Skip to content

graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs

Python
3.0 KB
Raw Blame

The harness learns: a closed evolution loop, bench-driven workers, and a sharper prompt

v0.0.234 is one arc: delegation and evolution become default properties of the harness, and worker assignment starts reasoning about capability, cost, and what you're actually logged into. Every behavior below was verified with live runs against real provider accounts before merge.

Workers: effort pins complete the persona story (#292 follow-up)

  • Personas can pin reasoning depth (effort: max) and a single spawn can override it — an axis independent of model/tier, so an effort-only override keeps the persona's model pin. model: gpt-5.6-luna + effort: max is exactly a "Luna Max" worker.
  • /agents now shows each persona's operational pins; /agents promote preserves them when the learning loop rewrites a persona file.

Evolution: the local DGM loop actually closes

  • Eval-driven scores now land in the local trajectory archive (genome + niche + score), not just backend telemetry — a default-local-privacy session finally feeds its own learning.
  • A target-met new best auto-promotes the niche champion and hot-reloads it mid-session: scored work breeds better workers with zero commands. Hand-written personas are never clobbered; only promote-written files are fair game.
  • VERIFIER RED now grants an in-loop repair turn (bounded, budget refunded on green) instead of ending a -p run at score 0 — two of three frontier models fell into the eval-as-baseline trap and now recover to green.

Worker economy: bench priors, subscriptions first, cost never escalates

  • A compiled-in DeepSWE v1.1 score/cost sheet seeds every install (./.harness/bench.json > ~/.harness/bench.json > builtin) and derives per-provider tier ladders: frontier by score, small by score-per-dollar — only for providers you are actually logged into or keyed for.
  • Explicit tier asks route to a logged-in flat-rate subscription first: marginal cost zero outranks any metered model, and the login is your standing consent for that vendor. A metered rung may descend price but never raise it; naming a model outright remains the escalation door.
  • --subagent-provider now loads that provider's login/stored key even when the root model can't route there — root gpt-5.6-sol with k3 workers used to fatal despite a valid Kimi login.

Prompt: the good parts of Codex's base instructions

  • The root prompt gains persistence-to-verified-done, work preambles, todo status discipline, root-cause editing constraints, and final-message formatting (path:line citations, verbosity scaled to the change) — in graff's own compact voice. Ultracode stays off by default.
  • todo_write and eval tool descriptions now teach the same discipline at the point of use, including the new RED repair-turn contract.

Also

  • SDKs regenerated for the subagent schema (effort) and tool-description changes.

Full Changelog: https://github.com/justrach/codegraff/compare/v0.0.233...v0.0.234