[BUG] browser tool cannot initialize new session after close action

[daviddmc]

Checks

• I have updated to the lastest minor and patch version of Strands
• I have checked the documentation and this is not expected behavior
• I have searched ./issues and there are no duplicates of my issue

Strands Version

1.10.0

Tools Package Version

0.2.9

Tools used

AgentCoreBrowser

Python Version

3.12

Operating System

macOS

Installation Method

pip

Steps to Reproduce

browser = AgentCoreBrowser()

print(browser.browser({'action':{'type':'init_session', 'session_name':'session-name-001', 'description': 'description'}}))
print(browser.browser({'action':{'type':'close', 'session_name':'session-name-001'}}))
print(browser.browser({'action':{'type':'init_session', 'session_name':'session-name-002', 'description': 'description'}}))
print(browser.browser({'action':{'type':'close', 'session_name':'session-name-002'}}))

output

{'status': 'success', 'content': [{'json': {'sessionName': 'session-name-001', 'description': 'description'}}]}
{'status': 'success', 'content': [{'text': 'Browser closed'}]}
{'status': 'error', 'content': [{'text': 'Failed to initialize session: Playwright not initialized'}]}
{'status': 'success', 'content': [{'text': 'Browser closed'}]}

Expected Behavior

Should be able to create new session after close

Actual Behavior

Failed to initialize session: Playwright not initialized

Additional Context

No response

Possible Solution

set self._started = False in the close action.

Related Issues

No response

[ethanhinson]

Confirming this still reproduces, and that it also affects the managed Bedrock AgentCore Harness, where the built-in browser tool is invoked by the model rather than called directly.

Reproduction, all inside a single harness invocation so there is no cold start between the steps:

1. init_session with session_name s1-session -> success
2. navigate to https://example.com -> success
3. get_text with selector h1 -> Text content: Example Domain
4. close session s1-session -> Browser closed
5. init_session with session_name s2-session -> Failed to initialize session: Playwright not initialized

Every later init_session on that same container keeps failing the same way. It does not recover on its own, and it survives an UpdateHarness re-provision, presumably because the warm container is reused. The only recovery observed was getting a new container: a brand new harness worked on the first try, and a previously broken one only worked again after a long idle gap.

This lines up with the source. close() calls _async_cleanup() directly, which sets self._playwright = None but never resets self._started. Only _cleanup() resets _started, and close() does not go through it, so _start() skips re-initializing Playwright and create_browser_session() then raises RuntimeError("Playwright not initialized").

One additional observation that may be worth folding into the same fix: the close action also leaves the AgentCore browser session alive on the AWS side. After step 4 returned Browser closed, ListBrowserSessions still reported that session as READY, and it stayed READY until it was stopped explicitly with StopBrowserSession. Since AgentCoreBrowser passes session_timeout_seconds=3600, each closed session can linger for up to an hour. So close currently tears down the local Playwright driver while leaking the remote session.

Environment: managed AgentCore Harness, us-west-2, built-in browser aws.browser.v1.

[kimnamu]

Hit this from the AgentCore side as well. It reproduces on current main (db4cf76) with the AgentCore control-plane client and the CDP connect stubbed out, so nothing in this run touches AWS — the failure is purely local bookkeeping:

1 init_session session-name-001 -> {'status': 'success', 'content': [{'json': {'sessionName': 'session-name-001', 'description': 'd'}}]}
                     state       -> _started=True, _playwright=<set>
2 close        session-name-001 -> {'status': 'success', 'content': [{'text': 'Browser closed'}]}
                     state       -> _started=True, _playwright=None
3 init_session session-name-002 -> {'status': 'error', 'content': [{'text': 'Failed to initialize session: Playwright not initialized'}]}
                     state       -> _started=True, _playwright=None
4 init_session session-name-003 -> {'status': 'error', 'content': [{'text': 'Failed to initialize session: Playwright not initialized'}]}

@daviddmc's Possible Solution is the right one, and @ethanhinson's reading matches the source line for line. close() (browser.py L940-946) runs _async_cleanup() directly, which drops self._playwright = None (L975), but _started is only reset in _cleanup() (L198) and close() never goes through it. That leaves the object at _started=True, _playwright=None, and the auto-start guard at L128 (if not self._started: self._start()) is the only caller of _start(), so Playwright is never brought back.

close() and _cleanup() are the only two callers of _async_cleanup(), so resetting the flag where _playwright is dropped covers both paths:

        self._playwright = None
        self._started = False

Same script with that one line, current main:

1 init_session session-name-001 -> {'status': 'success', 'content': [{'json': {'sessionName': 'session-name-001', 'description': 'd'}}]}
                     state       -> _started=True, _playwright=<set>
2 close        session-name-001 -> {'status': 'success', 'content': [{'text': 'Browser closed'}]}
                     state       -> _started=False, _playwright=None
3 init_session session-name-002 -> {'status': 'success', 'content': [{'json': {'sessionName': 'session-name-002', 'description': 'd'}}]}
                     state       -> _started=True, _playwright=<set>
4 init_session session-name-003 -> {'status': 'success', 'content': [{'json': {'sessionName': 'session-name-003', 'description': 'd'}}]}

tests/browser/ is 25 passed before and after, and ruff format --check / ruff check are clean on the file. Repeated closes are fine too — I ran init → close → init → close → close → init and all six return success.

Since close() sits in the base class this isn't AgentCore-specific: local_chromium_browser.py L65 has the same if not self._playwright guard raising the same message, so a local browser ends up in the same state after close.

PR #319 already contains that exact line (Closes #286, Closes #205), so this may be a review-bandwidth question rather than a missing fix.

On the AgentCore session left READY — that's a second, separate defect. _client_dict is assigned in AgentCoreBrowser.__init__ (L38) and read in close_platform() (L92), and never written anywhere else under src/, so close_platform() iterates an empty dict and client.stop() is never called. My run prints _client_dict == {} after a session was created, which fits the ListBrowserSessions output above. That one is #205, with PR #365 as the standalone one-line hydration; #319 bundles it with the _started fix.

Small thing to fold in whichever of those lands: close_platform() doesn't clear _client_dict, unlike _async_cleanup() which does self._sessions.clear(). With both #319 source lines applied, the six-step sequence above ended with stop() call counts of [3, 2, 0] across the three session clients — each earlier session gets re-stopped on every later close. self._client_dict.clear() after the loop keeps it to once.

Caveat on my side of the evidence: the network is stubbed in my repro, so it pins the local state machine and says nothing about the live service. @ethanhinson's managed-harness run is the half that covers that.