Skip to content

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

Python
3.9 KB
Raw Blame

v0.0.267 (2026-08-19)

This release makes long-running jobs dramatically cheaper to supervise and makes prompt-cache behavior visible instead of implicit. It also lets /btw reuse the parent turn's cached prefix and strengthens the executable specification behind that cache contract.

Wait for exit, not the next byte

Background jobs now wait like grok-build: one bash_output(wait_ms>0) blocks until the process exits, rather than returning on each new output chunk or after 30 seconds. A measured CI watch had turned the old polling loop into hundreds of model calls; the model now sleeps while the process runs.

  • wait_ms = 0 (or omitted) remains an immediate snapshot.
  • wait_ms > 0 waits for exit or cancellation, with a 10-hour ceiling.
  • Existing 1…30000 values are promoted to that ceiling, so older 30-second callers automatically get completion semantics instead of a poll loop.
  • agent_output follows the same contract.
  • Tool schemas and spawn guidance explicitly say that one wait covers exit.

See ADR 0010 for the measured failure mode and trade-offs.

See what the prompt cache is doing

The new /cache HUD exposes the cache posture without revealing prompt contents, paths, or tool bodies:

  • stable-prefix hash and the latest cache read/write counts;
  • session hit rate, catalog mode, and the last named bust reason;
  • xAI conversation, session, and prompt_cache_key affinity;
  • the remaining opt-in levers for maximizing reuse.

/debug includes a compact row from the same snapshot. Prefix-changing mutations such as goal, playbook, compaction, persona, tool, and MCP updates are recorded where they happen, but count as a bust only when the next request's system-and-tools prefix actually changes. Stable catalog mode stays opt-in because earlier live measurement showed no benefit from making it the default.

/btw reuses the parent prefix

Side questions now send the same system prompt, tools JSON, and prompt_cache_key as their parent request, then append the side question as a user message. Tools remain advertised to preserve identical prefix bytes but are not executed by /btw.

In a live Grok 4.6 measurement, an append-only second turn reused 3,712 of 3,721 prior input tokens, and the following /btw reused the same 3,712 tokens (99.8% of the prior prompt). xAI requests also send the project-derived x-grok-session-id alongside the existing conversation and prompt-cache identifiers.

See ADR 0011 for the visibility and default-policy decision.

Cache behavior is executable and proved

Two new reference kernels, conformance suites, and Lean proofs document the cache contract independently of the production implementation:

  • Prompt prefix models the maximizing walk that preserves stable bytes and identifies the cacheable boundary.
  • Prompt stability proves the must-nots that would silently invalidate an otherwise reusable prefix.

The kernels are part of the normal conformance manifest, so future changes are checked against the same contract rather than relying on release-time measurements alone.

Reliability fixes included

  • Streamable HTTP MCP startup performs one modern tools/list probe and enters legacy initialization only when the server returns a legacy-shaped response, avoiding duplicate discovery and initialization traffic.
  • CodeDB Pro prepared paths use native separators on Windows.
  • CI parsers understand folded native-tool schema listings.
  • Hosted TUI cancellation and goal-loop checks are isolated from unrelated recap behavior.

Distribution

Prebuilt archives are available for Linux, macOS, and Windows on arm64 and x86_64. Both macOS binaries are Developer ID signed, hardened-runtime enabled, and notarized by Apple; SHA256SUMS covers all six platform archives.