCEPTED-OK.

C-05 create --env __proto__=X / --metadata __proto__=X silently drop the key (BUG, LOW) - open, ADV-commands-06

• Expected body {"metadata":{"__proto__":"2"},"envVars":{"__proto__":"1"}}; actual {"metadata":{},"envVars":{}}.
• Root cause: src/cli/commands/sandbox.ts:912-920 _parsePairs assigns into a plain {} (the setter for __proto__ swallows strings).
• Fix: Object.create(null) is not enough for JSON.stringify of own __proto__; use Object.defineProperty(result, key, {value, enumerable: true, writable: true, configurable: true}) or a Map.

C-06 FORCE_COLOR together with NO_COLOR leaks a Node warning (BUG, LOW) - open, ADV-commands-08

• Repro: FORCE_COLOR=1 NO_COLOR=1 abx config list. Actual stderr: (node:PID) Warning: The 'NO_COLOR' env is ignored due to the 'FORCE_COLOR' env being set.; Python prints nothing.
• Root cause: Node's own startup check on the real environment (before src/bin/abx.ts runs).
• Fix: in the launcher/shim (bin/abx.js wrapper) call process.removeAllListeners("warning") is too late; use NODE_NO_WARNINGS=1/--no-warnings in the re-exec, or process.emitWarning filter via --disable-warning=... in the shebang; the compiled binary can strip it at build time.

C-07 scheme-less E2B_API_URL message differs (BUG, LOW) - open, ADV-commands-09

• Repro: E2B_API_URL=127.0.0.1:9 abx list (or "not a url"). Expected Request URL is missing an 'http://' or 'https://' protocol.; actual Unexpected error: Invalid URL "127.0.0.1:9".
• Root cause: src/transport/_httpx.ts:472 throws InvalidURL with Node's wording, and the scheme-less case should be the UnsupportedProtocol message (httpx _transports/default.py/httpcore); InvalidURL is only for unparsable authority.
• Fix: when the URL has no scheme (or scheme is not http/https) raise UnsupportedProtocol("Request URL is missing an 'http://' or 'https://' protocol.") / "Request URL has an unsupported protocol 'ftp://'." and let the CLI map it like Python.

C-08 download write error carries a filename (BUG, LOW) - open, ADV-commands-10

• Repro: abx download SBX /home/user/hello.txt /dev/full. Expected OS error: [Errno 28] No space left on device; actual appends : '/dev/full'.
• Root cause: src/api/files.ts:90 fs.writeFileSync error carries path; the generic OSError mapper uses it. Python's write_bytes failure has no filename in this path (the open succeeds, the write fails). Only the write-failure case differs; open failures (ENOENT, EACCES) correctly include the filename.
• Fix: map write failures with toOSError(e, null) (open the file with openSync first, then writeSync, so open and write errors are distinguishable).

C-09 truncated Content-Length body message (BUG, LOW) - open, ADV-commands-11

• Expected ... peer closed connection without sending complete message body (received 1000 bytes, expected 5000); actual (incomplete chunked read).
• Root cause: src/transport/_httpx.ts:1499 hard-codes the chunked wording for ECONNRESET, and "received 0 bytes, expected more" otherwise.
• Fix: count bytes read and compare with content-length (chunked only when transfer-encoding: chunked).

C-10 Integer / timeout parsing: FIXED while reviewing

• C-10a Unicode digits (list --limit ٣; Python int() accepts them): src/framework/py.ts parseIntStrict. ADV-commands-01 now passes.
• C-10b ints above 2**53 (list --limit 99999999999999999999, process info <id> 9007199254740993, process signal --signal 99999999999999999999 "Integer exceeds 64-bit range"): INT now keeps precision. ADV-commands-02 and -14 pass.
• C-10c setTimeout overflow: exec --timeout 2147483648, create --request-timeout 99999999999|inf fired immediately and printed TimeoutOverflowWarning (src/transport/_httpx.ts withTimeout and 5 more setTimeouts). Fixed; ADV-commands-03 and -15 pass.
• C-10d negative effective timeout (exec --timeout -5, http_timeout = timeout + 5 <= 0): Python fails at once with E3001 and sends no Start request. Fixed; ADV-commands-04 passes.

Minor, not covered by a test

• C-11 (BUG, LOW) sandbox REPL on TERM=dumb: Node readline uses its dumb handler (no line editing, Ctrl-D with partial text closes the REPL, Up arrow ignored); Python's readline still edits. No Ctrl-Y yank (echo foo bar, Ctrl-W, Ctrl-Y: py echo foo barbar-style yank, node nothing). Tab and Ctrl-L differ. src/cli/repl.ts _editedInput.
• C-12 (BUG, LOW) empty-Enter prompt redraw differs when type-ahead is pending (cosmetic; subsumed by C-01).
• C-13 (GAP-UNJUSTIFIED, LOW) the lone-surrogate argument/cell ("\ud83d"): Python fails Unexpected error: 'utf-8' codec can't encode ... surrogates not allowed, Node prints U+FFFD and exits 0. Not reproducible from a real terminal (argv is bytes), only via programmatic data; documented only.
• C-14 (ACCEPTED-OK/LOW) info . requests /sandboxes/ in Node vs /sandboxes in Python (WHATWG dot-segment normalisation vs httpx).

Held up (no diff found)

• Full option/argument/command metadata of every command (names, flags, types, defaults, envvars, hidden, help, epilog, show_default, choices) - identical to the oracle; only the no-op callbacks of groups config, mcp, template are missing in Node (harmless).
• --help for every command at widths 40/80/200 (281 cases, help_all.mjs also 0 diff).
• Option-parsing fuzz derived from the oracle tree: unknown options, abbreviations, clusters, --opt=value, --, duplicates, envvars, missing values, bad choices (2385 cases, only C-10a/b).
• Every runnable command x --json/--quiet/--ci/--verbose/--log-level x 401/429/500/503/unreachable/no credentials (1517 cases; 0 diffs once C-04 is filtered out).
• Shortcut configuration edge cases and unknown-command hints/"did you mean" (994 cases).
• Output channel discipline (stdout vs stderr), exit codes, handle_errors formatting, closed stdin/stdout, SIGINT/SIGTERM/SIGHUP/SIGQUIT handling, locale variants, REPL over pipes (26 cases).
• Rich replacement: tables/panels/markup escapes/CJK/emoji/ZWJ/flags/combining/tabs/control characters at widths 1..300 and with COLUMNS/TTY variants (456 cases, 0 diffs).
• Click/Rich TTY + NO_COLOR/FORCE_COLOR/TERM/CLICOLOR matrix (375 cases; one timing flake in the kill spinner frame count).
• Uploads (file/dir/symlink/FIFO/unicode names/12 MB/binary) and downloads, create/kill/list flag semantics, config wizard paths (1450 existing cases).

ACCEPTED-OK (Python-specific or Node-inherent)

• SIGUSR1 starts the Node inspector (Python dies of SIGUSR1); could add --disable-sigusr1.
• PYTHONIOENCODING=ascii, PYTHONUTF8=0 LC_ALL=C: Python-only encoding behaviour. stdin being a directory (Python fatal init error). Invalid UTF-8 argv (Python surrogateescape vs Node U+FFFD).
• Closed-stderr corner cases. A 1e-3 s --request-timeout race. template search --token 日本語🎉: both fail, different text (ascii codec vs invalid header char).
• httpcore connect_tcp.*/send_request_* DEBUG lines and Using selector: EpollSelector; escape-sequence encodings of arrow keys inside the REPL.