graff — a fast agentic coding harness in Zig: multi-provider, MCP, workflows, DGM evolution loop, TS/Python SDKs
Python
Document release and ratchet test count
Rach Pradhan4732296d786e
diff --git a/docs/releases/assets/v0.0.302.4/release-cover.png b/docs/releases/assets/v0.0.302.4/release-cover.pngnew file mode 100644index 00000000..86550e3fBinary files /dev/null and b/docs/releases/assets/v0.0.302.4/release-cover.png differdiff --git a/docs/releases/v0.0.302.4.md b/docs/releases/v0.0.302.4.mdindex e6183799..12b6d416 100644--- a/docs/releases/v0.0.302.4.md+++ b/docs/releases/v0.0.302.4.md@@ -4,7 +4,7 @@ This release makes long-running work more dependable, improves request recoveryand model selection, and brings the terminal and desktop sessions closer to thesame behavior. It also expands the offline checks used before a release.-+## Work that continues after a turn@@ -18,6 +18,13 @@ Background commands retain durable handles. When one finishes, its owningsession receives the exit status and output. A stale handle cannot accidentallyaddress a newer command. Worker resume also keeps the conversation's workspaceand settings. Saved sessions restore their history without rerunning old tools.+Positive worker-output waits now wait for completion; a zero wait still returns+a snapshot, and cancellation can interrupt either path. Aggregate tool limits+apply across descendant workers rather than resetting for each child.++Loaded MCP tools are available to RLM through their named tool calls, and+their ordering survives session save and restore. Independent workflow items+remain isolated while dependent stages share the workspace selected for them.Tool previews now preserve useful complete diagnostic lines from their sourceranges. Directory scans disclose when a bounded result is partial. Streamed@@ -40,28 +47,41 @@ original origin.Usage displays distinguish reported tokens and cost from attempts whose usageor charge was not reported. An incomplete total remains explicitly unknown;it is not shown as zero or as a complete subtotal.+An optional effort-selection charge is marked settled only when its response+contains a validated settlement receipt. A missing receipt leaves that charge+unknown, and a mock request does not enter billing totals.## Models and effortOpening the model picker refreshes the available catalog with a boundedrequest and a cached fallback. Saved or explicit selections keep their chosen-route when a refreshed catalog omits a model; changing routes remains an-explicit choice. `graff refresh` is a short form of `graff models refresh`.+route even during startup or when a refreshed catalog omits a model; changing+routes remains an explicit choice, never a silent fallback. `graff refresh`+is a short form of `graff models refresh`.Effort controls show the values supported by the active model. Models with anon/off thinking control show **Off** and **On** instead of an artificialintensity ladder. An unsupported explicit setting is rejected before a request,and changing models visibly resets a saved setting that no longer applies.-The optional effort-selection tool can queue a supported setting for its owning-session; it does not approve actions or judge the answer.+The optional effort-selection tool is available only when the active route and+model advertise compatible choices. While a selection is in progress or+waiting to apply, another call cannot overwrite it. A valid selection changes+the owning session at the next inference+boundary, not in the middle of the current request. A manual effort change or+model switch invalidates a pending selection. Errors leave the existing+effort unchanged. The tool does not run automatically on each turn, approve+actions, or judge the answer.## Desktop and ACPACP sessions expose their available effort choices on creation and load.Clients can change the same state used by `/effort`, and agent-originated-changes update the client. A change requested during an active turn applies-after that turn. Tool approval belongs to the active request and client;-cancellation cannot grant approval.+changes update the client. A client setting change requested during an active+turn is queued for the next turn; an agent-originated selection waits for the+next inference boundary. Neither rewrites a request already in flight. Tool+approval belongs to the active request and client; cancellation cannot grant+approval. The [ACP compatibility contract](https://github.com/justrach/codegraff/blob/v0.0.302.4/docs/acp-compatibility.md)+lists supported protocol behavior and client acceptance checks.The desktop restores composer focus when switching chats, supports transcriptselection and response copying, and keeps browser navigation errors accurate@@ -81,7 +101,7 @@ The image below previews a proposed replacement interface. **It is not thedesktop interface shipped in v0.0.302.4.** This release still ships theCodegraff desktop interface described above.-+## Install and verification@@ -90,6 +110,13 @@ desktop download is the signed, notarized DMG attached to the same release;the stable desktop update archive is checked against its manifest before therelease is published. Linux desktop packaging is included where supported.+The newest versioned release branch produces unique beta builds from its exact+source commit. Those prereleases do not enter the stable desktop update feed.+The stable tag triggers release checks and creates a draft. The release is+published only after those checks and desktop signing gates pass. The+[full changelog](https://github.com/justrach/codegraff/blob/v0.0.302.4/CHANGELOG.md)+lists additional fixes and their release context.+Offline checks cover formatting, source limits, test reachability, generatedSDK consistency, terminal behavior, desktop regressions, and scripted harnessbehavior. Local transport and ACP fixtures cover retries, cancellation, tooldiff --git a/scripts/eval/tier1-manifest.json b/scripts/eval/tier1-manifest.jsonindex 20da3b1f..5021495d 100644--- a/scripts/eval/tier1-manifest.json+++ b/scripts/eval/tier1-manifest.json@@ -15,7 +15,7 @@"src/repl.zig","TUI/root.zig"],- "test_count_baseline": 2776,+ "test_count_baseline": 2781,"test_count_slack": 25,"required_invariants": [{