Hoist llm.ask prompt binding to avoid elevated `p`; prefer reasoning_content when content is empty; point live/tests at in-tree MARQDO_EXT; add USTC live gates and mock agent.step selection+resume evidence; view document --out for read-only browser embed checks. Design stays Partial (not Verified). Co-authored-by: Cursor <cursoragent@cursor.com>
2.1 KiB
markup, title, description, import
| markup | title | description | import | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| occlusion | workbench F02 agent selection (mock upstream) | Explicit llm.create selection must drive agent.step; process env default model must not win. Upstream URL/key/model come from env set by the Rust harness. |
|
main
base = > sys.env_get name="OPENAI_BASE_URL" key = > sys.env_get name="OPENAI_API_KEY" mdl = > sys.env_get name="OPENAI_MODEL"
- not
base
print text=missing-OPENAI_BASE_URL sys.exit code=2
_ = 1
- not
key
print text=missing-OPENAI_API_KEY sys.exit code=2
_ = 1
- not
mdl
print text=missing-OPENAI_MODEL sys.exit code=2
_ = 1
Poison ambient default — selection on the handle must still win at the plugin.
sys.env_set name="OPENAI_MODEL" value="env-poison-model"
model = > llm.create model=mdl base_url=base api_key=key backend="openai-compatible" name=f02-agent
tools =
| 工具 |
|---|
ag = > agent.agent model=model tools=tools standing="F02 selection probe; answer with one word only"
step = > ag.step task="Reply with exactly one English word pong" writeback=False max_reads=1
st = > json.get value=step key="status"
ans = > json.get value=step key="result"
- not
ansans = > json.get value=stepkey="answer" -
_ = 1
- not
ansans = > json.get value=stepkey="text" -
_ = 1
print text=
stprint text=ans
dir = ".marqdo/workbench-f02-resume"
saved = > agent.resume_save id="f02-sel" goal="pong" status="running" round=1 resume_dir=dir
ok = > json.get value=saved key="ok"
ok
print text=resume-save-ok
print text=resume-save-fail print text=
savedsys.exit code=1
loaded = > agent.resume id="f02-sel" resume_dir=dir
found = > json.get value=loaded key="found"
found
print text=resume-load-ok
print text=resume-load-fail print text=
loadedsys.exit code=1
agent.resume_clear id="f02-sel" resume_dir=
dirprint text=f02-agent-selection-ok