Module fingerprint 

Deterministic fingerprints: the same error seen by two drones (or two sessions weeks apart) must hash the same, regardless of line numbers, temp paths, ids or durations.

Functions

bash_read_paths

Files a shell command pulls into context (cat a.rs, sed -n 1,80p src/x.rs | …). Agents read through Bash far more often than through the Read tool, so recall-on-touch must see these too.

command_shape

Program plus subcommand (cargo test, npm run, git status), the unit at which “how do I build/test this repo” knowledge transfers.

error_fingerprint

Fingerprint of the most specific error line (see error_line).

error_line

The most specific error line of a failed command’s output: the last line naming the error (Python tracebacks end with it), else the first error-looking line that is not a generic header. None for harness messages (permission denials, rejections).

invocation

The invocation structure of a command: program, subcommand, env assignments and flags, without test selectors, paths or quoted values. Two commands that differ only in which tests/files they target have the same invocation.

is_test_failure

Whether an error line is a failing test assertion rather than an environment/tooling error.

is_tool_shape

Whether a command shape is a reusable tool invocation (build, test, package, env sourcing), as opposed to an ad-hoc script or a one-off inspection that teaches nothing to the next drone.

looks_failed

A command can fail behind a pipe (pytest | tail exits 0), so judge by output too: tracebacks, test-runner failures, compiler errors.

normalize_error_line

Normalize one error line so incidental details don’t split fingerprints.

only_reads

True when every segment of the command just reads or lists files.