Lewati ke konten utama
Versi: Saat ini

v0.0.69 — Structured agents + /release command

Released: 2026-05-14. Six commits.

Agent Builder — structured visual editor​

The Agent Builder gains a structured visual mode alongside the existing chat-style editor: a block-graph canvas with routing / core_loop / reflection / output blocks.

The runtime lands here too. POST /api/chat now branches when builderMode='structured_visual' and walks the graph via a new services/structured_agent_runtime.py:

  • Block executor with an 8-hop cap.
  • Routing uses temperature-0 label matching.
  • Reflection only revises when the model returns non-OK text.
  • chat.py loads the prior block from chat_sessions and persists current_block_id for subsequentTurnBehavior=continue. The workbench debug panel receives a structured_trace payload.
  • init_schema.sql gets an idempotent ADD COLUMN current_block_id.
  • A Beta badge sits on the structured editor until the runtime is validated end-to-end in staging.

Sixteen new unit tests cover start-block resolution, routing label match, reflection passthrough, hop cap, and a stubbed flow run.

/release slash command​

Wraps the full release flow from CLAUDE.md "Release & Tagging" so the next tag-and-deploy is one command instead of a manual checklist. Steps: pre-flight checks → annotated tag on both remotes → wait for CircleCI tarball → sha256 verify both sides → deploy test → confirm again before prod → write release memory entry.

Hard guards baked in: never push to one remote (CircleCI watches hubstudio), never honeyframe update without sha verify (v0.0.65 scp-truncation lesson), never deploy prod when test failed, never touch pilot from this command (VPN-gated). Discoverable from Codex via the .codex/prompts → .claude/commands symlink.

Research — competitive matrix​

Two Codex-produced flat feature inventories land in docs/_research/:

  • Dataiku DSS — 149 rows.
  • Databricks — 147 rows.

Both carry source URLs per row, captured 2026-05-14 against the official docs.

A new docs/competitive-feature-matrix.md maps both inventories against honeyframe primitives. Opinionated against "Dataiku UX + dbt engine" positioning — ML training, Spark, Delta Lake, Feature Store all marked out-of-scope by design rather than as gaps. The matrix calls out 8 honeyframe-only features (mock-data cards, width buckets, Figma Make import, agent dependency graph, etc.) and a 4-batch sequencing recommendation for post-v0.0.68 work.