hankweave adds harness switching for Agents SDK, Codex, and Gemini aliases

• Hankweave added short model aliases that switch the same job across different agent harnesses: hrishioa's launch thread says sonnet routes into Anthropic's Agents SDK, gpt-5.3-codex into Codex, and Gemini-style aliases into pi or opencode flows.
• The release makes harness choice a first-class config layer instead of a rewrite: the launch thread describes a "unified input interface" plus "unified logs, tracing and control interface" across backends.
• The practical motivation is orchestration, not model branding: in the same launch thread, hrishioa argues that shell calls, sandboxing, and file editing over "hundreds of loops" are hard enough that teams should "use the sdks" rather than rebuild each harness.
• Hankweave also added validation and a runnable multi-harness demo: according to the follow-up post, validation shows "exactly what's running where," and the demo repo backs a sample job that chains multiple harnesses with loops and budgets.

The new workflow is a translation layer over multiple agent runtimes. Hrishioa's launch thread says declarative "hanks" now map into different harnesses while keeping one control surface for prompts, behavior, logs, and tracing, with NDJSON as the shared output format. That narrows one recurring integration problem: ACP is described there as "rarely fully supported," so hankweave normalizes inputs and outputs instead of assuming protocol compatibility.

The release also treats backend selection as something you can swap with "a few characters" rather than a separate integration path. In the validation post, hrishioa says validation will show "exactly what's running where," which matters when one workflow mixes Agents SDK, Codex, and Gemini-adjacent harnesses. The sample multi-harness repo demonstrates that pattern with a single job that uses four harnesses, including loops and budgets, as a concrete example of cross-runtime orchestration rather than a single-model demo.