feat: Add aggregate byte cap to subprocess output capture

[mabry1985]

Summary

The _run() helper in tools.py previously buffered a browser command's entire stdout+stderr via subprocess.run(capture_output=True), so untrusted page content (via get text, get html, eval) could consume process memory and flood the model's context with no plugin-owned bound.

_run() now uses subprocess.Popen with two concurrent drain threads (one per pipe, so a full stderr can't deadlock stdout) that enforce an aggregate stdout+stderr byte cap. The cap is a new config key, max_response_bytes (default 200000), read at tool-registration time and exposed as an operator-editable Settings field. On overflow the child is killed and the tool returns a bounded Error: output exceeded {cap} bytes (truncated); on timeout the child is terminated — and in both cases reaped, so no zombies are left behind. Within-cap behavior is byte-for-byte unchanged (raw stdout stripped; stderr on failure still sliced to 500 chars), and the existing timeout_s continues to work alongside the new cap. Version bumped to 0.6.5 in both protoagent.plugin.yaml and pyproject.toml.

Tests were retargeted from mocking subprocess.run to a Popen stand-in with BytesIO pipes, plus new coverage for: over-cap truncation + child kill, exact-cap pass-through, aggregate stdout+stderr bounding, configured-cap override, the 200KB default when unconfigured, and timeout terminate-and-reap.

Fixes #18

[protoreview]

QA panel review — PASS

code-review-structural · head 7448a84b1f47 · formal

Low-risk PR: the core change (switching tools.py from subprocess.run to subprocess.Popen with drain threads, and updating the test mocks accordingly) was reviewed by two finders and found sound. The single actionable item is a stale documentation reference — PROTO.md's "Test patterns" section still describes the old conftest.fake_run / CompletedProcess mechanism that no longer exists. Fix first: update that bullet to describe the new fake_popen / _FakeProc pattern. No panel disagreement; the verifier confirmed the finding by re-reading PROTO.md on main and cross-checking the diff. Coverage note: tools.py (the main production file) drew zero findings, which is consistent with the finders' assessment that the Popen refactor is correct.

Findings

	Severity	Location	Finding	Verified
🟡	minor	PROTO.md	The PR moves CLI mocking from subprocess.run/conftest.fake_run to subprocess.Popen with a test-local fake_popen but leaves PROTO.md's 'Test patterns' section d…	confirmed

findings JSON (machine-readable)

[
  {
    "file": "PROTO.md",
    "line": 0,
    "severity": "minor",
    "category": "conventions",
    "claim": "The PR moves CLI mocking from subprocess.run/conftest.fake_run to subprocess.Popen with a test-local fake_popen but leaves PROTO.md's 'Test patterns' section describing the old mechanism, so the repo's canonical grounding doc now misdescribes how the tool tests mock the CLI.",
    "evidence": "PROTO.md (base ref): \"Mock the CLI \u2014 monkeypatch `subprocess.run` with `conftest.fake_run(...)`. It records the argv it was called with and returns a canned `CompletedProcess`\" \u2014 but the diff replaces every such call with `monkeypatch.setattr(tools.subprocess, \"Popen\", fake_popen(out=\"OPENED\", record=rec))` and adds a new local `def fake_popen(out=b\"\", err=b\"\", rc=0, timeout=False, record=None, procs=None):` in tests/test_agent_browser.py; PROTO.md is not touched by the PR.",
    "verdict": "confirmed",
    "note": "Re-read PROTO.md on main: the 'Test patterns' bullet verbatim says 'monkeypatch subprocess.run with conftest.fake_run(...) \u2026 returns a canned CompletedProcess'. The PR diff shows all test calls now use subprocess.Popen + local fake_popen (returning _FakeProc with BytesIO pipes), and PROTO.md is absent from the diff. The doc is stale."
  }
]

---

1 finding(s) excluded from the verdict by in-diff confinement (file not among this PR's changed paths):

• PROTO.md (minor) — The PR moves CLI mocking from subprocess.run/conftest.fake_run to subprocess.Popen with a test-local fake_popen but leaves PROTO.md's 'Test patterns' section de