Skip to content

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

Python
4.3 KB
Raw Blame

The routing release: every model call lands where you'd want, and shows its work

v0.0.236 closes the loop the last release opened: the harness now learns its delegation policy, traces every routing decision it makes, honors your flat-rate subscriptions in every spelling, and answers "where would this model land?" in two seconds without spending a token. Plus: graff speaks ACP — point Zed's agent panel at it.

Every routing decision has a paper trail (#372)

  • Every spawned worker — subagent tool, workflow phase, pipeline stage — emits shape= role= tier= resolved_model= source= policy_or_genome_id= to the tracer, the local DGM archive, and a new additive agent_route --json event. source distinguishes explicit-pin / persona / learned-policy / workflow-override / session-default / ladder, so "which decisions were static, which the root authored, which the fleet learned" is no longer a guess.
  • On top of the trace, a learned tier policy: scored runs accumulate (shape, role, tier, model) observations; tier resolution consults the exact cell, then the shape pool, then the ladder. A learned answer may only dominate the ladder's (≥ quality, strictly cheaper) — it can never trade up, and explicit pins bypass it entirely.

Phases learn as one; fitness stops mixing rungs (#376)

  • Workflow phases take ONE learned seat, resolved before fan-out — every worker in the phase runs the same model, so within-phase prompt comparisons stay clean while the phase itself re-seats onto a dominating rung. Live causal proof: two identical runs differing only in the local score archive — the seeded run re-seats every worker to gpt-5.6-luna, source=learned-policy, and executes there flat-rate.
  • Fitness rows are stratified by the resolved model: genome comparisons and promotions rank only within a stratum (sol/terra/luna previously bucketed identically), with legacy rows honestly quarantined as unknown.

Subscriptions win in every spelling (#377)

  • --model kimi-k3 used to seat the codegraff gateway — silently metered while a flat-rate kimi login sat unused, and the whole session's workers inherited the wrong provider. Provider resolution now matches family-prefixed spellings against the direct provider's native row (kimi-k3 → kimi's k3) before the gateway is considered; the same rule covers every deepseek-*/grok-*-style gateway spelling, and it resolves from the compiled table alone — a fresh machine with only a kimi login no longer errors on the obvious spelling.

graff route: the two-second billing answer

  • graff route <model>… dry-runs provider seating with zero API calls, through the same credential resolution, seat chooser, and billing classifier a real session uses: kimi-k3 → kimi/k3 · auth: OAuth/login · billing: subscription, flat-rate. Metered models print their $/1M; unknown and catalogued-but-uncredentialed models get distinct, actionable errors.

graff speaks ACP (#375, first half)

  • graff acp is a full Agent Client Protocol agent over stdio JSON-RPC: initialize / session/new / session/prompt with streamed agent_message_chunk updates and honest stop reasons, one conversation per session across prompts. Point any ACP host (Zed's agent panel) at the binary. Protocol-clean stdout — banners and emitters are fully suppressed in agent mode.

Sharper edges filed off (#369, #370)

  • Blocking pre_tool hooks can name their sanctioned replacement: an optional suggest field lands in the denial the model sees ("blocked by pre_tool hook: … — use instead: …"), turning a guess-and- probe spiral into a one-call recovery. Hooks got their first README documentation while we were there.
  • SDK generation is deterministic by construction: generate.py runs the harness against a scrubbed HOME, so per-machine catalog caches can never drift the committed MODELS list again.

Filed for what's next

  • #378: render mermaid fences as terminal diagrams (Sugiyama-style layout, grok-build parity) and generate architecture diagrams straight from codedb's call/dep graph — ground truth, not model recollection.
  • #374: extract the headless engine; #375's second half: A2A federation.

Full Changelog: https://github.com/justrach/codegraff/compare/v0.0.235...v0.0.236