docs(AGENTS.md): add model selection & cost discipline section

Adds guidance on choosing between premium and cheap models for code-assistant work, including escalation/de-escalation triggers, sub-agent routing defaults, /fleet rules, and session-hygiene tips. Keeps premium reasoning for genuinely hard problems and routes mechanical work to cheaper models.

Assisted-by: GitHub Copilot:claude-opus-4.7

Author: felipebalbi

AGENTS.md

@@ -230,3 +230,92 @@ From `CONTRIBUTING.md` and observed history
 - `rustfmt.toml` references nightly‑only options; CI runs `cargo fmt
   --check` on stable, which warns about (but does not fail on) those
   options.
+
+## Model selection & cost discipline
+
+Premium models (Opus, GPT-5 family, "high"/"xhigh" reasoning variants)
+cost an order of magnitude more than standard models (Sonnet, Haiku,
+mini). Most steps in a typical task do not need premium reasoning,
+and over-using premium models wastes credits without improving
+outcomes. The rules below apply to *all* model selection: your own
+session, sub-agents launched via the `task` tool, and parallel work
+launched via `/fleet`.
+
+### Default posture
+
+- **Default to the cheapest model that can do the job.** Reach for a
+  premium model only when one of the escalation triggers below is hit.
+- **Plan with premium, execute with cheap.** Spend at most one or two
+  premium turns on design / planning, then downshift to a cheaper
+  model for mechanical execution of the plan.
+- **Never bump the model "just in case."** If you cannot articulate
+  *why* a cheaper model would fail, use the cheaper model.
+
+### Escalation triggers (use a premium model)
+
+Reach for a premium model when *any* of these are true:
+
+- Cross-module refactor, architectural design, or API design from
+  scratch.
+- Subtle correctness reasoning: concurrency, lifetimes, `unsafe`,
+  FFI ABI, cryptography, safety-critical control paths.
+- Debugging a failure that survived one prior cheap-model attempt.
+- Reviewing code on a safety-, security-, or money-critical path.
+- The diff cannot be predicted in advance — i.e. there is genuine
+  creative or design work to do, not just typing.
+
+### De-escalation triggers (use a cheap model)
+
+Use the cheapest available model when *any* of these are true:
+
+- Searching, reading, summarising files or docs.
+- Single-file mechanical edits: rename, format, lint fix, dependency
+  bump, boilerplate, scaffolding from a known template.
+- Generating tests for code that already works.
+- Running builds, tests, linters, or other commands where the model
+  only needs to report success/failure.
+- Routine commits, PR descriptions, changelog entries.
+- The diff is essentially predictable before generation.
+
+### Sub-agent routing (the `task` tool)
+
+When delegating with the `task` tool, set `model:` explicitly. Do not
+let sub-agents inherit a premium default for cheap work.
+
+| Sub-agent type    | Default model             | Override to                                     |
+|-------------------|---------------------------|-------------------------------------------------|
+| `explore`         | cheap                     | keep cheap (`claude-haiku-4.5` or `gpt-5-mini`) |
+| `task` (run cmd)  | cheap                     | keep cheap                                      |
+| `research`        | cheap for breadth         | premium only for the final synthesis            |
+| `general-purpose` | match task                | cheap for mechanical work; premium for design   |
+| `rubber-duck`     | premium                   | keep premium — this is where reasoning pays off |
+| `code-review`     | premium on critical paths | cheap on cosmetic / mechanical diffs            |
+
+### `/fleet` (parallel sub-agents) rules
+
+- Fleet mode multiplies cost by the fleet width. Apply the rules
+  above *per worker*, not in aggregate.
+- Split a fleet job along complexity lines: route the cheap,
+  parallelisable workers (file edits, test runs, doc updates) to a
+  cheap model; reserve premium models for the small number of
+  workers that need real reasoning.
+- If every worker in a fleet would need a premium model, the work is
+  probably not a good fit for fleet mode — reconsider the
+  decomposition before paying N× premium.
+
+### Session hygiene
+
+- Keep sessions short and focused. Long premium sessions are the
+  single largest source of waste because every turn re-processes the
+  full history.
+- Use `/compact` when the conversation grows long, and `/new` for
+  unrelated work.
+- Prefer `/ask` for one-off side questions so they don't extend the
+  main session.
+
+### When in doubt
+
+Ask: *"If a cheaper model produced the wrong answer here, would I
+catch it in seconds (compiler, tests, my own review) or in
+weeks (production incident)?"* If the former, use the cheap model
+and let the feedback loop do its job.