# Behavioural tests: free to rewrite the code > I think we need a larger project to turn most of the tests into > behavioural ones so that we are free to rewrite the code. Planned 2026-11-03 by the user and the coordinator, from an inventory agent's survey. The `e2e` line (kbdin, mousein) is its first step. ## Decided by the user - **A test goes in and comes out through public surfaces only:** the 9P tree (pane files, `ctl`, `look`/`exec`, `log`, `/screen`, `kbdin`, `mousein`), the process (flags, daemon, attach, signals, terminal bytes), and files on disk and effects. A rewrite that keeps those keeps its tests. - **T1, in-process 9P, is the main tier.** The core is driven only through `serveFs` (src/ninep/testing.zig's `call`, `rd`, `wr`): as behavioural as a client on a socket, and milliseconds a case. - **Process-level tests (T2) get a Zig runner, not Python.** The Python suites (test/fs.py and the rest) stay as examples of scripting a session and are not grown. - **Later, not now:** - The reference as the spec: one test per documented behaviour, named by its section (docs/typ/reference.typ for the 9P surface, guide.typ and the cheatsheet for keys and mouse). A sample of 20 reference rows found 18 in some test; `restoredcol` and the EROFS mapping in none. - Phase 2 as a multi-agent workflow, an agent per area and a reviewer. ## Tiers | tier | drives | for | |---|---|---| | T1 | the core through `serveFs`, in process | most behaviour; each case also through a replica that must agree | | T2 | a real `pardes` over its socket, a Zig runner | daemon, attach, mounts, shells, flags | | T3 | terminal bytes, the `.snap` scripts | tty rendering and keys | | T4 | the window: `pardes-gui` hidden (`PARDES_TEST`, offscreen), its input through the same 9P files, frames captured once its status file says idle | what only pixels show | | perf | T1 or T4 gestures, timed | latency gates against baselines | The window (T4): its core is the one T1 drives, so a behaviour the cells show is tested at T1; T4 is for what only pixels show (scene effects, images, fonts, edges). test/gui_golden.py's 17 frame hashes (test/gui-goldens.txt) are an exact gate for one machine, never a CI one, and move to the Zig runner with T2. gui.zig's own tests: 19 M, 17 B, 47 U. Performance: test/perf.zig times gestures through `update` and `render` with a few internal reads (12), against test/perf-baseline-*.json (the perf-gate step); the PDF, rollback, locations and fs benches time components. A gesture is to be written as T1 or T4 input (kbdin, mousein), so a rewrite is timed by the same case it is tested by; a bench of one component stays a U-level bench of that component. U tests (a stable library at its own pub functions with no core: regexp, diff, sam addresses, walk, wire, Text, modal's parse) stay as they are. ## Inventory (2026-11-03, a regex classifier: about ±15%) | class | tests | lines | |---|---|---| | B, behavioural already | 252 | 6.3k | | U, a library's unit | 682 | 15.4k | | M, to migrate | 722 | 26.7k | | X, pinned to the implementation | 30 | 1.6k | Most M: pardes.zig 116, test/panes.zig 104 (of 124), ninep/ctl.zig 82 (in through 9P, checked through `p.panes`), test/output.zig 34, ninep/pane.zig 28, test/pdf.zig 25. What the M tests reach into, and the public surface for it: | internal | tests | surface | |---|---|---| | `p.active` | 242 | `/focus` | | a file's text | 173 | `pane/n/body` | | `body.cur_*`, `sel` | 126 | `dot`, `sel`, `addr` | | path, dirty | 121 | `name`, `dirty`, `/index` | | click cells from `p.rects` | 158 | **missing**: pane geometry | | `body.mode` | 90 | **missing**: a mode file | | Alt, Shift keys | 18 | **missing**: modifiers in `kbdin` | | notices, places | 28 | **missing** | | `.pdf.?` | 25 | **missing**: page, viewport | | terminal cursor, overlay | 60 | partial | ## Phases 0. Inventory. Done, above. 1. The T1 `Session` API (open, type, click, read pane files and screen text, expect; through a replica too), the missing surfaces in the table's order, and a lint that fails a behavioural test touching `p.panes[`, `.file.?`, `p.rects` and the like. 2. Migrate, in the order the code is to be rewritten: ninep/ctl.zig (its checks only), test/pdf.zig (step 3 of the pane-kinds plan), pardes.zig and test/panes.zig, replica, the gui and tty shells, host_io. A migrated test must fail on a build broken on purpose. 3. Gaps: a test for each documented behaviour with none. 4. Delete the M and X tests the new ones cover.