main
John Lauer Expand Fusion heatsinks, publish narrated thermal comparison and independent Kimi/Fable handoffs f7b1942 29d ago

Reproduce the Fusion + Astra demonstration

Start with the latest Adom Hydrogen build for your release channel. Check Settings → updates and compare the running and latest versions; apply an available update before testing prompt delivery. Install or update adom/codex (adom-wiki pkg install adom/codex) and open Codex from the dock. The first starter prompt is Codex Astra, reproduce this Fusion demo. Use Send to AI on Hydrogen 1.0.233 or newer. The picker defaults to Codex for Enter or the countdown, while letting you select another AI. The dashboard blocks older builds before sending; update Hydrogen or copy the prompt into a verified Codex conversation. Confirm Astra in Codex’s model selector before asking for fresh routing. A model subscription/login, Fusion installation and a connected Adom Bridge are required; the package does not supply those accounts.

The assistant should read this guide and the installed electrical-routing, ablation-routing and routing-video skills. Discover your desktop and bridge capabilities; never reuse CRR, another user's relay, HWND or document IDs. Preserve any pinned development bridge. Find the public adom/bq25792-charger reference and work on a new copy. The source/replay evidence lives on adom/fusion-bridge, branch feature/per-net-trace-routing, under demo/routing/bq25792/thermal/.

Choose one of two explicit outcomes:

  1. Replay the demonstrated result: use the accepted thermal-final/candidate-plan.json, with the original footprint geometry and matching comparison rules. Replaying saved coordinates is not a new routing calculation. Verify the empty copy, draw native copper trace by trace through the discovered Fusion bridge, refill, run native DRC and audit the saved result.
  2. Calculate a fresh result with Astra: inspect schematic/net membership, operating voltages and currents, stackup, thermal pad nets and existing vias first. Ask for missing manufacturing/load constraints while reviewing the board. Recalculate obstacle-aware horizontal/vertical/45-degree paths; reserve useful connected copper before ordinary routing; refine and recheck the plan. Do not substitute a replay for this request or invoke an external autorouter. Results may differ and must pass their own checks.

For the BQ example, keep the existing four layers and 35 MOSFET footprint thermal holes. The bottom thermal nets are VAC1 for Q1, VBUS for Q2/Q4, VAC2 for Q3 and VBAT for Q5. Ground fills the remaining permissible space. The demonstrated 0.1 mm isolation is this user's process floor, not a universal manufacturing or voltage rule. Retained floating islands require explicit permission and do not count as connected heatsinking. Switching nodes and sensitive signals need their own constraints.

Record the real Fusion canvas and existing Adom activity palette with native Bridge recording. Show an untouched reference and a stripped copy when demonstrating independent routing. Validate actual exported geometry, not only command acceptance. Preserve raw footage and failed candidates. Narrate with Adom TTS, clearly label precomputed planning/replay, 3x action and inspection holds, and deliver the MP4 to the selected desktop. Report inherited errors, rule-profile changes and unresolved electrical/thermal qualification.

Expected historical result for comparison: 576 segments, 89 vias, eight pours, no off-angle segments, zero airwires, zero warnings and one inherited clearance error. About 82.2% projected copper retention. These numbers are evidence for this board and rule profile, not a pass condition to force by suppressing rules.

Prompt delivery checks

The dashboard vendors the shared Adom newbie panel. Its version check and asynchronous send/status calls pass through the dashboard to the current host control API. A queued job is not delivery: wait for the reported delivery state and verify the prompt in the new thread. If a call times out, inspect the thread/job before retrying. Never overwrite another draft. The Codex dashboard uses a dedicated adapter that drops any incoming type and adds prefer:codex and background:true. The shared panel receives ai:codex and sends prefer itself. The generic proxy passes bodies and responses unchanged; the dedicated adapter changes only the outgoing provider request and preserves the upstream response. Astra still needs to be selected in Codex’s model selector; changing prompt wording does not change the model.

Desktop verification, September 6, 2026

CRR was updated to the latest available insider build, Hydrogen 1.0.230. The native dock launch and shared panel checks passed. Provider delivery did not: both picker selection and explicit Codex requests reached Claude, including a response that reported Codex. Hydrogen 1.0.232 corrected background-job intent scope, and 1.0.233 corrected Codex selection entering the Claude fast path. Codex package 0.1.16 now requires at least 1.0.233 and preserves the AI picker with Codex as its preferred default. The earlier 1.0.230 run remains failed evidence; it is not retroactively a successful test.

Latest continuation: expand and compare the heatsinks

Video 11 continues the video-10 routes with a separate thermal-copper allocation pass. Read the expansion report. The actual saved result keeps 576 segments and 89 vias and adds four larger native pours (12 total), expanding the four MOSFET-connected bottom regions from 126.7 to 765.7 mm². It preserves dedicated ground areas for IC1/U2 and both internal pours. This is a redistribution of copper; total bottom coverage decreases slightly to provide isolation. The assumed air-cooled comparison keeps solder mask. For a fresh reproduction, establish real loads and cooling conditions before choosing the allocation, and recompute the model rather than treating these example watts as a board rating.