Skip to content

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

Python
3.2 KB
Raw Blame

The probe release: eight defects found by experiment, fixed the same day

v0.0.235 came from turning the harness on itself — three simulated coding exercises against v0.0.234 plus a field report from a real session, every finding filed with a repro and fixed with a live verification. Along the way main's CI went fully green for the first time since before v0.0.233.

Eval loop: scores you can trust (#367)

  • parseEvalScore no longer scrapes a leading number off prose: pytest -q's "4 passed in 0.01s" used to score a green suite 4/100 (an unwinnable loop), and "1 failed" parsed as 1 → the fraction rule → a false 100/100. The bare-number fallback now accepts only a whole-line number, and a command that prints no score falls back to exit-code semantics (0 → 100, else 0) with a verdict note — plain test runners work as eval commands out of the box.

The harness owns its last line (#368, #369, #366)

  • Budget exhaustion in -p now reports what ran out, that the work is PARTIAL and not rolled back, and the run's trace path — instead of a bare "turn failed: RunBudgetExhausted".
  • Unattended tool denials name what was denied and forbid the observed dead ends (retrying, probing harness config) — one model burned its whole call budget spiraling on an unexplained denial.
  • .harness/settings.json is refused to write_file/edit_file for root and subagents alike: a worker was caught trying to grant itself pre_approved_tools. A model never widens its own permissions.

Worker routing: the ladder finally behaves (#371, #373, #374 follow-on)

  • /model re-derives the default worker ladder — switching to sol no longer leaves children on the previous provider's rung (or with no descent at all, as a real 0.0.233 ultracode session showed). Explicit --subagent-* choices stay frozen.
  • Ladder rungs come off the Pareto front: no rung can be dominated by a cheaper rung below it. The builtin sheet now seats codex as sol / — / luna, so a sol root's default workers land on luna instead of a model luna beats on both score and price.
  • The DGM feeds the front: locally scored runs re-weight the sheet under pseudo-counts before seating, so lived experience reroutes future delegation — verified by test: bad luna runs demote it until terra re-seats as mid.

Quieter, greener infrastructure (#364, #365, #370)

  • The autosave skip-mark no longer leaks its path on every one-shot exit (it once aborted a run mid-flight).
  • The pty cancel test's exit window was the CI red since v0.0.233 — every cancel assertion had been passing all along. 30s headroom + transcript dumps on failure; the serve-resume test now tolerates torn reads of the deliberately in-place session autosave.
  • SDKs regenerate against a scrubbed HOME so local catalog caches cannot drift the committed MODELS list (#370 tracks making generation deterministic by construction).

Filed for what's next

  • #374: extract a headless engine — CLI/-p, REPL/TUI, --json, serve, GUI, iOS as adapters over one command core.
  • #375: speak ACP (editor agent mode) and A2A (agent-to-agent federation) as adapters over that seam.

Full Changelog: https://github.com/justrach/codegraff/compare/v0.0.234...v0.0.235