Troubleshoot Before Halt

This sprint had the riskiest repair boundary: ordinary red gates, post-review red gates, stale channel state, and first-rung escalation all had to move at once. The brief kept that work inside the inner-unit ladder, which made it small enough for a fresh agent to reason about without also touching halt policy. Its checks ran cleanly in the measured record, so the later sprints could treat the contract as established.

After the inner ladder moved, this sprint had a narrower job: bring the standing loop repair path under the same contract. That cut isolated generic prompt rendering from the sprint-specific path, so a fresh worker only had to prove the shared prompt could be filled outside the usual unit pipeline. The fact record shows no recovery for the sprint, which fits a port bounded by an already written contract.

This sprint protected old evidence while new runs changed their vocabulary. The brief separated compatibility parsing from live emission, which kept the agent from treating a historical stage rename as a product rewrite. It passed without a recorded recovery, leaving the documentation sprint a settled reporting contract to describe.

This was the consistency pass for the repair milestone. The brief deliberately pushed prose, prompt catalog entries, pattern comments, and generated snapshots after the behavior had landed, keeping the agent out of speculative documentation. The run facts show a green boundary, so no later re-plan had to revisit the repair story.

This sprint was a language-contract cut, not a recovery behavior cut. It asked the agent to add the header declaration and template-ready failure context while leaving patterns without the declaration unchanged. That made later dispatch work possible without giving the first sprint authority to decide stop policy.

This was the runtime dispatch sprint, where the hook stopped being metadata and became a guarded attempt before surfacing a halt. The brief narrowed the agent's authority to declared outcomes, green verification, handover reports, and recursion refusal. The measured sprint did not need a recorded recovery, which matters because the behavior itself is about only recovering when the engine can prove it.

This sprint applied the hook contract to the shipped build policy and to reviewer and executor routing. Its boundary was policy rather than syntax: decide which failures deserve troubleshoot, which blocked verdicts remain deliberate stops, and when verification is needless. The run record is green for the sprint, so the policy table landed without a later repair loop.

This sprint closed the halt-routing milestone by updating the prose contract for the agents that now receive hook context. Because the implementation was already split across declaration, dispatch, and default policy, the documentation worker could focus on authority boundaries and placeholder completeness. The measured record shows no recovery, so the wording pass did not uncover a structural gap.

This sprint established the pause-now contract at the lowest level: kill the in-flight process tree, preserve the worktree, and classify the stop as resumable operator pause. The brief kept it away from UI work so the fresh agent could concentrate on process and run-state semantics. The run facts show the sprint's stages completed even though the child finish row is not isolated with a commit, so the feature evidence has to come from the tree.

This sprint completed the engine-side pause contract by adding the wait-at-boundary mode. The brief narrowed the problem to recording intent, letting current work finish, and stopping before the next step begins, which is the right grain for a memoryless worker. Like the preceding pause sprint, its individual finish is not cleanly separable in the fact record, but the current implementation shows the contract landed.

This sprint put the completed pause contract into the live control surface. The brief asked for one action model across menus, keys, clicks, and enablement, so the agent could wire controls without redefining pause semantics. It passed without a recorded recovery, which left the UI as a thin layer over engine commands rather than a separate pause path.

This sprint swept the active prose surface after the implementation milestones had settled. Its scope was broad in files but narrow in judgment: remove active claims about retired prompts, channels, stage emission, and reviewer pass-through behavior while preserving historical compatibility notes. The green fact record suggests the sweep matched the already shipped contracts rather than exposing new behavior work.

This sprint handled the plan queue after the vocabulary sweep, which kept repository housekeeping from preceding the proof that the replacement design existed. The brief gave the agent concrete artifacts to remove or archive and a simple correctness test: the live queue should describe pending work only. It completed without a recorded recovery.