[Maya-Kid]

The record #470 asked for before any code: the six tasks and the fallback chain.

Why

Cut 1 of #470 was accepted on 2026-10-04 with "A short ADR first (the next free number is 0067) with the task list and the fallback chain, then the code." 0067 was held for #470 on #1130. #1213, which also wrote 0067, is moving to another number.

What changes

• docs/decisions/0067-each-task-chooses-its-model.md covers:
  • the problem and what a profile holds;
  • the six tasks, with the calls each one covers;
  • the chain that model_for(kb, task) follows, and model_for as the only way to a model;
  • the migration, in two sentences;
  • the cuts, what was not chosen, and three open questions.
• The index row under chat-and-mcp, with status Proposed.

No code, schema or behaviour changes. The two code PRs of cut 1 follow:

1. every call site goes through model_for, still reading today's row;
2. the profiles and the migration.

Open questions in the record

• Where profiles live. One model does every job in every base #470 says the deployment level, but each workspace has its own row and keys today.
• implication::read_phrases. Does it go under alignment?
• A base's unset tasks in cut 2. Do they follow the defaults, or does a base's own chat override cover them?

[WaylandYang]

Thank you, this records what we agreed on #470 and the chain is clear. embedding never falling back to chat, and a set-but-not-ready link ending the search, are both right: data should never reach an endpoint nobody chose for that task.

Answers to the open questions:

• Profiles stay per workspace. Keys and admins are per workspace today, and a deployment-level profile would let one workspace's data go to another's endpoint.
• implication::read_phrases goes under alignment.
• A base's unset tasks follow the task defaults, as #470 writes it.

Please fold those into the decisions. Two small additions while you're there: in decision 5, say that chat_reasoning_effort moves onto the migrated chat profile and what happens to embed_dim; in decision 4, note the one exception, the connection test, which builds a client from the profile being edited. Then the status can be Accepted and I'll merge it.

[WaylandYang]

Still waiting on the changes from the 2026-10-19 review: fold the three open-question answers into the decisions, add chat_reasoning_effort and embed_dim to decision 5 (an embedding profile needs its dimension, 0035), note the connection-test exception in decision 4, then mark the record Accepted here and in the README index. The numbering is fine: 0067 is reserved for #470, 0068 is #1222, 0072 is #1213. No commits since 2026-10-15.

[WaylandYang]

Friendly ping on the 2026-10-19 review: the record still has the three open questions that were answered (per-workspace profiles, read_phrases under alignment, unset base tasks follow the defaults), decision 5 needs chat_reasoning_effort and embed_dim (0035), and decision 4 the connection-test exception. With those folded in and the status set to Accepted in the record and the index, this can merge. 0067 is still free; if you can't get to it this week, say so and we'll make the edits on the branch.