fix(apodex): preflight search credentials from the profile's tools - #35
fix(apodex): preflight search credentials from the profile's tools#35Samurai007AK wants to merge 3 commits into
Conversation
The SERPER_API_KEY / JINA_API_KEY check was keyed on the research mode, which the terminal cannot select, so it never fired for react or agent_team, both of which bind web_search and web_fetch. A blank key first surfaced as an error string inside a tool result. Gate on the tools the profile declares, and drop the Serper issue to a warning so a coding session with no search key still starts. Closes ApodexAI#34
|
Thanks for the fix — the tool-based direction makes sense, but I see two remaining issues:
|
… in the TUI
Two review findings on the tool-based preflight.
Warnings were printed to stderr moments before Textual takes the alternate
screen, so under the default TUI they were gone before the session started.
cli.py now prints them only in line mode and hands them to FrontierAgentApp,
which renders them in the transcript after mount.
AgentProfile.tool_names read the top-level `tools:` of the apodex YAML, but
react and agent_team dispatch to a native workflow and never bind that list.
It is now a property resolving agent_tools / main_agent_tools /
sub_agent_tools from the workflow profile, through that workflow's own loader
so aliases and ${VAR} resolve as they do at dispatch. The duplicated `tools:`
lists are deleted rather than worked around: nothing read them, and a dead
copy is what drifts. agent_team's coordinator binds no web_search while its
sub-agents do, which the old list could not express either way.
The loaders log their own config diagnostics, and dispatch loads the same
profile again and logs them where the renderer routes them, so the preflight
read is muted to avoid printing each one twice.
|
@zhanghanduo Thanks for the follow-up. I addressed both points in the latest commit.
I also removed the duplicated The relevant tests were updated to cover the tool-based preflight and TUI warning visibility. |
|
One minor remaining issue: the new preflight tool resolution does not account for closed-book filtering. With Could you apply the same closed-book filtering before checking credentials and add coverage for both modes? |
|
@dq-ai-dev Addressed - thanks for catching it . Applied the same closed-book filtering before checking credentials:
Added coverage for both modes: unit cases for each flag + wrong-flag isolation, and end-to-end over the real |
3 participants
inspect_runtime_config() only checked SERPER_API_KEY and JINA_API_KEY when the
active mode was
research, and the terminal cannot select that mode:_TERMINAL_WORKFLOW_MODES caps the list at
reactandagent_team, and cli.pyrejects anything else before the preflight runs. Both selectable profiles bind
web_search and web_fetch, so the check never fired for the profiles that need
it. A blank key first surfaced as
[ERROR]: SERPER_API_KEY environment variable not set.handed to the model as a tool result, which reads as a failed call andgets improvised around rather than reported.
Gate the credentials on the tools the profile declares instead of the mode name,
and split them so each key is checked against the tool that needs it. Whether
web_search is bound is the actual dependency; the mode name was only ever a
proxy for it, and a wrong one.
The names come from a new AgentProfile.tool_names rather than from
profile.tools(). That field is a factory which imports apodex.agent_tools,
builds the whole registry, and raises KeyError on an unrecognised name — more
machinery and more failure modes than a local, secret-free preflight should
carry. tool_names is populated from the same YAML list the factory closes over.
missing_serper_api_key drops to a warning. It was blocking, but it never ran, so
nothing depended on that: making it fire as a blocker would newly refuse startup
(cli.py turns a blocking issue into exit 2) and break switch_session for anyone
running these profiles as a coding agent without a search key. web_search is one
of the seven tools these profiles bind, and the README describes the key as
optional — which stays true this way. missing_jina_api_key already warned.
One limitation worth flagging: cli.py prints warnings to stderr just before the
TUI takes the alternate screen buffer, so under the full TUI they may not remain
visible. Line mode (--no-tui) shows them normally. Surfacing them inside the TUI
means touching the renderer and seemed like your call rather than mine.
Closes #34