ldy/JobSourceAgent
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

agent-fallback

Browse Source
This commit is contained in:
ldy
2026-06-17 23:32:14 -04:00
parent 41037fb2f5
commit ed3ec1d2c6
5 changed files with 735 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
CLAUDE.md
View File

@@ -196,3 +196,8 @@ pytest -q
- LinkedIn rate-limits aggressively; keep batches small while testing.
- Standard pip struggles with pydantic dependency resolution in this stack; always use uv pip install instead.
- The system Python is protected by PEP 668 (externally-managed-environment). Always use explicit virtual environment paths (e.g., .venv/bin/python, .venv/bin/pytest) for all terminal commands instead of relying on global commands.
- **browser-use 0.8.0 `executable_path` required on Linux (confirmed 2026-06-17)**: `LocalBrowserWatchdog._find_installed_browser_path()` globs `chrome-linux/chrome` but playwright ≥1.40 installs to `chrome-linux64/chrome`. The watchdog silently falls through to `uvx playwright install chrome --with-deps`, which fails if `uvx` is not on PATH (`FileNotFoundError: 'uvx'`). Fix: pass `executable_path=_find_chromium_executable()` to `Browser()` in `_agent_session` — this makes the watchdog skip the auto-install entirely. `_find_chromium_executable()` in `agent_fallback.py` includes the `chrome-linux64` pattern.
- **browser-use 0.8.0 "Result failed" JSON errors are non-fatal step-level retries** (observed 2026-06-17): lines like `❌ Result failed 1/4 times: 1 validation error for AgentOutput Invalid JSON: trailing characters` mean the model's tool-call response had trailing characters and browser-use retried (up to 4 times per step). These do not abort the agent run; they appear in `history.errors()` but `is_done()` can still be `True`. Expect a handful per run.
- **Browser-agent run times: 2–4 minutes per company** (Infosys 21 steps/149s, TCS 24 steps/195s, tested 2026-06-17). This tier is a true last resort — only fires when all static tiers miss. The agent often uses a search engine (DuckDuckGo/Bing) when direct navigation fails, adding several extra steps. Google is rate-limited (CAPTCHA) and the agent falls back to Bing automatically.
- **Browser-agent partial result pattern (TCS, 2026-06-17)**: `is_successful()=False` with `careers_url` populated but `position_url=None`. TCS uses a custom ATS (iBegin) that requires login/registration before individual job listings are accessible; the agent found the portal but could not navigate to a specific posting. Such records land in `needs_review` as expected. Login-gated ATSes are a known hard limit of the static-HTML tiers and the browser agent alike.
- **Browser-agent live results (2026-06-17)**: Infosys → `careers_url=https://career.infosys.com/joblist` + `position_url=https://career.infosys.com/jobdesc?jobReferenceCode=INFSYS-EXTERNAL-247232` (both HTTP 200). TCS → `careers_url=https://ibegin.tcsapps.com/candidate/` (HTTP 200) + `position_url=None`. Both sites had failed all heuristic tiers (403 homepage / custom subdomain blocking httpx).