Reactive Agents repairs the almost-right tool call so the run keeps going

The FC-dialect probe is the part that makes the one-line model swap work. It picks native function calling where a provider supports it, and falls back to a tiered text-parsing driver where it does not. The dev.to post calls that "the actual reason a small open model and Claude can share one code path at all" ^[11].

Two layers sit on top of it. Model-adaptive context profiles tune prompt density and compaction per model tier, on the reasoning that a small model drowns in the same verbose prompt a frontier model handles easily ^[9]. A healing pipeline sits in front of every tool call and repairs the almost-right ones: a tool name off by a naming convention, a parameter sent under an alias, a malformed path. The call gets repaired and runs ^[10]. Those three failure modes are what you get when a call is parsed out of prose, so the probe and the healing layer are handling one weakness at two points.

The comparison on offer is a single incident. In both configurations the agent calls the health and recent-deploys tools, ties a degraded error rate to a deploy from twelve minutes ago, and recommends a rollback ^[6]. The post does not report a pass rate or a repeat count, and does not say how often the healing pipeline fires ^[17]. It is direct about the capability gap: "A 4B local model is obviously not as capable as Claude, and nobody working on this framework would tell you otherwise," the post said ^[7].

For that result to transfer, your tools would have to resemble those two. Both are read-only. Both are cheap to call twice. The conclusion is one hop from the evidence, because a deploy twelve minutes old is a strong signal in a short context ^[6]. Put a write tool, six services and two disagreeing metrics in the same loop, and the claim I would test first is that the harness finishes the loop regardless of which model is behind it ^[8].

The control claim gets concrete in the phase list. A run moves through bootstrap, guardrail, cost-route, think, act, observe and verify, continuing through termination, and every phase exposes hooks before and after it runs ^[12]. Seven named phases at two timings each is fourteen interception points before you write any provider code ^[16]. The example hook fires after the act phase, reads ctx.toolResults.at(-1), and logs the tool name ^[13].

The comment on the ollama line in the builder reads "your laptop, $0" ^[5]. That figure prices the API calls; the afternoon spent on prompt density sits outside it.