# M8 — model and context window in the agent view: acceptance (2026-11-03)

Requests:

- In the agent view, set the model of new sessions and of running sessions.
- Each row shows the running session's model and its context window use.

## Shipped

| Item | Where | Check |
|------|-------|-------|
| Model and thinking for new sessions | M6b (`/model`, `/thinking` in the dispatch input, `dispatchModel`) | unchanged |
| `/model`, `/thinking` in a peek switch that session; bare opens a picker; an unknown model is refused in the view | `extensions/agent-view.ts` (`sessionSetting`) | test; live: `/model anthropic/claude-haiku-4-5` → the row showed it within one report, window 1M → 200k |
| A turn in progress keeps its current request | worker `ReplyQueue.applySetting` → `pi.setModel` | live: `/model` sent during `sleep 6`: the `toolUse` message came from claude-haiku-5-5, the final one from claude-haiku-4-5 |
| Worker reports `thinking` and `context` (`ctx.getContextUsage()`) | `lib/worker-state.ts`, `extensions/worker.ts` | tests; refreshed at `session_start`, `message_end`, `agent_settled`, `model_select`, `thinking_level_select`, `session_compact` |
| Thinking changes go out at once; context use alone is throttled like row text | `Reporter.schedule`, `stateChanged` | test |
| Daemon validates and stores them; old workers that send neither keep the last values | `daemon/main.ts` `acceptReport`, `contextFrom` | test |
| Row column `model · 43%`, one width per render so it lines up (right-aligned, ages padded); warning ≥ 80 %, error ≥ 95 %; share only, then nothing, when narrow (summary keeps 24 columns) | `lib/view-model.ts`, `AgentView.columnsFor` / `detailText` | tests at 100, 64 and 44 columns; alignment test |
| Peek line: `provider/id · thinking high · ctx 86k / 200k (43%)`; `ctx ? / 200k` after a compaction | `modelFacts` | test |
| `rowDetails: false` hides the column | `lib/config.ts` | test |
| Fix: respawn kept the dispatch's `--model`/`--thinking`, undoing a switch | `lib/store.ts` `respawnCommand` drops `--model`, `--provider`, `--thinking` for a written session (pi restores them from its `model_change` / `thinking_level_change` entries) | test; live: switched to claude-haiku-4-5 / off, stopped, replied → respawned with both kept |

Thresholds are fixed (80 / 95 %), not settings.

## pi API used (for forks)

- `ctx.getContextUsage()`: pi ≥ 0.49.0
- `thinking_level_select` event, `pi.getThinkingLevel()`: pi ≥ 0.71.0

Both are older than what pi-roost already needs (`parentToolCallId`, 0.99.0), so the
minimum pi version does not change.

## Review fixes (PR #15)

Two independent reviews (data path; agent view). Fixed:

- A switch is no longer reported as done when it is only sent. The peek says "asked … to
  use …", then "now uses …" once a report shows it. If no report shows it within 8 s, it
  says "still on …" (for a model, most likely no credentials in the background session's
  environment). A thinking level that pi lowered is shown as the closest one supported. The
  worker also checks `pi.setModel`'s `false` result.
- Context use after an assistant message is read once pi has saved the message (pi runs
  `message_end` handlers first), so the share is no longer one request behind. `/tree`
  refreshes it too.
- `/bg`, quitting into the background, `/fork` and `/resume` (origins `adopt`/`fork`)
  no longer pass the foreground pi's `--model`/`--provider`/`--thinking`: the session's own
  model wins, as on respawn. An explicit `/model` from the view still wins.
- The peek footer fits 80 columns again (`/model` is under `?`).
- The tone follows the rounded share (79.6 % shows as 80 % and is a warning).
- `formatTokens(999_500)` is `1M`, not `1000k`. `contextFrom` rejects a window that
  rounds to 0.
- The row is re-read when a picker choice is sent. A session that is starting gets
  "try again in a moment". `/model default` in a peek explains that a running session has
  no default.
- Tests: the alignment test really has ages of different widths and a CJK model name.
  The reporter test waits for the throttled send instead of flushing with `close()`. New
  tests cover the switch confirmation, adjustment and timeout; shell rows; and
  adopt/fork flags.

Second review, fixed:

- The first round's adopt/fork change could lose a model or thinking level that pi never
  recorded in the session. pi saves a resumed session's model only once it answers, and a
  thinking level only when it changes. `/bg`, quit to background and `/fork` now pass the
  foreground's live model and thinking level. A respawn passes what the worker last
  reported running, checked before it becomes a flag.
- Context use is also read at `tool_execution_start`, after the save whatever other
  extensions do in `message_end`. `warn` survives a stale ctx.
- Switch checks are kept per session *and* setting, so a `/thinking` no longer cancels the
  check of a failing `/model`. A switch that replaces a pending one ignores the first one's
  report. The timeout checks the row again, and a reconnect checks the listed rows. The
  timeout notice fits 80 columns. A row deleted while the picker is open is refused.
- Tests: a refused `setModel` (`false` and a throw), a concurrent model and thinking
  switch, a superseded switch, the 80-column footer, the "starting" and "gone" messages,
  and the last-reported respawn flags.

Third review, fixed:

- If a hand-off's worker ends before its first report, the respawn falls back to the
  model and thinking level it was dispatched with (`rec.model` / `rec.thinking`).
- A test checks that `/bg` and quit to background pass the live model and thinking
  level, and not this pi's original `--model` / `--provider`. The respawn wiring has a
  daemon test.
- The timeout notice puts its hint first (`model switch failed (credentials?
  envAllowList): still …, not …`), so a long model id is what gets cut.
- Superseded switches are a list. With three switches in a row, only the last value
  counts as the outcome, and the same switch sent twice confirms at once.

Known limit: if a model's provider disappears (removed from `models.json`, or its
extension is uninstalled), a respawn's `--model` makes pi exit with "not found". The
crash restart is bounded to one try. Dispatch flags have always behaved this way. Start
the session again with another model, or remove the job.

Left as is: column widths are measured over all rows, as `nameWidth` is, so they do not
jump while you filter.

## Tests

Automated: 17 files, 177 tests (with M9 merged in). Live runs against pi 1.5.0 with an isolated
`PI_ROOST_HOME` and claude-haiku-5-5 / claude-haiku-4-5.