Burndown Full

What problem this solves

A coding agent is given a plan, generates it, starts executing, and then stops before the change is actually complete across the codebase. The user is left with a half-migrated repo. This is not a motivation problem — it is a well-documented set of failure modes in long-horizon LLM agents:

• Context rot / long-context degradation. Model reasoning quality drops as the input grows, even within the advertised window. Information in the middle of a long context is attended to least reliably ("lost in the middle"). So as execution proceeds, the agent's grip on "what's left to do" erodes.
• Premature termination. When tool outputs and history get pruned or summarized to save context, agents lose task-level awareness ("how many items remain, am I near done?") and declare completion early. The fix shown in the research is to keep a condensed, persistent record of remaining work visible at all times.
• Plan-as-boundary error. A plan is a hypothesis about scope produced before the repo was fully explored. The real set of affected files is almost always larger than the plan's list (barrels, tests, stories, configs, re-exports, dynamic usages). Treating the plan's file list as the boundary guarantees an incomplete burndown.

The countermeasures below are drawn from how robust coding agents (Claude Code's TodoWrite, etc.) actually achieve reliable completion: enumerate before executing, externalize the worklist to a file so it survives context truncation, work in small focused batches, and define "done" by a fresh verification pass rather than by the agent's memory.

The core principle

The plan is a starting hypothesis about scope, never the limit of scope. Memory is not a source of truth — a fresh repo-wide search is. You are not done when the planned files are edited; you are done when a from-scratch search for the old pattern returns zero, and the whole project type-checks, lints, and tests clean.

Internalize this loop: enumerate → batch-execute → persist progress → verify from scratch → if anything remains, loop. Never exit the loop on a feeling of completion.