Module tool_evidence 

ACP tool calls as extra evidence for a task check. The node remembers the last finished tool calls of each task worktree (task-<id>) session: the command, whether the provider said it failed, how long it took (the node’s own clock: ACP sends no duration) and why it did not run, if it did not. When the node sends its own task.check_run, it attaches the last call that ran the same command (see take). The proof stays the node’s run: a tool call is only what the agent saw and never passes a check by itself.

ACP turns every tool_call_update (also in_progress) into a completion, so the last update of a call wins. Command text is kept only for matching; it is never logged.

Structs

ToolCall

A finished tool call of a session in a task worktree.

Functions

command_of

The command a tool call runs: rawInput.command as one shell line or command + args (or an argv array), else a Bash-style title in backticks. None: not a command.

forget_open

A session ended: its unfinished calls go; finished ones stay for the next check.

observe

Feed one provider event of an ACP session (only task worktrees are kept).

runs

The call ran check: the same line, or the line after a cd ... && / ; prefix.

take

The last finished call in workspace that ran check, removed so it backs one check run.