fix(local_browser): drop final_script.py workflow and bound ARIA history

[adamlu123]

What

local_browser.yaml's prompts were copy-pasted from base.yaml's bash_command-driven workspace flow, which doesn't fit a live persistent Playwright session. Two concrete bugs:

1. The agent was told to launch its own browser inside final_script.py while also being told never to import or launch Playwright.
2. It was told to shell out via run_command(...) — a helper that doesn't exist anywhere in the codebase. Any agent step calling it would die with NameError.

This PR reframes the live-browser mode around its actual capabilities and adds a context-bounding knob so ARIA snapshots don't dominate token usage as the loop runs.

Changes

local_browser.yaml (−221 / +127)

• Drop final_script.py, plan.md, self_reflect_config.json, final_runs/, image_qa, and self_reflection from the prompts. There is no workspace directory in this mode; the agent drives the live page and reports its answer in final_response when done.
• Override agent.require_self_reflection_success: false so the completion gate (which checks for final_runs/run_<id>/self_reflect_result.json) does not block done: true.
• Override agent.summary_user_prompt to talk about live browser state (URL, controls, selectors) instead of workspace artifacts (plan.md, final_script.py, final_runs/) that don't exist in this mode.
• Document in the yaml header what each step's observation contains (URL / title / stdout / ARIA snapshot / screenshot path) and note that the screenshot is NOT visually attached by default — flipping model.attach_observation_screenshot: true enables multimodal image input.
• Opt into ARIA-history pruning with agent.keep_last_n_observations: 1.

AgentConfig.keep_last_n_observations (+33 / -1)

• New knob in agents/default.py. Default -1 (disabled) — opt in per config.
• When > 0: strips the ARIA snapshot payload from observation messages older than the most recent N, both in the rendered text content and in the extra.observation dict. URL / title / printed stdout are preserved so the agent still knows what it did.
• Implementation: walks self.messages after each add_messages call, finds observation messages by extra.observation, replaces the saved aria_snapshot string in the text content with (ARIA snapshot pruned; see most recent observation).

Validation

Live end-to-end run against task: "Search for flights from SEA to JFK on 2026-08-15 to 2026-08-20" via base.yaml + local_browser.yaml + model_openai.yaml.

Metric	Without pruning (N disabled)	N=1 (this PR's default for local_browser)
Steps	9	7
Model calls	10	8
Exit	Submitted	Submitted
Final URL	google.com/travel/flights/search?tfs=…	(same shape)
Final response	"Google Flights results … SEA → JFK, round trip, Aug 15, 2026 to Aug 20, 2026."	"Flights searched successfully on Google Flights for SEA to JFK … Results page shows 11 results."
Trajectory size on disk	7.5 MB	2.2 MB

Trajectory inspection of the N=1 run confirms: observation #1–6 have ARIA stripped from both the text content and extra.observation.aria_snapshot; only observation #7 retains its full 15,282-char ARIA snapshot.

Test plan

• Existing unit tests pass (7/7).
• Live e2e against Google Flights with model_openai.yaml succeeds and reaches the results page in 7 steps with N=1.
• Trajectory inspection confirms ARIA pruning fires as designed.
• Suggested follow-up: apply the same prompt cleanup to persistent_browser.yaml, persistent_browser_always.yaml, crafted_cli.yaml, and task_showcase.yaml (same run_command / final_script.py issues are copy-pasted there).