We should probably move this to a shared package, as it is used by aibridgepricesgen and https://github.com/coder/coder/blob/main/coderd/aibridge/prices/prices.go#L26C6-L26C13

And possibly other structs/method. Can be done in a follow-up cleanup PR 👍

yeah, I can fix it in a follow-up PR

P2 [CRF-2] Fourth verbatim copy of the price-seed schema; two adjacent sync-note comments still say the schema lives in three places. (Netero P2, Robin P2, Zoro P2, Ryosuke P3, Razor P3, Hisoka P4)

coderd/aibridge/prices/prices.go:23-25 and scripts/aibridgepricesgen/main.go:68-71 both declare that "the JSON contract for the price seed lives in three places that must stay in sync": the seeder struct, the generator struct, and the batch SQL upsert in coderd/database/queries/aicostcontrol.sql. This PR adds priceRow here as a silent fourth. json.Unmarshal ignores unknown fields, so when a new price column lands upstream (say cache_hit_price), readRows will drop it, samePrices will not compare it, and compare will report 0 changed for a real repricing.

Consequence: the weekly refresh renders "No price changes." over a real reprice, and the summary the reviewer is asked to trust is quietly narrower than the diff below it. The tool exists precisely so reviewers see what moved; this bug hazard defeats its purpose on the first schema extension.

Two fixes:

1. Minimal, matches existing pattern: update the two sibling NOTE comments to say four places and enumerate the diff tool.
2. Structural: extract the row struct into a leaf schema package (e.g., coderd/aibridge/prices/priceseed) that owns only the type and JSON tags, import it from both scripts, and delete the drift-warning comments. go build then catches the drift the comment was asking humans to catch.

Ryosuke's framing: three-comments-in-sync is a rule; one imported type is a fact. @ssncferreira's earlier thread suggested the same structural move as a follow-up; the panel is elevating this to P2 because the consequence (silent "No price changes" on future column additions) is real and customer-visible, and there is no ticket linked to the follow-up commitment. Needs a fix here, a linked ticket, or an explicit human decision to accept the gap.

🤖

P3 [CRF-3] Catalog-only refresh opens a PR titled "chore: refresh AI model price book" whose body reads "No price changes." (Hisoka P3, Kite P3, Meruem P3, Ryosuke P3, plus Mafuuu Nit, Pariston Note, Razor Note, Zoro Note)

Detect changes flips changed=true when either ${PRICES_FILE} or ${CATALOG_FILE} differs, yet Build pull request body runs aibridgepricesdiff against the two prices.json snapshots only. knownModelsGenerated.json carries contextLimit, maxOutputTokens, reasoningEffort, displayName, and aliases, none of which are in prices.json; when upstream changes one of those without touching prices, a PR opens whose body reads ## Price book changes\n\nNo price changes. The review note two lines further down claims "The summary above lists what moved", which is now false.

The class of bug is that "did anything change" lives in two places (the git-diff gate and the tool's price-only rendering) and the two can disagree. Two fixes eliminate the gap rather than paper over it:

1. Snapshot ${CATALOG_FILE} too and teach aibridgepricesdiff to name catalog moves alongside prices; rename the heading (## Snapshot changes or similar) so the label does not promise a price-only view.
2. Gate the PR-open decision on PRICES_FILE changing alone, letting a catalog-only drift wait for the next price move. Cheaper but delays catalog updates.

Option 1 is the more honest fix. Reviewers on this PR are the same for both artifacts, so the routing is fine; the mismatch is cognitive but predictable and repeated weekly.

🤖

P2 [CRF-4] Slack failure alert asserts one specific root cause for every failure mode in the job. (Leorio P2, Chopper P3, Mafuuu P3, Meruem P3, Pariston Note, Razor Nit)

At 2 AM:

⚠️ AI model price book refresh failed.
The generator fails by design when upstream drops a model pinned in scripts/aibridgepricesgen/overrides.jq or curated in curation.json. Logs: <run_url>

The step's if: is failure() || steps.detect.outputs.changed == 'false', so this branch runs when anything failed: Harden Runner, Checkout, Set up mise tools, Install pnpm dependencies, make gen/aibridge-prices for reasons unrelated to model drops, git push, the gh api calls (the PR description itself records that the verification runs hit HTTP 503 from api.github.com/graphql). The on-call reader is anchored to overrides.jq / curation.json before opening the run log, burns time confirming those files are fine, and only then reads the actual failure.

Diagnose, don't guess. Name the operation and run URL first; if a leading hypothesis is worth keeping, present it as one common cause, not the cause:

:warning: *AI Gateway price book refresh failed.* Logs: %s
A common cause is upstream dropping a model pinned in scripts/aibridgepricesgen/overrides.jq or scripts/aibridgepricesgen/curation.json. If the logs point elsewhere, that is the real cause; treat this hint as a starting point, not a diagnosis.

🤖

P3 [CRF-6] Checkout has no ref:, so a workflow_dispatch from a non-main branch smuggles that branch's history into the shared refresh PR. (Hisoka P3, Ryosuke Note)

Checkout has no ref:, so it defaults to github.ref. On schedule that is fine. On workflow_dispatch GitHub lets any actor with write pick a branch, and github.ref becomes that branch. Line 123 then git checkout -b "${REFRESH_BRANCH}" off that HEAD, git add only the two artifact paths, git commit, and git push --force refs/heads/bot/aigateway-prices-refresh. The remote refresh branch is now that topic branch's history plus one price commit. Line 142 creates the PR with base=main, so the diff shown to reviewers is main..<topic> for every file the topic branch touched, not just prices.

The implementation plan explicitly says workflow_dispatch "only fire from the default branch." It does not. Only maintainers can dispatch, so this is unlikely by accident, but the round-1 P1 (a temporary push: trigger for a personal branch, added because the author wanted to test from a branch) shows the pattern where testing from a non-main branch is real. A successor dispatching to test a change would open a refresh PR shipping unrelated commits. Reviewer opens a many-file diff titled "chore: refresh AI model price book" and either closes it in confusion or, worst case, merges it.

Two mechanical fixes, either sufficient:

1. Pin the checkout ref: add ref: main under with: on the Checkout step, or git fetch origin main && git checkout -b "${REFRESH_BRANCH}" origin/main before staging.
2. Guard the job: jobs.refresh.if: github.ref == 'refs/heads/main'. Costs a manual main re-dispatch to test.

🤖

Nit [CRF-5] Prose name drift: "AI model price book" here and in the Slack failure text (line 168) versus "AI Gateway price book" in the workflow header (line 1) and the heartbeat text (line 170). (Chopper, Gon, Leorio, Ryosuke)

The PR description states the naming policy explicitly: identifiers stay aibridge, prose says "AI Gateway". The file header at line 1 follows that ("Refreshes the AI Gateway price book"), and the no-change Slack text at line 170 follows it too.

PR_TITLE: "chore: refresh AI model price book" produces the weekly commit message and PR title, so every future refresh commit lands in main's history spelling the object differently from the workflow that produced it. Line 168 has the same drift right next to the compliant heartbeat two lines down, which also fragments Slack search ("AI Gateway price book" misses the failure alerts).

Change both to "AI Gateway price book".

🤖

Note [CRF-14] Cancelled runs go silent because the notification step gates on failure() || changed == 'false', and failure() does not fire on cancellation. (Mafuuu)

If a scheduled run is manually cancelled or superseded by the concurrency group, no Slack signal is emitted. The workflow's stated intent is "Failures are announced in Slack"; cancellation is not a failure, but for a weekly cron whose whole point is to not be silent, on-call may want to know a run was killed mid-flight instead of finding out at the next heartbeat. Consider if: always() && steps.detect.outputs.changed != 'true' with a cancelled branch, or accept the gap and document it in the header comment.

🤖

Note [CRF-7] New rule silently displaces @DanielleMaywood as an owner of knownModelsGenerated.json. (Netero)

CODEOWNERS uses last-match-wins. Line 31 (site/src/pages/AgentsPage/ @DanielleMaywood, unchanged by this PR) previously covered the frontend catalog. The new rule at line 49 lists only @evgeniy-scherbina @ssncferreira @johnstcn, so any change to that file (workflow or manual) now requests review only from those three. The PR body discusses ownership of curation.json and overrides.jq but not this narrowing.

Likely intentional (the file is generated and the new owners are the pricing-data reviewers), but the surrounding comment does not state the intent and the displacement is invisible in the diff. Confirm this is the intent, and add a line to the comment noting that the frontend owner is deliberately excluded from generated catalog updates.

🤖

Nit [CRF-8] sort.Slice is superseded by slices.SortFunc since Go 1.21; .claude/docs/GO.md names slices.SortFunc as the modern replacement. (Ging-go, Meruem)

go.mod is on go 1.26.5. Drop the less helper and the sort import:

slices.SortFunc(keys, func(a, b modelKey) int {
    return cmp.Or(cmp.Compare(a.provider, b.provider), cmp.Compare(a.model, b.model))
})

🤖

Nit [CRF-9] diff type doc is missing a terminating period. (Gon, Leorio)

// diff is the full comparison between two snapshots diverges from the sibling docs in the same file (priceRow at 27-29, modelKey at 39). Add the period.

🤖

Nit [CRF-11] compare doc omits the deterministic-sort contract. (Leorio)

The comment says "classifies every model as added, removed, or changed." The function also sorts each output slice by (provider, model), and TestCompareSortsDeterministically locks that behavior in. render iterates each slice in order and depends on it for stable PR bodies. A reader working only from the doc would think the output order is iteration-dependent. Extend:

// compare classifies every model as added, removed, or changed, and sorts
// each category by (provider, model) so the rendered summary is stable.

🤖