Configure a step to explicitly allow failure

[ELLIOTTCABLE]

I'd really like a more granular passthrough than --no-verify.

In particular, in a lot of cases, I prefer my agents commit broken state that doesn't compile; but I also don't want that to ever happen blindly. Ideally, on a random normal commit, an unexpected build-failure stops the commit (as a tripwire to ensure somebody knows something is wrong); but when in the middle of work, I much prefer granular commits, even if the work cannot be brought green in a single cohesive commit.

However, other gates are absolutely not optional; I do not want builders skipping the critical deterministic checks that must be true of every stage of work. (For horrible YOOL 2026 'agentic development', this is usually "stop LLMs from doing stupid things", frankly.)

I can build some custom tooling around this, of course; but it feels like a natural hk feature. Declaring, per-tool/step, whether it blocks the commit, or simply reports/prints any errors but proceeds to allow the commit. (For my use-case, I'd like to configure that with something like a KNOWN_BROKEN=true git commit ... flag; but I don't much care how it's implemented, perhaps the ability to configure it that way is outside hk's tight scope.)

(Love hk and mise, you make good stuff, keep it tf up. o7)

[jdx]

Opened #1291 to implement this. It adds per-step allow_failure configuration, including an expression form for the KNOWN_BROKEN=true workflow. The step still runs and reports its non-zero exit, while other steps remain blocking. The implementation only allows tool command failures; hk configuration, rendering, spawning, and staging errors still stop the hook. Structured output marks the failed step with failure_allowed=true.

AI-assisted — Tool: Codex; model: openai/gpt-5; version: unavailable.