Promethist

Agent Debugger

Recent
{{ connectionStatus.label }}

Turns

{{ turns.length }} total

{{ loadError }}

Login Error
{{ authError }}
Access Denied
Your account is authenticated, but it does not have project permissions for the session you tried to open.
Authentication
Redirecting to Microsoft sign-in for live debugger access...
Loading session...
{{ hasAttemptedSessionLoad ? 'No turns found.' : 'Enter a session id to inspect a debugger session.' }}
Select a turn from the left pane to inspect components.
Tool Calls

{{ toolCallProject ? toolCallProject.name : 'How agents call their tools' }}

{{ toolCallProject.ref }}

Open a project to define when its tools should be called and to see how its agents actually call them.

{{ toolCallProjectError }}

Tool calls recorded in LLM logs for the sessions of the project started within the period. Relevance (false positives and negatives) is judged per turn against the published criteria of each tool.

Components and tools accept comma-separated names; * matches any characters, so handover_* selects the whole handover tool box.

{{ toolCallStatsError }}

Choose a period and press Show to see how the agents of the project call their tools.

Criteria

When a tool should and should not be called, per tool or tool pattern, for the whole project or one agent. Edits are saved as a draft; turns are classified only against the published revision, so a draft takes effect once it is published.

{{ toolCallCriteriaError }}

{{ toolCallCriteriaLoading ? 'Loading criteria…' : 'No criteria in this project yet.' }}
  • {{ item.toolPattern }} {{ item.agentRef ? `agent ${item.agentRef}` : 'whole project' }}
    {{ toolCallCriteriaStatus(item).label }} updated {{ formatDateTime((item.draft || item.published).created) }}
    Should call
    {{ truncate((item.draft || item.published).shouldCall, 600) }}
    Should not call
    {{ truncate((item.draft || item.published).shouldNotCall, 600) || '(not specified)' }}

    Shown: draft rev. {{ item.draft.revision }}. Classification uses published rev. {{ item.published.revision }} until the draft is published.

Markdown Editor
{{ markdownEditor.title }}
{{ markdownEditor.mode === 'wysiwyg' ? 'Use toolbar to format.' : 'Switch to WYSIWYG for formatting tools.' }}
Shortcuts: Esc close, Ctrl/Cmd+S apply.
{{ toolCallCriteriaEditor.draftRevision !== null ? `Editing draft rev. ${toolCallCriteriaEditor.draftRevision}` : toolCallCriteriaEditor.ref ? `New draft rev. ${toolCallCriteriaEditor.nextRevision}` : 'New criteria' }}

Tool call criteria

The judge reads these together with the conversation and the tool description to decide whether the tool should have been called in a turn.

{{ toolCallCriteriaEditor.publishedRevision !== null ? `Published rev. ${toolCallCriteriaEditor.publishedRevision} stays in use for classification until this draft is published.` : 'Not published yet: turns are not classified against these criteria until they are published.' }}

{{ toolCallCriteriaEditor.error }}

Criteria revisions

{{ toolCallCriteriaHistory.error }}

Loading…
  • Revision {{ revision.revision }} · {{ formatDateTime(revision.created) }} {{ revision.state === 'Archived' && revision.changeNote === 'Removed' ? 'Removed' : revision.state }} · {{ revision.changeNote }}
    {{ revision.toolPattern }} · {{ revision.agentRef ? `agent ${revision.agentRef}` : 'whole project' }}
    Call: {{ revision.shouldCall }}
    Don't call: {{ revision.shouldNotCall }}