long-task-control

Long Task Control

Run a control loop that preserves verified work while correcting the course. Its governing principle: constrain the process, leave the tactics free — the gates below are hard, but how the agent improvises within them is not.

Reconstruct the target on every invocation from the original request, later authorized changes, and acceptance criteria. Newer authorized changes override older intent; keep unresolved conflicts visible.

1. Observe. Read the actual artifacts, checks, remaining work, history or rollout, and active agents — and trust these, not an agent's reported activity or self-claimed progress. Distinguish delegate liveness, observable progress, and completion: silence or elapsed time proves neither progress nor failure; active commands, dirty artifacts, and new check results are progress; only a terminal result backed by acceptance evidence is completion. Read that state from direct telemetry, and wait on defined checkpoints rather than prompting a delegate for running narration. Do not redo accepted work; optional work may be stopped once it can no longer change the outcome.

2. Compare. At each boundary, compare artifacts and trajectory against the target: missing outcomes, unsupported completion claims, scope drift, repeated loops, or effort that no longer moves acceptance. Redo this after every major phase, not only at the end.

3. Challenge — run an adversarial pass by default. At each major phase boundary challenge the work on two axes:

   • Design Challenger: ROI, design soundness, and gold-plating — scope creep, over-engineering, and redundant complexity the executor added on its own (加戏).
   • Drift Watchdog: alignment with the original intent, real progress since the last boundary, evidence quality, repeated work, and stale delegation.

   Finish any challenge that can change the contract, interface, or acceptance boundary before a writer begins; run only non-blocking guards concurrently. Run these as separate clean-context, read-only reviewers whenever the surface is non-trivial or the phase touched shared state; collapse to a single inline check only for a small, low-risk surface. Then give one owner the job of merging all findings: deduplicate, resolve conflicts, and classify each as actionable, boundary-only, overdesign, or false-positive. For each finding handed to a human, give its rationale with a code snippet so the call is reviewable item by item, not a bare label. Never staple reviewer outputs together as the decision, and let the agent that fixes differ from the one that reviewed.

4. Correct. Disposition each material finding — keep, prune, repair, replan, or escalate. When the Design Challenger flags gold-plating or drift, re-decompose the remaining work into commit-sized units before continuing, so the executor cannot smuggle in unrequested scope. Before interrupting an active delegate, inspect its status, dirty diff, active processes, and generated artifacts directly, and preserve and consume recoverable work; if an execution delegate with valuable context terminates with only a plan, give it one narrower outcome and exit condition before replacement. Reuse an agent whose context still serves the target; retire stale agents and their descendant work once it has no remaining outcome. Keep the three evidence types distinct — review yields candidate findings, runtime checks yield behavioral evidence, walkthroughs recover human understanding — and never let one stand in for another. After a correction, recheck the affected surface and any broader acceptance surface it may have disturbed.

5. Continue, and close on evidence. Resume work, but stop and replan the moment retries, polling, or agent activity produce no new acceptance evidence since the last boundary — the gate is "no new evidence," not a retry count. Interim passes review the current phase's diff; the terminal pass before closing reviews the entire cumulative diff, not just the last phase's. A typical rhythm: architecture-review + fix, then general-review + fix, once per module, then the same pair twice over the full diff at the end — adjust to the work, this is illustrative, not a fixed count. Before a turn, continuation, or handoff boundary, make every required delegated result terminal and consume it rather than assuming child-agent state will persist. Close only when that work is consumed and acceptance evidence supports completion; otherwise continue or surface the real blocker.