ir-refresh-aliases

name	ir:refresh-aliases
description	Sync irrlicht's model-name alias map against codeburn's upstream BUILTIN_ALIASES. Fetches codeburn's src/models.ts, diffs entries against core/pkg/capacity/aliases.go, and proposes additions/changes as a PR. Use when user says '/ir:refresh-aliases', 'refresh aliases', 'sync codeburn aliases', 'check alias map', or when a session prices at $0 for a known model.

Refresh Model Aliases from Codeburn

Irrlicht's core/pkg/capacity/aliases.go is a hand-translated port of codeburn's BUILTIN_ALIASES map. New frontends (Cursor releases, OMP variants, Antigravity Gemini models, etc.) appear upstream first; this skill keeps our map in sync.

Modes

This skill runs in two modes — the workflow below covers both; only the apply step (Step 7) differs.

• Standalone (default — invoked as /ir:refresh-aliases or "refresh aliases"): produce a diff report, edit aliases.go, run tests, and open a PR. Maintainer reviews the PR before it lands on main.
• Release-inline (invoked from /ir:release Step 1.5): produce the same diff, but edit aliases.go directly on the active release branch — no separate PR. The release commit ships the alias updates alongside the version bump. Fail-soft: if the fetch errors out (offline, upstream 5xx), continue the release with the existing map.

When to run (standalone)

• Quarterly cadence (no schedule — maintainer triggers).
• On demand whenever a session shows EstimatedCostUSD: 0 for a model the maintainer recognizes as a real provider (likely a new alias upstream).
• Before cutting a release, if /ir:release's Step 1.5 wasn't run for some reason.

Workflow

1. Fetch upstream

curl -fsSL https://raw.githubusercontent.com/getagentseal/codeburn/main/src/models.ts > /tmp/codeburn-models.ts

If the fetch fails, stop and report the error — don't proceed with a stale copy.

2. Extract upstream BUILTIN_ALIASES

awk '/^const BUILTIN_ALIASES/,/^}$/' /tmp/codeburn-models.ts

The block is a single TypeScript object literal: 'key': 'value', per line, with grouping comments. Parse into an alias → canonical map. Ignore comment lines.

3. Load current Irrlicht map

Read core/pkg/capacity/aliases.go and extract entries from var modelAliases. Each line has the form:

"key": "value",

Strip whitespace, comments, and the surrounding map[string]string{…} boilerplate.

4. Diff

Compute four sets:

• Added — keys present upstream but not in irrlicht. Default action: append to the Go map.
• Changed — keys present in both but with a different canonical target. Rare; flag for maintainer review (could indicate a model rename or a codeburn correction).
• Removed — keys in irrlicht but not upstream. Don't auto-delete; codeburn may have dropped something we still need. Flag with the entry contents.
• Unchanged — no action.

Skip LOCAL_OVERRIDE entries when classifying Changed. Lines in aliases.go ending with // LOCAL_OVERRIDE: <reason> intentionally deviate from codeburn because the codeburn-chosen canonical isn't in LiteLLM's pricing table. Treat the local value as authoritative; do not propose reverting to codeburn's value. Do still report the override entry under a separate Overrides section so the maintainer sees it once per sync, and if the codeburn-side value has itself changed (e.g., codeburn re-points the alias to a different canonical) — re-evaluate whether the override is still needed, since the new codeburn canonical may now exist in LiteLLM.

5. Validate canonical targets against LiteLLM

For each Added/Changed entry, check that the canonical value is a key in the running daemon's LiteLLM-derived table. The daemon caches it at ~/.local/share/irrlicht/model-capacity-cache.json:

jq -r '.config.models | keys[]' ~/.local/share/irrlicht/model-capacity-cache.json | grep -F "<canonical>"

Surface any aliases whose canonical isn't in LiteLLM — they will still resolve to ModelCapacity{} and need either a follow-up upstream LiteLLM entry or a different canonical choice.

6. Propose changes

Report to the maintainer:

codeburn BUILTIN_ALIASES sync (commit <sha-or-date>)

Added (N):
  "<alias>" → "<canonical>"        [LiteLLM: ok | missing]
  ...

Changed (N):
  "<alias>": "<old-canonical>" → "<new-canonical>"   [LiteLLM: ok | missing]
  ...

Removed (N) — not auto-deleted:
  "<alias>" → "<canonical>"
  ...

Overrides (N) — still diverging from codeburn:
  "<alias>" → "<local-canonical>" (codeburn: "<upstream-canonical>") — <reason>
  ...

If diff is empty: report "no changes" and exit.

7. Apply

Edit core/pkg/capacity/aliases.go — add new entries in the appropriate grouping section, update changed canonicals. Preserve the existing per-group comments. Don't delete Removed entries unless explicitly approved.

Run go test ./core/pkg/capacity/... -race -count=1 — the table-driven test enumerates every entry; new aliases need a canonical seed already covered, otherwise the test catches it.

Then, depending on mode:

• Standalone: commit on a fresh branch (chore(capacity): sync BUILTIN_ALIASES with codeburn) and open a PR with the diff report as the body.
• Release-inline: leave the change staged on the active release branch (the /ir:release Step 7a commit will pick it up). Mention added entries under "Fixed" in the release notes — /ir:release Step 1.5 has the suggested phrasing.

Constraints

• Exact-match only. Never propose prefix/fuzzy logic — see core/pkg/capacity/aliases.go header comment and the TestModelAliases_UnknownReturnsUnchanged test.
• Don't touch NormalizeModelName in core/pkg/tailer/parser.go. That's a separate (Anthropic-shorthand) normalizer with different semantics.
• Don't add runtime-fetch logic. The map is bundled at build time on purpose — pricing must be deterministic and offline-safe.