# v2.1 implementation plan: fixes found while running v2

> **For the implementing model:** do not work from this file. The owner gives you one task file at
> a time. This file is the index for the owner and the reviewer.

**Goal:** close the two defects and one gap found while v2 ran, without new features.

- **01-cancel-record** — a delivered response is never recorded as a 499; cancellation is what
  the reverse proxy observed. Found 2026-09-25 by the intermittent `Errors:2` in
  `TestQueueFullIs503`; verified with a diagnostic build; the given
  `internal/proxy/served_test.go` reproduces it on every run.
- **02-props-loaded-only** (to be written after v2 merges) — the poller asks `/props?model=X`
  only for models `/v1/models` lists as loaded, because llama-server's router autoloads a model
  named in that query (`models_autoload`). See the v2 README's note of 2026-09-25.
- ~~03-timing-margins~~ — done by the owner directly (test-only work, no implementer task): the
  remaining ordering sleeps in `limiter_test.go`, `proxy_test.go` and `cancel_test.go` now wait on
  limiter state (`Queued`/`InFlight`); the one sleep left is a deliberate lower bound. Five clean
  `-race` runs of both packages except the 499 defect task 01 fixes.

**How this plan was made:** acceptance tests first; no reference implementation. The given test
for task 01 was run against the v2 tree (fails six of six) and against a throwaway fix that
follows the task's rule (passes four full package runs with the v1 cancel tests); the throwaway
was discarded.

## Global constraints

- Everything in `AGENTS.md`. Branch `v2.1` from `master` after v2 merges. One task, one fresh
  OpenCode session, one commit. Given files are copied and never edited.

## Changes during the run

(none yet)