Skip to content

AbstractObserver 0.6.0

Choose a tag to compare

@github-actions github-actions released this 01 Oct 21:22
· 156 commits to main since this release

Changed

  • Launch / Automate: the workflow picker lists only what you can run. It is the kit's WorkflowPicker
    (ui-kit, unreleased) in its any-interface mode over GET /api/gateway/bundles: "Gateway default" (the coding
    agent's), then Shared (made available by your admin) and Mine (your own), each with its interface and
    version. Bundles your admin made unavailable are no longer offered; runs of them still show by name.
  • Release: a v* tag now also gets a GitHub release page carrying its CHANGELOG section (job
    github-release), like the framework's other packages. The page for v0.5.0 was created by hand.
  • Observe: the run opens on the run view (the Ledger tab), grouped by agent cycle. The Story tab is
    removed; what only it showed stays above the steps as one summary: status, duration, LLM calls, tokens, tool
    calls (and how many failed), the outcome in one paragraph (the error, or the final answer, or the saved
    summary), and Artifacts, Folder and Answer (when the run waits for you). Session files, the metric
    tiles, the chronology and the Summarize button are gone (artifacts stay reachable from System).
  • Steps are collapsed cards: kind, node, status, duration, the model of an LLM call, and one line (tokens
    in · out, the tool and its arguments, the event name, the child run). Expanded, an LLM call shows System, Messages (as sent),
    Tools offered, Response, Reasoning and Raw JSON, each foldable (▸/▾) with Copy; a tool step shows its calls and
    results. Replaces the old ledger cards and the cycles panel.
  • Prompts stored once are shown in full: a $slim placeholder in a finished step is rebuilt from the step's
    start record in the same ledger and checked against its sha256. When the start record is not loaded the step
    says "Prompt body not kept: "; when the checksum differs it says so.
  • Housekeeping hidden by default: status/progress events, waits, subworkflow resumes, memory bookkeeping,
    model loads and effect-less steps appear only with the All steps switch (one table,
    src/ui/run_step_kinds.ts). A failed step is always shown; a tool step with a failed result counts as failed.
  • Filters and search, in one toolbar row under the step count (wraps on phones): Cycles / Steps,
    All / LLM calls / Tools / Failed, a search over card text and expanded content
    (plain text, any case). The view, filter, All steps and search are kept in the address
    (#run/<run_id>?only=llm&all=1&q=…); #run/<run_id> links keep working.
  • The run picker counts one LLM call per call (its start and end records were counted twice).