npm/opencode-aiopencode
does not hang when stdin is a non-delivering pipe: HANGS - agent blocks forever
npm/opencode-aiagent trajectory
Awaiting three identical trials
Deterministic observations can identify likely friction, but they cannot license the claim that an agent passed or failed.
deterministic evidence · not the rank
Every check, its evidence, and the fix
N/A is excluded from both numerator and denominator. Open a check to inspect what the probe observed.
Legibility
can an agent learn this tool from its own surface?
failD1.1machine-readable schema (`schema` / `--describe` returns JSON)0/8
no capability document; agent must parse help prose
Add a `schema` or `--describe` command that prints the complete command, option, input, and output schema as JSON.
passD1.2exposes an MCP surface4/4
yes
failD1.3help contains usage examples0/3
no
Add copy-pasteable examples to top-level and subcommand help, including one successful machine-readable invocation.
Structured I/O
is the output machine-parseable?
failD2.1documents a JSON output mode0/10
no machine-parseable output documented
Add and document `--json` or `--output json`, then keep its keys and types stable for scripts and agents.
failD2.2no ANSI escapes when the stream is piped0/5
leaks colour codes into agent context
Detect non-TTY output and disable color, cursor movement, spinners, and other ANSI escapes automatically.
failD2.3has a quiet mode0/5
no
Add `--quiet` (or `--no-progress`) so stdout can contain only the requested result.
Non-blocking
does it ever hang when no human is present?
failD3.1does not hang when stdin is a non-delivering pipe0/8
HANGS - agent blocks forever
When stdin is not a TTY, fail fast instead of prompting; document a `--non-interactive` or `--yes` escape hatch.
passD3.2--help does not hang4/4
ok
failD3.3documents a non-interactive escape hatch0/4
no
Document a non-interactive mode that supplies defaults or exits with an actionable error whenever input is missing.
failD3.4cold invocation under 3s0/4
8.01s
Defer plugins, updates, auth, and network initialization until after argument parsing so a cold invocation returns within 3 seconds.
Context economy
how many tokens does it cost to use?
partialD4.1help fits an agent context budget3/8
4285 bytes (~1071 tokens)
Keep top-level help concise and move detail into subcommand help or machine-readable schema output.
passD4.2does not reprint full help on every mistake7/7
99 bytes
Safety rails
can an agent preview before it destroys?
failD6.1offers a dry-run0/5
no
Add `--dry-run` to mutating commands; validate locally and print the exact planned changes without making network calls.
failD6.2offers an explicit confirmation bypass0/5
no
Add an explicit `--yes` confirmation bypass that is safe to use only when every required value is supplied.
Error recovery
does a mistake produce an honest, actionable signal?
failD7.1unknown subcommand exits non-zero0/8
exit 0 — SILENT FAILURE
Return a non-zero exit code for unknown subcommands and name the closest valid subcommand on stderr.
passD7.2unknown flag exits non-zero8/8
exit 1
passD7.3error message is terse4/4
99 bytes
Cold credentials
can an agent get authenticated without a human?
failD9.1names its credential env var in its own help surface0/6
agent must already know the variable name; undiscoverable from --help
Name the supported credential environment variable in `--help`, including its precedence over config files.
failD9.2offers a no-account or temporary-account path0/8
no; the agent stops at the signup wall
Offer a documented temporary, guest, or claim-later path so a cold agent can complete one useful task without signup.
passD9.3does not launch a browser auth flow in a headless context6/6
ok