Module external 

External executors: another coding agent — Claude Code (claude -p), Codex (codex exec) or any command — makes the attempt; speccoder keeps everything around it.

Guarantees are diff-level, not action-level: the agent edits the worktree with its own tools and its own permission system, so speccoder cannot classify each call. What it still does: the isolated worktree, frozen and spec-derived acceptance tests written before the attempt, the integrity audit of the diff (protected files, deleted or weakened tests), the checks and the verdict, retries with the verifier’s digest, the gaming monitor, ticking and committing. Where the agent has a permission system, speccoder’s command allowances are passed to it.

Structs

Executor

[executor] in speccoder.toml.

Outcome

What an external attempt produced, in the shape of a native attempt’s output.

Enums

Kind

Which agent.

Functions

argv

The argv for an executor; allow are the commands speccoder allows (checks and [commands] allow), passed to agents with a permission system; withheld is a directory the agent is told not to read (the held-out tests) — withheld, not sealed: only Claude Code takes such a rule.

interpret

Reads an agent’s final message and usage from its stdout (and, for Codex, its last-message file).

parse

Reads [executor]; None (or kind = "native") is speccoder’s own loop.

prompt

What the external agent is told: the same context a native attempt gets, the rules speccoder will check afterwards, and how to stop honestly.

prompt_in

prompt for a project whose format is worded w.

run

Runs one attempt with exec in worktree.