From 5b73af3aef0484c6b2db3ff40f24370fd0a2bba1 Mon Sep 17 00:00:00 2001 From: Tibo De Peuter Date: Thu, 10 Sep 2026 16:29:25 +0200 Subject: [PATCH] docs(copilot): add model selection reference --- copilot/model-guide.md | 25 +++++++++++++++++++++++++ 1 file changed, 25 insertions(+) create mode 100644 copilot/model-guide.md diff --git a/copilot/model-guide.md b/copilot/model-guide.md new file mode 100644 index 0000000..9a5ee0b --- /dev/null +++ b/copilot/model-guide.md @@ -0,0 +1,25 @@ +# Model selection guide (Copilot CLI) + +Reference for picking a model/agent per plan phase. Referenced from +`instructions/workflow.instructions.md` rather than inlined there, so this table can be +updated independently whenever the available model lineup or pricing changes. + +Update this table directly when the model lineup changes — no need to touch the workflow +instructions to do so. + +| Model | When to use it | 💡 Smartness | ⚡ Speed | 💲 Cost | +| --- | --- | --- | --- | --- | +| _(fill in current lineup here)_ | | | | | + +## Rough guidance + +- **Hardest, most open-ended tasks** (complex refactoring, ambiguous debugging, deep + architectural analysis, long multi-step reasoning): pick the highest-smartness model + available, accept lower speed/higher cost. +- **Routine, well-specified work** (mechanical edits, bulk reads/exploration, formatting, + straightforward test/lint fixes): pick a lower-cost, faster model — this is most of a + typical phase's work. +- **PR-feedback loops and unresolved-comment rechecks**: usually mechanical; only escalate + to a higher-smartness model if a review comment surfaces genuine ambiguity. +- When in doubt, the executing agent should state which tier it picked and why, so it can + be verified rather than silently over- or under-spending.