AI Flow
Public Made by Adomby adom
Adom's AI Flow: a tool to help the AI follow all of the steps it takes to build a board.
parts and sourcing: when a part's component page cannot be made, retry once with better inputs, then assess the part's long-term supply risk and propose well-documented alternatives before reporting failure
adom-aiflow 0.2.62 (latest) · adom-wiki 1.8.15 · adom-parts-search 2.1.5 · adom/core 4.15.3 · Adom cloud container (Linux)
Ask
In my words: AI Flow should always retry at least once instead of being so quick to report failure. If we cannot get the data needed for the wiki, we should assess whether the parts we are using are actually reliable for a long-term product, or whether we need to seek a more popular alternative that we can count on for years to come.
This is for the sourcing and parts steps (sourcing check, parts, the aiflow-sourcing skill): what happens when a part's component page cannot be made.
What happened
On 2026-11-10 my agent asked Adom's part factory (adom-wiki part request, one call per MPN, each with a datasheet or Mouser link) to build component pages for the 24 parts of a production board's BOM that had none. 10 were built, 2 are still queued, and 12 were refused. The agent re-requested two of them under their orderable MPN (both refused again) and reported all 12 as failures.
The refusal reasons, as the factory gave them:
| Reason | Parts |
|---|---|
| The datasheet has no land-pattern drawing | 2 |
| Two independent readings of the datasheet's land pattern disagreed | 2 |
| No datasheet found from the manufacturer or the distributors (one of the two requests carried a datasheet link) | 2 |
| No distributor sells this exact part number (one had a packaging suffix, one was not the orderable number; re-requested under the orderable MPN, one was again "no distributor", the other then hit "no land-pattern drawing") | 2 |
| The footprint drawn from the datasheet failed a geometry check (pads overlap) | 1 |
| The 3D model does not sit on the footprint (rotated or offset) | 1 |
| Stopped at the 3D model step (factory-internal) | 1 |
| Stopped at publishing the page (factory-internal) | 1 |
Most of these are worth a second try with different inputs: the factory-internal stops, the missing datasheet, the land-pattern disagreement (a manufacturer's own footprint or package drawing settles it), the missing land pattern (it is often in a separate package drawing or the manufacturer's CAD download). Some are also a signal about the part itself: a part no distributor sells under its number, or whose maker publishes no land pattern, may not be one to build a product on for years.
Today parts says only: a part with no page fails; make it (wiki-component skill) and run again, or give it from the project library. Nothing asks for a retry, and nothing asks whether the part is a good long-term choice.
Proposed behaviour
When a part's component page cannot be made (a part request ends cannot_build or review, or no datasheet, land pattern or 3D model can be found):
Retry once, with better inputs. Exactly one more attempt per part, with at least one input changed, never the same request again:
- another datasheet source (the manufacturer's product page instead of a distributor PDF, or the reverse);
- the orderable MPN (
adom-wiki part candidates <MPN>), without a packaging or reel suffix where the factory names pages that way; - the manufacturer's land-pattern drawing or CAD download (aiflow-sourcing section 5 order), or the manufacturer page as the link.
Both attempts and their reasons go in the run. This is one bounded retry, not a loop: the "one search per question, never retry" quota rule for vendor searches stays as it is.
If it still fails, flag the part as a long-term supply risk and assess it. Write a short assessment per part into the run (for example
design/supply-risk.json), each figure with its source and date:- lifecycle status (active, NRND, last time buy, obsolete);
- how many distributors stock it, and the stock depth at each against the build quantity and the expected product life;
- lead time;
- second sources: pin-compatible or footprint-compatible parts from other makers;
- how widely the package is used (common package, or one maker's own);
- whether the datasheet and a land pattern are published at all.
Propose alternatives before reporting failure. Up to three more popular, well-documented parts that do the same job: active, multi-sourced, in stock at the profile's preferred source, ideally already on the wiki or in Adom Basic Parts, with what each changes on the board (footprint, values, firmware). The human chooses; never substitute silently.
Only then report the part as failed, and lead the report with the decision the human has to make, for example: "U7: refused twice (no land pattern); supply risk high (one distributor, 300 in stock, single source); alternatives A, B, C".
Acceptance criteria
- A part whose page request is refused gets exactly one retry with at least one changed input, recorded in the run with both reasons; a retry with identical inputs is refused by the tool.
- A part still without a page after the retry is flagged in the
partsoutcome (for examplesupplyRisk: high | medium | low) with the assessment file beside it; every figure carries its source and date. - The outcome lists up to three alternatives per flagged part, each with lifecycle, stock, source and whether it already has a wiki page, and what it changes on the board.
partsandfinishreport a refused part only together with its retry and its assessment; the hint names the next command to run.- Nothing is substituted without the human's choice, and the choice is recorded.
- The same assessment can run on a part that came from a project library (
--project-lib), since a part with no public page is often the same part that is hard to buy. - The aiflow-sourcing skill describes the three steps, and the
partsfailure hint points to them.
5 Replies
Log in to reply.