I contribute fixes upstream, not only to my own repositories.

The problem

The documented startup command for the SPI v2 backend named the wrong ASGI target, so a reader following the README exactly could not start the service.

What I changed

Corrected the Uvicorn target in the service README to the FastAPI application object. A one-line change, and the smallest contribution here — but a wrong getting-started command is the first thing a new contributor hits.

// resolves AcademySoftwareFoundation/dna#34

The problem

The MCP server's loop_estimate_cost re-implemented the loop-cost realistic-mix table by hand instead of importing it, and the copy never read pattern.cost.early_exit_required. For early-exit patterns that overestimated the realistic cost by roughly three times against the loop-cost library the documentation points users at.

What I changed

Added the loop-cost package as a sibling file: dependency — the same pattern mcp-server already uses for loop-audit, loop-gate and loop-context — and called its estimateCost() directly instead of maintaining a second mix table, so the two tools cannot drift apart again.

// resolves cobusgreyling/loop-engineering#434

The problem

The action spliced inputs.command into its run: script as raw text, before bash ever parsed it. For a multi-line command: | block that silently broke sandbox isolation — only the first line reached loop-sandbox run --, and every later line executed as a bare top-level statement on the runner, outside both the sandbox and the circuit breaker.

What I changed

Wrote the command to a temporary script passed through env:, so the workflow YAML never re-templates it, and invoked it as a single quoted bash "$SCRIPT" call on both the sandboxed and the direct path. The README gained safe-usage documentation and an explicit warning against embedding untrusted input in command.

The problem

A ModelRequest carrying both a ToolReturnPart and a UserPromptPart came out of AG-UI dumping in the wrong order — the user message first, the tool message after it. Providers that require a tool result to follow its tool call immediately, Anthropic and Bedrock among them, then rejected the replayed history.

What I changed

Preserved the original ModelRequest.parts ordering through dumping, with a _flush_user_content() helper that flushes buffered user content at ordering boundaries rather than at the end, and added regression coverage for the ToolReturnPart-then-UserPromptPart case.

// resolves pydantic/pydantic-ai#5964

The problem

Repeat count could only be set globally. Measuring how non-deterministic one test case was therefore meant repeating the entire evaluation, paying for every other test to run again as well.

What I changed

Added tests[].options.repeat, which overrides the global repeat for that test only. Each repeat index gets its own cache namespace — including index 0 — so repeated runs neither reuse nor overwrite the ordinary non-repeat cache entry. Scenario config now inherits all options fields, matching the existing vars/metadata/assert inheritance, with scenario test options still taking precedence. Invalid values fall back to the effective global repeat and are diagnosed by the public Zod schema and the generated JSON schema.

The problem

The doctor command gave no way to tell a missing optional dependency apart from a broken core installation: absent optional tooling failed the check rather than reporting itself, and Rich markup meant the extras being discussed did not even render literally in the output.

What I changed

Added deep-mode checks for the optional extras and for the semgrep, pip-audit and opa command-line tools, plus an API-key readiness warning. Missing optional tooling now warns instead of failing a core-only install, and the output labels are escaped so the extras render as written.

// resolves MCP-Audit/MCTS#212

The problem

The diagnostic reported a collapsed vertical channel — predicted y nearly constant while target y swept most of the screen — as a uniform offset, and advised the user to recalibrate. Recalibration cannot repair a collapsed axis. The function checked only direction consistency and magnitude relative to accuracy, and both failure shapes point every bias arrow the same way, so the uniformity test could not separate them.

What I changed

Regressed bias against target position independently per axis, which does separate them: a uniform offset gives a slope near zero, a collapsed axis a slope near minus one. The slope check runs before the uniformity check, a collapsed axis is now named as such per axis, and the message points at the eye-zoom and resolvable-step diagnostics instead of recommending recalibration. Existing thresholds were left untouched, and the one new threshold is justified against the two data points actually available.

// resolves shaal/eye-tracker#58