Freerouting's candidate routes caused collateral damage (opens went from 38 to 11 only after the team kept just the four good routes as templates); a gated router adoption check refused the full candidate on all four adoption checks before it reached the board, and schematic parity passed with 53 nets agreeing both ways.
Recurring agent failure issue in balloon-fresh (felixfelix-bot)
What engineers report
A recurring agent failure problem in balloon-fresh (felixfelix-bot) that mostly hits the electrical engineer.
Failure pattern: validation / failure
Analyst note
From the Grok pain-radar digests, Nov 8-9 2026 (Nov 8 8:41 AM ET through Nov 9 6:41 AM ET digests plus the Haitch editions). Verified live 2026-11-09: issue/PR read; title and state confirmed.
What it costs a team
Modeled annual cost
$30,000
For a team of 5
Prototype delay
3 days
Modeled schedule slip per occurrence
Time lost
240 min
Per engineer per month, about 1 incidents
How the cost is modeled
240 min/month × 5 engineers × 12 months = 240 hours × $125/hour = $30,000
A modeled estimate with medium confidence. Swap in your own team size and loaded rate to size it for your team.
Engineering tools that address it
No engineering tool profile lists this error yet. Browse tools that connect to balloon-fresh (felixfelix-bot).
Evidence
More recurring errors in balloon-fresh (felixfelix-bot)
- A base-station board had 143 electrical errors. 74 came from a copper pour that was never refilled, and the shipped board was built from a superseded netlist: the ERC-gated netlist has 53 parts while the board had 44, with twelve missing including the whole MCU decoupling ring and three removed resistors still on it. A netlist integrity test already failed on master and nobody acted on it, and refilling the 44-part board would have produced a board that looked clean while missing those caps.ECAD · impact 74
- A dead boot pin was invisible in every count the project published. /IO0 was a single-pad net because the netlist was grown by text transform, and a one-pad net cannot be reported as unconnected. A renamed copy of the board silently lost its .kicad_dru rules and reported 537 violations instead of 16, and a sibling commit reported 15 to 12 from a work copy with a different baseline while the board of record went from 11 to 12. Four more nets have the same defect with no free GPIOs to fix them.ECAD · impact 74
- An agent working on a fabricated hub board traced 3 failing connectivity tests to the tests and tools, not the board. The test and the pin registry both pointed at a superseded board file that was never fabricated, which produced 18 phantom 'pin absent' and 18 phantom 'not on board' drifts. One real drift remained: a commit changed U1.27 on the board without updating the registry. Worse, the suite had never run in CI, because the CI job did not install KiCad and 8 of 10 tests skipped quietly.ECAD · impact 74
- Both CPL generators used KiCad's footprint anchor, which sits on pad 1 for pin headers, instead of the pad-array centre, so J_HUB_PV was 5.080 mm off and J_PROG 6.350 mm off and JLCPCB would have placed them where the pins cannot enter the holes. The existing check claimed verification to <0.001 mm, converted to JLC coordinates and back to the same anchor value, and the two generators also disagreed on board size, 72.10 x 80.10 vs 72.00 x 80.00.ECAD · impact 74
- Every one of 14 pads referenced F.Mask but the board had none, so the export gave 9 files and zero mask gerbers, with SS24 diodes on 1.6 x 1.9 mm lands next to the cell string. Regenerating the file silently reduced the board from 32 track segments to 7, and the fix was to restore from git and patch the layer table directly.ECAD · impact 74
- kicad-cli DRC 0 was unreachable by construction because 16 two-face wing pads and 3 GND zone stitcher records were counted as opens. A release gate that can never pass becomes a formality people learn to override, so exclusions are now derived from board geometry instead of a hand-written list.ECAD · impact 74
Sourced from public community reports. Costs and delays are modeled estimates, not measurements. Know a better source or a fix? Tell us at hello@productflo.io.