feat: stealth launch option + adopt the plugin's own running browser session (0.3.0)

[mabry1985]

Why

merchantAgent has never returned a live eBay sold comp. Two causes, both in this plugin's launch path:

1. Google refuses the one-time sign-in. Chrome under CDP control carries navigator.webdriver, and Google's sign-in page answers "this browser or app may not be secure". The eBay account signs in through Google, so the profile never got signed in (zero eBay cookies from Aug 5 until this week).
2. The daemon guard misfires on our own session. The ebay session daemon outlives the agent process. After a restart, or from a subagent that built its own tool set, ensure_session re-ran open --profile … --headed, the CLI printed ⚠ --profile, --headed ignored: daemon already running, and the plugin raised — telling the operator to agent-browser close --all, which kills the very session it was protecting (and every other plugin's browser). Hit twice in one evening.

What

• stealth: false (manifest, ships off) → --args --disable-blink-features=AutomationControlled at launch. Same flag core's agent_browser plugin uses for its stealth option; nothing else changes. Documented in README and the skill as a Google-sign-in fix, not eBay evasion.
• A clean launch writes <profile>/.protoagent-ebay-session.json with {session, profile, headed, stealth, launched_at}. When the CLI reports the daemon already running, the plugin reads it back: same options → adopt the session (log line, no error); different options → error naming exactly which option differs; no/garbled marker → the old foreign-daemon error. Every message says close --session ebay, never --all. close() removes the marker. No profile configured → no identity at stake → unchanged pass-through.
• ebay_session_status adds the stealth instruction to next_step when not signed in and stealth is off; drops it once on.
• SKILL.md: the "there is no Amazon data here" bullet was stale since 0.2.0 added amazon_price_check — it now says what the Amazon number is (asking, not sold). Also documents the new "different options" error.
• Version 0.3.0 (manifest + pyproject).

Tests

126 host-free tests (was 114), ruff check + ruff format --check clean. New: foreign daemon still refused (and --all absent from the message), own session adopted across a fresh instance, option drift named (stealth), garbled marker = no marker, close clears the marker, no-profile pass-through, stealth flag present only when asked and still session-scoped, unwritable profile dir doesn't fail the launch, config → Browser wiring, manifest ships stealth off, session_status hint on/off/signed-in. The CLI stub now puts the warning on stderr with exit 0, which is what the real CLI does (verified live against agent-browser 0.27.1).

Not in this PR

• No lifecycle hook to close the daemon on agent shutdown: a persistent daemon is now handled gracefully and keeps the signed-in window alive across restarts, which is the behavior we want.
• The adopt decision trusts the marker; a daemon that died and was recreated under the same session name by something else would be adopted. Session names are per plugin, so this is accepted and documented in browser.py.

🤖 Generated with Claude Code

https://claude.ai/code/session_01X1Mou2gxw6Wz9dLCmVocoL

[protoreview]

QA panel review — WARN

code-review-structural · head 288302f8a2e3 · formal

Low-risk change: a single minor correctness issue in browser.py's close() path. The fix is straightforward — check the _run result (or wrap in try/except) before calling _clear_marker(). No panel disagreement; the lone finding was confirmed by the verifier. Verification was diff-only (direct file reads unavailable), so the structural pass was effectively skipped — the finding's grounding rests on the diff hunk alone.

Findings

	Severity	Location	Finding	Verified
🟡	minor	browser.py	close() unconditionally clears the launch marker after _run(self._cmd("close")) without checking whether the close succeeded; if the daemon is still running, t…	confirmed

findings JSON (machine-readable)

[
  {
    "file": "browser.py",
    "line": 0,
    "severity": "minor",
    "category": "correctness",
    "claim": "close() unconditionally clears the launch marker after _run(self._cmd(\"close\")) without checking whether the close succeeded; if the daemon is still running, the next ensure_session() finds no marker and misidentifies our own session as a foreign one, raising a confusing 'foreign daemon' error.",
    "evidence": "def close(self) -> None:\n    _run(self._cmd(\"close\"), timeout=self.timeout_s)\n    self._clear_marker()\n    self._session_ready = False",
    "verdict": "confirmed",
    "note": "Code confirmed in PR diff: close() discards _run's result and unconditionally calls _clear_marker(). Since _run returns a result (not raises) per the ensure_session() pattern, a failed close still clears the marker, causing _adopt_running_session() to hit the marker-is-None branch on the next call."
  }
]

[protoreview]

Promoting the WARN verdict for head 288302f8a2e3: all checks terminal-green, zero unresolved review threads. (approve-on-green)

Open findings carried by this approval — non-blocking, but they did not go away:

• minor browser.py:0 — close() unconditionally clears the launch marker after _run(self._cmd("close")) without checking whether the close succeeded; if the daemon is still running, the next ensure_session() finds no marker

Approving a WARN does not resolve its findings (issue #22).

[mabry1985]

Redesign after review (head cb6ca34). The first cut assumed the CLI's "ignored: daemon already running" meant the flags were dropped. The source and a live check show the opposite and worse: a URL-less open with flags sends a full-options launch and then a second, option-less one, so the daemon relaunches Chrome on a temp profile — every window 0.2.0 ever opened was abandoned within seconds. Fix: launch via open about:blank with the flags and carry the flags on every open <url>; the daemon reconciles (reused / relaunch with new options / respawn). The marker + adopt + session-list machinery is gone, along with Vera's close() finding.

Live-verified on the running ebay daemon (agent-browser 0.27.1): one relaunch onto ~/.protoagent/ebay-profile with --disable-blink-features=AutomationControlled, navigator.webdriver === false, repeat launch reused (same Chrome pid), navigation with flags does not relaunch.

[protoreview]

QA panel review — PASS

code-review-structural · head cb6ca3457731 · formal

Low-risk change: all four LLM finders and the structural pass found no defects. The prior-round minor finding about close() unconditionally clearing the launch marker was confirmed moot by the verifier — the code path it described no longer exists in the current head. No panel disagreement. No coverage gaps: the structural pass ran and returned clean. Nothing to fix.

No findings — the review came back clean.

findings JSON (machine-readable)

[]

[protoreview]

Promoting the PASS verdict for head cb6ca3457731: all checks terminal-green, zero unresolved review threads. (approve-on-green)

[protoreview]

QA panel review — PASS

code-review-structural · head e043044c247b · formal

Low-risk change: the only surviving finding is a nit-level docstring inconsistency in _as_bool where the documented falsy-string set omits "off". Fix first: add "off" to the docstring (or drop it from the code set if it was accidental). The panel did not disagree on any point. Verification confirmed the single finding without modification. No structural pass was skipped.

Findings

	Severity	Location	Finding	Verified
⚪	nit	tools.py:183	_as_bool's docstring lists only "false"/"no"/"0" as the falsy strings, but the implementation also treats "off" as False, so the documented contract understate…	confirmed

findings JSON (machine-readable)

[
  {
    "file": "tools.py",
    "line": 183,
    "severity": "nit",
    "category": "conventions",
    "claim": "_as_bool's docstring lists only \"false\"/\"no\"/\"0\" as the falsy strings, but the implementation also treats \"off\" as False, so the documented contract understates the accepted set (and no test covers \"off\").",
    "evidence": "Docstring: `\"\"\"A YAML/console flag: real bools pass through; the strings \"false\"/\"no\"/\"0\" mean False.\"\"\"` \u2014 but the code is `return text not in {\"false\", \"no\", \"0\", \"off\"}`, which also maps \"off\" to False; the test test_string_flags_from_a_settings_form_are_read_as_booleans exercises \"false\"/\"true\"/\"yes\"/\"\" but never \"off\".",
    "verdict": "confirmed",
    "note": "Verified in the PR diff: the docstring lists three falsy strings (\"false\", \"no\", \"0\") while the code's set includes a fourth (\"off\"); the test exercises \"false\", \"true\", \"yes\", and \"\" but never \"off\". All three sub-claims are accurate."
  }
]

[protoreview]

QA panel review — PASS

code-review-structural · head 5238ea6224bc · formal

The change is low-risk: it adds a stealth launch option, refactors flag assembly into _launch_flags/_open_cmd, retires the now-obsolete _IGNORED_MARKER check in favour of the new daemon-reconciliation model, and updates the docs to match. No defects surfaced across the panel or the verifier's structural pass. The two prior-round items (a minor on browser.py close-marker handling and a nit on _as_bool's docstring) were both below the blocker/major threshold and do not carry forward. Nothing to fix before merge.

No findings — the review came back clean.

findings JSON (machine-readable)

[]