Loading files
This file is pinned to the selected release. Download the original to open it in your design tool.
# Master state — the IDE for the last 1%
**Goal (reset by Dee 2026-08-16, supersedes the fab-readiness goal below):**
Autonomous Circuit is **Cursor for electronics**. Agents already do ~99% of a
board. The 1% left is the hard part — routing is NP-hard, and taste is human —
so the app needs a board IDE an electrical engineer can drop into. Quality bar:
**familiar to an EE**, i.e. Altium/KiCad muscle memory. Not full-featured;
enough to edit.
Standing instructions that came with the goal:
1. Keep building and testing IDE features.
2. The three example boards must come out of the app good to go, needing only a
couple of small human edits.
3. **Commit and push frequently.**
4. Run a standing board of judge/reviewers who *use every feature* and score it.
Keep iterating to raise the score.
The metric is the panel score: `docs/reviews/ide-panel-rubric.md` — seven
lenses, 70 points, ship bar written into that file. It replaces `fab.ready` as
*the* number this loop moves, without replacing `fab.ready` as a floor.
## Where things stand (2026-08-16, evening)
**The IDE surface is committed and green.** `6a209dc` landed the editor: edit
engine with a byte-exact inverse, semantic server edit + fast verdict
(`board_edit_apply`, `board_fast_check`), keymap arbiter, pointer arbiter,
right-click menu, shortcut sheet derived from the resolvers, Properties
editing, rotate. `ec67807` before it gave the suite a real DOM, which is the
only reason the wiring bugs of the last three days are now catchable.
- viewer: **958 pass, 0 fail, 9.4s**
- circuitpy: **391 pass + 58 subtests**
- All three example boards: `fab.ready: true`, 0 blocking, in the repo **and**
in the app's own project workspaces.
**The app workspaces were stale and are now synced.** The app reads
`~/.autonomous-circuit/projects/<uuid>`, not the repo. harness-puck and
terminal-keyboard were Aug 15 builds; hydrate-coaster was **Aug 11 with 3
blocking errors**. All three replaced with today's verified builds (old copies
moved, not deleted, to the session scratchpad backup).
**One real defect found and fixed doing that**: `board_fast_check` reported
`blocked` on all three fab-ready boards. Two legs report the same finding —
`harvest_circuit_json` grades a `pcb_trace_too_long_warning` element `warning`
from its suffix, `run_tscircuit_checks` grades everything `@tscircuit/checks`
returns `error` — and the build ends with `checks.dedupe` (first wins → the
warning copy) while the fast gate did not. Now it does. All three read `legal`.
## The fleet (2026-08-17, afternoon)
Twelve products, built with the app by AI electrical engineers who had never
seen this codebase. `scripts/fleet-status --write` regenerates the table in
`products/README.md` from each board's own sidecar, so it cannot flatter us.
**9 of 12 built boards are fab-ready — in the repo *and* in the app**, which
now agree. They did not before: a round-4 judge found the table claiming 9 of
12 while the app a human actually opens held 6 of 10, with two products not
installed at all. `scripts/install-fleet` syncs them and `fleet-status` prints
both counts side by side, plus a flag for any board whose verdict was earned by
a source that has since changed (compared by fingerprint, never by clock).
The three that are not ready — `macropad-6`, `pixel-badge`, `bench-i2c-scanner`
— all fail on the same defect, and it is inside `rp2040-core` rather than in
any board: a via in the QFN's 0.4mm-pitch pad field (ledger #53). Every
board-level lever is measured and ruled out: 5x, the automatic 10x escalation,
a hand-declared 10x, and a wider block gap, which at 9mm made a board *worse*.
It also blocks two green boards indirectly — `desk-air-monitor`'s dead button
and `sensor-node-mini`'s dead USB pair both have correct fixes that regress
into it (38, 12 and 3 blocking at the GPIOs tried), so both were reverted under
the floor rule.
Every builder also writes `work/ee-feedback/<slug>.md`, and that file is the
input to the next round of platform work — which is where nearly everything
below came from.
### What the fleet has cost us in defects, and what it bought
Seven fixes today, every one traceable to a board somebody was trying to
finish:
- **The router escalation was switched off for every board in the fleet.**
`_set_autorouter_effort` refused any source that already declared
`autorouterEffortLevel`, and SKILL.md tells every engineer to declare `"5x"`.
Three engineers hand-rebuilt at 10x, one at a time, guessing, while the
mechanism built to do exactly that sat disabled by the instruction meant to
make boards route. A declared effort is now a **floor** (ledger #51).
- **The planner's block gap was one number and `rp2040-core` needs another.**
Measured A/B on an identical composition: 2mm → **13 blocking**, 5mm → **1**
(ledger #52). Also measured, so nobody spends another build on it: **10x does
not move this class** — the escalation fires now and comes back 9 against 9.
- **A wire nobody drew leaves no element to be wrong about.**
`checks.floating_net_warnings` found dead wiring on four boards, three of
them `fab.ready: true`: `sensor-node-mini`'s USB pair terminates in air (it
composes `usb-c-power`, which has no data pins), `desk-air-monitor`'s button
is wired to itself, and two pixel chains never get their data in.
- **`board_fast_check` said "1 blocking" on a board whose real number was 3.**
It cannot run KiCad — that needs a compile — so `drc_violation` is invisible
to it. It now reads the sidecar and reports `lastBuild.invisibleHere`; the
IDE chip says "1 blocking · 2 unseen", and a *clean* answer on a board with
standing KiCad blockers is amber rather than green.
- **The effort pill never sent the effort.** `app_set_effort` landed
server-side with settings, driver flag and review loop behind it, and no
client code named the command: the pill read "Max" while every turn ran at
the CLI's default.
- **`project_create` ignored a flat body** and answered 200 with a project
called "New project" — two engineers named a board and got that instead.
- **Products' `blocks/` were gitignored** while the README inside them promised
"a frozen snapshot so gerbers stay reproducible". A fresh clone had no
snapshot at all, and a block one engineer patched so a board's two sensors
could take different I2C addresses would have vanished silently, leaving both
at 0x76 with no error anywhere. Now tracked, the way `examples/` always did.
### The pattern under five of those
Code that compiles, passes its own tests, is documented as landed, and is
**called by nothing**: `board_fast_check`, `onPlacementRotate`, `PadHeader`,
`editEngine.js`, `app_set_effort`. Four in one day. A unit test proves a piece
works; nothing proved the app reached it. `commandsAreWired.test.mjs` now reads
the server's command table and fails if a command has no caller in shipping
client code, or no written reason in `AGENT_ONLY`. Lesson H in
`docs/lessons.md`; the same question is still open one layer out, for a block
nothing composes and a check no gate calls.
## The score
**Round 1: 40/70. Round 2: 51/70. Round 3: 51/70. Round 4: 50/70** — rubric in
`docs/reviews/ide-panel-rubric.md`, write-ups in
`docs/reviews/ide-panel-2026-08-16.md` and `-08-17.md`, judges' own reports
under `work/ide-panel/round<N>/`.
Round 4 went **down one, and the tool is better for it.** Editing fell 8 → 6
and Board readiness 7 → 6, both because a judge *built* a case an earlier round
had only described, and both cases were real. Integrity went 6 → 8 with all
four claimed fixes re-driven and held. A score that only ever rises is
measuring the judges, not the product.
| Lens | R2 | R3 | R4 |
|---|---|---|---|
| Familiarity | 8 | 8 | 8 |
| Navigation & sight | 8 | 8 | 8 |
| Editing | 7 | 8 | **6** |
| Verdict | 8 | 8 | 7 |
| Integrity | 6 | 6 | **8** |
| Discoverability | 6 | 7 | 7 |
| Board readiness | 6 | 7 | **6** |
Every must-fix from round 4 was worked the same day. None of it counts as
score until someone who did not write it re-drives it — that rule is why the
number is trustworthy at all.
## The loop
`work/LOOP.md` still describes the fab-readiness machine and its invariants —
one build at a time, gate on parsed artifacts, never commit a board whose
sidecar is not `fab.ready: true`, `fab.ready` is a floor. **Those invariants
still hold.** What changed is the top of the loop: pick the next action from
the panel's ranked must-fix list, build it, test it, commit it, push it, and
re-run the panel to see whether the number moved.
## Log
- 2026-08-17 afternoon — Panel round 4 complete (50/70) and every must-fix
worked the same day. Fleet at **9 of 12 fab-ready, in the repo and in the
app**, which now agree because `scripts/install-fleet` exists and
`scripts/fleet-status` prints both numbers side by side — a judge found the
repo claiming 9 of 12 while the app a human opens held 6 of 10. Everything
converges on one defect: `rp2040-core`'s internal fanout (ledger #53), which
now blocks three products directly and two more indirectly, because fixing
their dead wiring regresses into it. A specialist is on the block itself.
- 2026-08-17 midday — Fleet to twelve products, 6 of 9 built boards green.
Seven platform fixes, all traced to a board somebody was trying to finish
(see the fleet section above). Ledger gained #51, #52, #53 and lesson H.
Suites: viewer **1009 pass**, circuitpy **424 pass + 58 subtests**. Five
agents in flight: two new products, one rebuild, one re-placement of the
three RP2040-blocked boards, one repair of the four boards with dead wiring.
- 2026-08-16 evening — Goal reset to the IDE. Suite floor re-established
(a hung `ShortcutSheet.render.test.js` was making the viewer suite take 615s
and report a failure; it pressed Escape on `window`, and Radix listens for it
in the capture phase on `document`). Five commits landed and pushed: test
harness, rail-width study, reserve+diffpair, fastcheck, the IDE surface.
App workspaces synced. Panel round 1 dispatched.
- Earlier history (three boards to fab-ready, the panel that kept dying on
context, the via-in-pad phantom): `docs/night-log-2026-08-16.md` and
`docs/lessons.md`.
26e259c5cThis file is pinned to the selected release. Download the original to open it in your design tool.