The bug

A fact the user states up front gets dropped from context after repeated compaction, and the agent then "forgets" it. Reproduced end-to-end against the real provider:

• A deploy token stated in the first user turn was present in every request through the first fold, then vanished from every request after the second fold — and the agent could no longer recall it (wrote no answer, never restated it).
• The smoking gun: the second fold's summary degraded to <compaction-summary>[assistant calls read_file] {"path":"data/f01.txt"}</compaction-summary> — the ## Goal was empty; the user's task and token were gone.

Root cause: planCompaction pinned only the system message (head = 1). The first user turn (the task brief + the user's stated facts/constraints) was always inside the summarized region, so its survival depended on a weak summarizer model re-extracting it correctly on every fold. One degenerate summary — which the model produces especially when re-summarizing an already-summarized region — drops the fact for good.

The fix

pinnedPrefixLen now keeps in the verbatim prefix:

• the system prompt,
• the first user turn when it's small enough to be a brief, and
• any prior compaction summaries.

So a fold never summarizes the task away, and a later fold never re-folds an earlier summary into nothing. A large first turn (pasted content) stays foldable — gated by an absolute token ceiling and a window fraction — so pinning never starves the context.

Verification

Same scenario, same small window, identical work:

Post-fold request structure after the fix: [system] [user: task verbatim, token present] [compaction-summary] [recent tail].

Tests: new TestPinnedPrefixLen (pins system+task+summaries; large/tiny-window turns stay foldable; a summary is not mistaken for the task turn); existing compaction/prune tests updated to the pinned-prefix behavior. go test ./internal/agent/..., control, and boot pass.