bash-no-match extension

Returns a no-match search to the agent as a normal result instead of an error (ticket #113).

pi's bash tool merges stdout and stderr into one stream and marks every non-zero exit as an error. For a search command, exit 1 with an empty stream is the normal "no lines matched" outcome, not a failure - and the error flag costs a small local model a recovery turn on a result that needs no recovery. The extension rewrites exactly that one case and nothing wider: in measured sessions, 55 of 86 no-output exit-1 bash errors were search no-matches, and the rewrite covers all of them.

Behavior

• On every bash tool result, the classifier checks the (exit code, result text, command) triple. It rewrites only when all three hold:
  • the result text is the no-output marker with the exit code line, byte for byte: (no output)\n\nCommand exited with code 1. The marker alone proves both streams were empty, so any preceding segment that printed output disqualifies the command by itself;
  • the exit code line says 1. Exit 2 and every other code stay errors, which also keeps the 7 observed exit-2 cases as errors;
  • the first word of the last && segment of the command is a search command: rg, grep, egrep, or fgrep. A single command is a one-segment chain.
• The rewrite keeps the stock marker and exit line, appends one "(no matches)" note, and clears the error flag. The rewritten result stays under 100 bytes.
• Never rewritten: ; lists, subshells, ls (the observed ls failures carried error messages and were real errors), commands with any output, and every exit code other than 1. A failing cargo test keeps its error status and its stock text, byte for byte.
• No extension writes files, re-issues tool calls, or alters commands in flight.

How it works

The decision lives in the engine-free classifier (extensions/bash-no-match/core.ts): a pure function over the (exit code, result text, command) triple, unit-tested on the observed failure shapes. The pi wiring (index.ts) is one tool_result hook scoped to the bash tool: it extracts the exit code from the result text and the command from the call input, and returns the classifier's rewrite or nothing.

Ordering with the output-limits extension: the two hooks do not interact in either order. This module is a separate extension, not output-limits, and the rewritten result is under 100 bytes, so the output-limits bound (which cuts only oversized results) never touches it; this hook fires only on the 39-byte no-output marker, a result output-limits passes through untouched.

Out of scope (ticket #113): kill/pkill no-match rewriting. They are legitimate exit-1-with-no-output cases, but a wider masking surface; revisit if the data says so.