Skip to content

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

Python
5.2 KB
Raw Blame

Codegraff v0.0.302.18

This release makes turns on the Responses wire shorter and cheaper, moves the fullscreen UI to its own client, and fixes session discovery and tool recovery. Measured with this release's code against the previous release and against Pi 1.0 with its codemode:

Evals

On a ChatGPT plan, interleaved over 11 tasks (five MCP tasks, a git digest and five solo coding tasks), 3 runs each, 8 at a time, with graff logged in to Codegraff as it is for its users. Cost is the list price of each run's tokens. Every arm passed every run.

gpt-6.1-sol: ahead of Pi with codemode on time, cost and cache hits

Pass Wall per task Median List cost per task Cache hit Output tokens per task
v0.0.302.17 33/33 40.8s 37.5s $0.0259 87.0% 874
v0.0.302.18 33/33 33.0s 26.8s $0.0192 89.8% 701
Pi 1.0 + codemode 33/33 34.2s 29.0s $0.0281 48.3% 612

gpt-6-astra: 14% less time than the previous release, the lowest cost of the three

Pass Wall per task Median List cost per task Cache hit Output tokens per task
v0.0.302.17 33/33 48.2s 38.2s $0.125 88.3% 767
v0.0.302.18 33/33 41.4s 31.7s $0.106 91.2% 693
Pi 1.0 + codemode 33/33 40.7s 32.9s $0.135 56.1% 696

On gpt-6-astra Pi is 0.7 seconds faster on the mean and graff's median run is faster; graff costs 21% less because most of its input is read from the cache. Logged in, graff also offers the model jev_effort, which asks the hosted Jev selector to choose the reasoning effort. Across 132 graff runs on the two models it was never called, so these are graff's default efforts.

Shorter turns

A model call's time follows the tokens it writes, not the size of the context it reads. graff's calls wrote more than they needed to:

  • Narration. Heads-ups are now one short line in the same response as their tool calls, instead of a short paragraph before most calls (ADR 0239).
  • Second derivations. A script that computes an answer is its own proof; the model no longer recomputes it a different way in the same script.
  • Reads of files about to be created. read_file is required only before editing an existing file.
  • rlm round trips. A slim view names the fields the whole value holds, the slim rule says to compute over results inside the same script, and an unbound call or each() prints its result (ADR 0240).
  • Version errors. The shell tool names the python3 the shell runs, read from where python3 on PATH resolves, never by running it (ADR 0242).
  • False stalls. Prose tightens the stall budget only while its output item is open. A model that writes a heads-up and then composes several tool calls in silence gets the full budget, instead of a reconnect that repeats the whole generation (ADR 0241).

The instructions cache across repos

The ChatGPT-plan backend renders the tools and then the instructions, and caches the instructions as one unit. They used to carry the project's instruction file and its layout, so a new repo's first call reused nothing of them. On the ChatGPT plan and the OpenAI platform those two blocks now ride as their own input item, and the instructions are identical in every repo (ADR 0243). Other providers send what they sent before.

The fullscreen UI is a separate client

The fullscreen UI is maintained in its own repository. graff tui now launches an installed graff-tui (the sibling binary first, then PATH) with its arguments unchanged, and never downloads one. In a terminal, graff and graff repl start the line REPL (ADR 0245).

Sessions and recovery

  • Live sessions show their activity age, list most recent first, and resume errors tell a live identifier from a saved key (#1459–#1464).
  • Saved workspaces stay discoverable after their processes exit, and cross-project resume enters the selected workspace first (ADR 0244, #1457, #1458).
  • Empty or format-only pastes recover a clipboard image or warn (#1468). Stale MCP registrations return recovery guidance, and disconnected tools leave the next catalog without any call being replayed (#1465, #1467). Account-backed model selections tolerate gaps in dynamic catalogs (#1466).