run 0.1.0
Start one program from the deployer's programs directory with exact arguments, no shell, a fixed environment, a deadline and an output cap; answers its exit status and output. Opt-in: it can run whatever the directory holds, and it does not sandbox it.
run [--root DIR] [--timeout-ms N] [--max-output N] [--stdin] [--max-stdin N] [--format json|text] -- PROGRAM [ARG...]
- output
- one JSON document
- schema
- run.v1
- reversibility
irreversible-consequential- guarantees
bounded_memory- concurrency
- none: it takes no lock and cannot see what the program it starts then does
- confinement
beneath: with --root, paths are checked lexically against it and then opened beneath it one component at a time, following no symbolic link (path.symlink), so nothing outside the root is reachable through one; without it, a reader opens the path as given, links and all, and a writer still follows no link in the file's own name- reads the environment
- no
- reads the clock
- no
- compiler
- de4e52fd8180
- for an agent
- SKILL.md (
run skill)
Authority
Derived by cancho authority from the source, embedded in the binary, and checked against tools.toml's ceiling in CI. Every label is bounded.
- args
- child_signal
- clock
- err_write
- exec("/opt/cancho-tools/programs")
- fs_read("")
- heap
- io_read
- io_write
- pipe_read
- pipe_write
- poll
What the row cannot say:
- path extent: a path from the command line is checked at run time, so the row says fs_read("")
- what a symlink reaches
- what standard output carries
- resource use beyond the tool's own caps (memory, wall time)
- per-invocation behaviour such as --dry-run, which the row cannot see
Operands
| Operand | Role | How many | Meaning |
|---|---|---|---|
PROGRAM | name | 1 | a program name in the programs directory, never a path |
ARG... | text | 0 or more | the program's arguments, passed byte for byte |
Flags
| Flag | Kind | Role | Default | Meaning |
|---|---|---|---|---|
--root | path | root | start the program in this directory (default: the working directory) and refuse a directory outside it | |
--timeout-ms | nat | none | 30000 | the longest the program may run, in milliseconds, after which it is killed (1 to 600000) |
--max-output | nat | none | 1048576 | the most bytes kept from each of standard output and standard error, and the program is killed at the first byte past it (ceiling 67108864) |
--stdin | bool | none | give the program this tool's standard input as its standard input (default: nothing) | |
--max-stdin | nat | none | 1048576 | the most standard input accepted with --stdin (ceiling 67108864) |
--format | choice:json/text | none | json | json for a program, text writes the program's standard output to standard output and its standard error to standard error |
Limits
| Limit | Default | Ceiling |
|---|---|---|
timeout-ms | 30,000 | 600,000 |
max-output | 1,048,576 | 67,108,864 |
max-stdin | 1,048,576 | 67,108,864 |
Exit codes
| Code | Name | Meaning |
|---|---|---|
| 0 | SUCCESS | what was asked was done; zero matches is success |
| 2 | INVALID_ARGS | the invocation is malformed or names something of the wrong kind; nothing was read |
| 3 | NOT_FOUND | a named input does not exist |
| 4 | PERMISSION_DENIED | the operating system refused, or the tool's own confinement did |
| 8 | PRECONDITION_FAILED | the answer is no, or a limit was reached |
Rules (21)
Match on rule, not on the message. A retry repair is an argv that can be run as it stands, and never widens what the tool may touch.
| Rule | Exit | Repair | When |
|---|---|---|---|
args.unknown-flag | 2 | sometimes | a flag that is not in the tool's table |
args.missing-value | 2 | never | a flag that takes a value was given none |
args.bad-value | 2 | never | a flag's value is not of the flag's kind |
args.unexpected-value | 2 | never | a flag that takes no value was given one |
args.duplicate-flag | 2 | never | the same flag was given twice |
args.missing-operand | 2 | never | an operand the tool needs was not given |
path.empty | 2 | never | an empty path |
path.dotdot | 2 | sometimes | a path with a `..` component |
path.absolute | 2 | always | an absolute path where one relative to --root is wanted |
path.outside-root | 4 | never | an absolute path that is not under --root |
path.too-long | 2 | never | a path longer than 4096 bytes |
path.symlink | 4 | never | a path that runs through a symbolic link, which the tool does not follow |
io.not-found | 3 | never | a named input does not exist |
io.not-a-directory | 3 | never | a path runs through something that is not a directory |
io.permission-denied | 4 | never | the operating system refused access |
limit.input-too-large | 8 | sometimes | an input longer than --max-bytes |
run.program-name | 2 | never | the program is not one plain name (it is empty, `.`, `..` or has a `/`) |
run.program-not-found | 3 | never | the programs directory holds no program of that name |
run.spawn-failed | 8 | never | the operating system could not start the program or its channels |
limit.timeout | 8 | never | the program ran past --timeout-ms and was killed |
limit.output-too-large | 8 | never | the program wrote more than --max-output to one stream and was killed |
Evidence
each gate is a test in cancho-tools/tests/conformance, run by CI on the commit this binary was built from:
- M1 schema conformance
- M2 determinism
- M3 rule coverage and hint soundness
- M4 fault injection
- M5 differential against GNU
- M6 authority
- M7 mutation behaviour
- M8 confinement
- M9 memory flatness
Agent in the loop: not run: no claim that this tool is more agent-friendly than the incumbents is made (cancho docs/agent-toolbox.md section 7.2).