clean tests run: pi 1.5.0 model-list shim, panes checked by process, board r refreshes detail, office . accepted

docs/CLEAN-TESTS.md

@@ -127,3 +127,102 @@ only `README.md`, and cancelling removes it from the board without touching `mai
 
 For each step: what you pressed, what appeared, and anything that did not match. That is worth
 more than a verdict, because it says where the interface and Paperclip disagree.
+
+---
+
+# Results, 2026-11-12
+
+Both tests were run on `.13`, driven with `tuios send-text`, `tuios send-keys` and
+`tuios capture-pane`, so every step below is what the screen actually showed.
+
+## Test 1 — passed, after one real defect was found and fixed
+
+| Step | What happened |
+|---|---|
+| folder + `git init` | ok |
+| `office --new "test-alpha"` | project, session and eight panes created |
+| `New task>new` → **AUD-172** | created and handed to Builder; the Assign field takes the word `yes` |
+| first three runs | **all failed in 4–5 s**, `errorCode: adapter_failed`, `logBytes: 0` |
+| the fix | see "Pi 1.5.0 and the model list" below |
+| the run after the fix | **succeeded in 69 s** |
+| `Work>change` | worktree, branch, base, commits, `STATUS.md \| 1 +`, and the diff |
+| `Work>review` (`v`) | `ocr` **selected 0 files** — it does not review Markdown |
+| `Work>merge` (`m`, `y`) | rebased, no verify command, review recorded as "nothing to review", `merged into main at 36acb8d` |
+| `main` | `STATUS.md` present, contents `alpha is running.` |
+
+**What did not work as the test expected**
+
+- **Cost is always zero.** `Project>costs` and `Board>the chosen task` report `0 cent(s)` and
+  `0 tokens out` over four runs. `usageJson` on the run is `null`: the `pi_local` adapter does not
+  capture token usage from Pi, so cost cannot be shown for these runs. Open item.
+- **`ocr` reviews code, not Markdown.** A documentation-only change gets `status: skipped`,
+  `selected: []`. `bin/merge` already treats this as "nothing to review, recorded, not hidden",
+  so it does not block a merge — but a reviewer reading the tab sees "nothing reviewable" while
+  the `change` tab lists one file.
+
+## Test 2 — passed, after one real defect was found and fixed
+
+| Step | What happened |
+|---|---|
+| `office . --project harness-repo --session harness-repo` | adopted the project and the panes; `.harness/` was not rewritten |
+| panes after the reboot | **restored shells**, and `office` called them "adopted" |
+| the fix | a pane is adopted only if its own process is the harness |
+| `New task>new` → **AUD-173** | created and handed to Builder; the run succeeded in 24 s |
+| the diff | `README.md \| 1 +`, nothing else, committed on the branch |
+| merge | **not merged on purpose** |
+| `c`, then `y` in `Board>board` | task `cancelled`; status becomes cancelled, thread and runs stay |
+| `main` | untouched — `README.md` has no new line, last commit still `b81cb91` |
+
+**What did not work as the test expected**
+
+- **After a reboot, every pane is a bare shell and the pane keeps its name.** `board` matched the
+  panes by name and reported them "adopted" while nothing was running in them. A page that looks
+  built but does nothing is the worst kind of wrong. Now `open_pane` reads
+  `/proc/<pid>/cmdline` and adopts a pane only when the harness is what is running; otherwise it
+  closes the pane and opens a fresh one. `shell` is exempt: it is a shell either way.
+- **`Board>board` `r` refreshed the task list but not the selected task's detail**, so
+  "last run running" stayed on screen minutes after the run finished. `r` now reloads both.
+- A cancelled task is grouped under **Done** and its row reads `gone`. That is deliberate
+  (Paperclip records it as finished) but it is worth knowing before you look for it.
+
+## Pi 1.5.0 and the model list — where names live, and what changed
+
+Measured on `.13`:
+
+| File | What it holds |
+|---|---|
+| `~/.agents/models.json` (symlinked as `~/.pi/agent/models.json`) | provider `omni` with **476 explicit model names**, and `openrouter` overrides |
+| `~/.pi/agent/settings.json` | five providers with `baseUrl`, `api`, `apiKey`: `omni`, `deepseek`, `openrouter`, `opencode-go`, `google` |
+| `~/.pi/agent/models-store.json` | **discovered** catalogues per provider with `etag` and `checkedAt`, e.g. `deepseek` → 2 models |
+| a project's `.pi/` | no model names at all; for an agent workspace it holds only `.pi/tasks` |
+
+So model names are global, not per project. An agent workspace's `.pi/` cannot hold a stale copy.
+
+The names are there: `deepseek deepseek-flash`, `deepseek deepseek-v4-pro`, and through the router
+`omni deepseek/deepseek-v4-flash`, `omni deepseek/deepseek-v4-pro`, `omni ds/…`,
+`omni oc/deepseek-v4-flash-free`, `omni opencode-go/…`. The configured agent model
+`omni/deepseek/deepseek-v4-flash` resolves.
+
+**The change that broke Paperclip is not the names, it is the stream.** Pi 1.5.0 prints
+`--list-models` on **stdout**; Pi 0.8x printed it on **stderr**. Paperclip's `pi_local` adapter
+reads stderr first:
+
+```js
+// Pi outputs model list to stderr, but fall back to stdout for older versions
+const output = result.stderr || result.stdout;
+```
+
+Five extension warnings now live on stderr, so the adapter parsed those, found no `provider model`
+rows, and refused every run with `Pi returned no models`. The shim that fixes it,
+`~/.local/share/paperclip-harness/wrappers/pi-paperclip.sh`, puts the table back on stderr for
+that one flag and passes everything else through unchanged; the Builder agent's
+`adapterConfig.command` points at it.
+
+**Consequences to keep in mind**
+
+- **omniroute** is the source of the `omni` names. A name that is missing from the router's
+  `/v1/models` will not appear in Pi, and `models-store.json` caches with an `etag`, so a stale
+  entry can hide a new name until the catalogue is refreshed.
+- **headroom** keeps its own model registry (LiteLLM) and still has no entries for the router's
+  ids, which is why it answers `502`. It is unrelated to Pi's storage, but when it is fixed it
+  will need the same names.