Skip to content

feat: mark a provider model retired when its provider stops offering it, and name the model that actually ran #852

Closed

Description

@HugoXK

Written-by: Fledge Platform Product

What a maintainer cannot do today

Tell that a model they assigned to a repository no longer exists at its provider. The gateway already recognises the case — src/gateway/modelFallback.ts:36-38 treats a 404 whose body says the model is "not found / not exist / unavailable / decommission / retired" as retryable — but that knowledge dies inside one call: the attempt is recorded, the call falls through to the next model in the repository's order, and nothing writes the fact anywhere a page reads. Settings keeps listing the model with a Delete button and no state (src/views/pages.ts:2231), the repository page keeps showing it as the default (src/views/pages.ts:1882), and the only trace is the single job that happened to fall back (src/views/pages.ts:3261). A maintainer with three repositories and dozens of jobs has no way to see that one of their models is gone and that jobs have been answering on a model they did not choose.

Why this card, why now

Competitor capability researched (daily 09:10): upstream providers now retire model ids on a published schedule and there is no quiet week. GitHub announced "Upcoming deprecation of selected GitHub Copilot models in December" on 2026-11-13 (https://github.blog/changelog/2026-11-13-upcoming-deprecation-of-selected-github-copilot-models-in-december), and in the same week Copilot swapped in Gemini 4 Carbon (2026-11-10), Claude Fable 5.2 (2026-11-09) and Gemini 4 Hydrogen (2026-11-13). Copilot's answer is an announced retirement plus a named replacement, so the user is told before their automation breaks. Fledge's answer today is a silent fallback.

Decision: BUILD, and this issue is the slice. Fledge composes with the primitive it already has: the attempt record (src/lib/modelRun.ts:10-15), the provider model table (src/db.ts:1491-1497) and the "Discover models" route that already reads the provider's live list (src/app.ts:4459-4478).

Environment

  • main = dddb518e5a598175b02f08b0cde33117204ce3d8 (live API read this round, Date: Mon, 16 Nov 2026 04:39:19 GMT).
  • Any provider whose /v1/models list has dropped an id the user added earlier (this is the normal state after a provider deprecation).
  • Schema at this tip: SCHEMA_VERSION = 44 (src/db.ts:65).

Resolved from the review

  1. retired on the attempt: built, not dropped. The red test keeps the assertion because the Expected paragraph needs it in one walkable step, so the slice extends ModelAttempt (src/lib/modelRun.ts:10-15), the attempt write (src/gateway/gateway.ts:321, serialised at :252) and the job model line (src/views/pages.ts:3261) with it. The Expected sentence is corrected too: today's job line already names the primary and the model that answered, so what is new is the word retired as the reason, not the two model names.
  2. Apply refusal: built here as slice item 6. bug: deleting a model provider silently unselects it from every repository, spends platform quota and never offers the restart #849 is closed — its work shipped as PR fix(ui): name the repositories that lost their model when a provider is deleted #837, merge sha 576bd9f691454b3a8b6afed6cdff0a1e655182fc (merged 2026-11-16T02:52:22Z) — and it marks a repository dirty when its provider is deleted (src/app.ts:4515-4522), so the apply control now appears. What it does not do is refuse an apply whose selected model is retired but still present; that is this card's item 6, and it is the reason a retired model is harmless once nothing can be applied with it.
  3. Schema move: stated in slice item 1 — SCHEMA_VERSION 44 → 45 (src/db.ts:65) and the four tests that pin 44 updated: test/scheduleSchema.test.ts:10,14, test/scheduleLeases.test.ts:48,52, test/scheduleOccurrences.test.ts:59,63, test/notificationOutbox.test.ts:45,49 (all four still read exactly as named at this sha).
  4. Non-blocking: the Discover route is src/app.ts:4459-4478 at this tip (the :2270-2285 range in an earlier body was the project-memory page).
  5. Second verdict (6906579127) folded in. The fifth pin moves to 46, not 45: test/db.test.ts:154-159 writes PRAGMA user_version = 45 at :156 under the comment "the first refused version is 45" at :154-155, then asserts the reopen throws /newer than this build/ at :159. At build 45 that row is accepted and the assertion goes red, so its value and comment become 46. And the Files: line below no longer says "one test": the slice edits five existing test files, test/db.test.ts included.

Expected vs actual

Expected: when a provider answers "this model does not exist" for a model a repository has assigned, Fledge marks that provider model as retired with the date, the Settings page and the repository's model line say so, the job that hit the 404 records the attempt as retired, the job's model line names retired as the reason, and a workspace cannot be applied with the retired model selected, src/app.ts:4326-4336.

Actual: the attempt is stored as {modelId, label, status: 404, retryable: true} (src/lib/modelRun.ts:10-15, pushed at src/gateway/gateway.ts:319-321, serialised at :252), the call continues to the next model, and no row, page or notice changes. A repository whose only model is the retired one fails the job with the provider's raw 404 text and still shows the retired model as its default.

Minimal reproduction

  1. Settings → add a provider and a model gpt-4o (the 1-token test testOwnModel at src/app.ts:501, called at :4545, passes, so the row is created at :4546 through src/db.ts:4873).
  2. On a repository, select gpt-4o as the default model and save (src/views/pages.ts:1882).
  3. Make the provider retire it: answer any chat call for that id with 404 {"error":{"code":"model_not_found","message":"The model gpt-4o does not exist"}}.
  4. Run a job. The gateway falls through to the repository's second model and the job succeeds; with exactly one model the job fails on the provider's 404 text.
  5. Open Settings → Models, the repository page and the job page: gpt-4o is listed as an ordinary model, still shown as the repository default, the job line says only gpt-4o (404), and nothing anywhere says it is retired.

Evidence at dddb518e

  • src/gateway/modelFallback.ts:36-38 — the retired-model 404 is already classified as a model-health failure ("A retired model is the one 'unavailable' a 404 can mean"), so the recognition exists and only the signal is missing.
  • src/lib/modelRun.ts:10-15 — ModelAttempt carries modelId, label, status, retryable; no field survives the call as "this id is retired".
  • src/gateway/gateway.ts:319-321 — the attempt is pushed and the loop continues; nothing writes back to provider_models. :252 serialises {model, attempts, fallback} on the job.
  • src/db.ts:1491-1497 — provider_models is id, provider_id, model, label, created_at; no retired column. Row type at src/db.ts:381-387, writer at src/db.ts:4873. The PRAGMA-guarded additive-column pattern is src/db.ts:2450-2451.
  • src/views/pages.ts:2231 — Settings renders each model as label <code>id</code> [Delete], with no state.
  • src/views/pages.ts:1882 — the repository default-model <select>; :1883 the per-bot selects, :1884 the "apply models and restart workspace" note, :1893 the apply form (still gated on repo.models_dirty === 1).
  • src/views/pages.ts:3261 — the job model line; triedText names each attempt as label (status) (:3244) and fellBackText names the earlier fallback (:3251-3258). Per job, not per model.
  • src/app.ts:4459-4478 — "Discover models" fetches the provider's live list and redirects with discovered=... (:4474); it never compares that list against stored models or selections.
  • src/app.ts:4326-4336 — the apply route: it calls instances.applyModels(repo.id) and surfaces a refusal through its existing ?error= redirect (:4330), which is where item 6 plugs in.

grep -rn retired_at src test scripts = 0 hits at this tip.

Proposed red test

test/modelRetire.test.ts (fakes only, no network, no live DB, no live SMTP):

  1. Fake provider answers 404 with code: model_not_found for model A and 200 for model B; one gateway call.
    • assert the job's model_run attempt for A carries retired: true and that B's does not;
    • assert the store.listProviderModels(provider.id) row for A has retired_at set and that a second 404 does not move the date;
    • assert the Settings HTML contains the retired badge and the date for A and not for B;
    • assert the repository page whose default is A renders "retired";
    • assert the job page's model line names retired for A.
  2. Guard the transient path: fake 429 and 500 answers leave retired_at NULL, leave retired unset, and keep the retryable set at src/gateway/modelFallback.ts:30-40 unchanged.
  3. Guard the apply path: with A retired and selected as the repository default, POST /repos/:id/models/apply refuses (the existing ?error= redirect, named message) and no restart happens; after the maintainer selects a live model, the same route applies.

Both halves to be kept: the test fails at dddb518e (no retired_at, no retired, no badge, the apply still succeeds) and passes with the fix.

Smallest mergeable slice (<= 1 week, T3)

  1. provider_models.retired_at TEXT NULL (+ the PRAGMA-guarded ALTER, pattern src/db.ts:2450-2451), a store method to set it once, SCHEMA_VERSION 44 → 45 (src/db.ts:65), the four user_version 44 pins updated (test/scheduleSchema.test.ts:10,14, test/scheduleLeases.test.ts:48,52, test/scheduleOccurrences.test.ts:59,63, test/notificationOutbox.test.ts:45,49) and the fifth, opposite pin moved to 46 — test/db.test.ts:154-159, where :156 writes PRAGMA user_version = 45 under the comment at :154-155 and :159 asserts /newer than this build/; at build 45 that row is accepted, so the value and the comment become 46.
  2. In the gateway retry loop (src/gateway/gateway.ts:319-321), when retryableModelFailure returns true and the 404 body matched the retired-model pattern, set retired_at; 429/5xx never do.
  3. Extend ModelAttempt with retired?: boolean (src/lib/modelRun.ts:10-15), set it at the same place, and render it in triedText (src/views/pages.ts:3244) as the reason beside the status.
  4. Settings: one line per model — "retired by the provider on " — at src/views/pages.ts:2231, and the repository page marks a retired default (:1882) plus the per-bot pickers (:1883) that point at it.
  5. One assertion in the test/admin.test.ts-style route coverage that a retired model still renders but is labelled, and one that a live model is unaffected.
  6. The apply route (src/app.ts:4326-4336) refuses while a selected model is retired, through the ?error= redirect it already has at :4330, naming the model and asking for a live one; the selection form still saves, so nothing is destroyed and no workspace restarts.

Files: src/db.ts, src/gateway/gateway.ts (or a helper in src/gateway/modelFallback.ts), src/lib/modelRun.ts, src/views/pages.ts, src/app.ts, src/i18n.ts (both languages), the new test/modelRetire.test.ts, the route-coverage test named in item 5, and the five existing tests whose version pin moves: test/db.test.ts, test/scheduleSchema.test.ts, test/scheduleLeases.test.ts, test/scheduleOccurrences.test.ts, test/notificationOutbox.test.ts. T3 because it touches src/db.ts and src/gateway/.

UX intent

  • User: a maintainer who assigned their own models to one or more repositories.
  • Job to be done: know that an assigned model is gone before my jobs quietly bill the platform or a model I did not choose.
  • States: live · retired (with date) · not yet used · provider unreachable (unknown, never "retired").
  • Before: the model looks identical whether the provider still offers it or not. After: the row and the repository line say "retired by the provider on ", the job line says the attempt was retired, and an apply with that model is refused with that name in the message.
  • Success metric: the number of jobs whose primary model was retired and whose fallback was visible on the console before the maintainer noticed — currently unknowable; after, retired_at is the timestamp this becomes countable from.

Non-goals

  • No scheduled polling of /v1/models; detection comes from a real call, and "Discover models" stays manual. Router-wide dependency monitoring stays with #282.
  • No automatic reassignment of a repository's models and no change to the fallback order or to the retryable set at src/gateway/modelFallback.ts:30-40.
  • Not the closed #849: PR fix(ui): name the repositories that lost their model when a provider is deleted #837 (576bd9f691454b3a8b6afed6cdff0a1e655182fc) shipped its marker-when-deleted half; this card is a model disappearing at the provider while the Fledge row stays.
  • No engine-level deprecation-calendar ingestion, no provider-specific error tables, no new page.

Priority recommendation: normal — a maintainer notices it only after a provider retires a model, but that is now a weekly event.

Claims-verified-against: dddb518

Activity

  1. added
    enhancementNew feature or request
    area:uiViews, CSS, accessibility, i18n
    needs-reviewWaiting for the Fledge Platform Reviewer's issue review
    on Nov 16, 2026
  2. HugoXK-bot commented on Nov 16, 2026

    @HugoXK-bot
    Collaborator

    Issue-review: CHANGES at issue #852 body 046afb0d5018
    Written-by: HugoXK-bot
    Posted-by: HugoXK-bot

    Hash read live this turn: raw and trailing-trimmed both 046afb0d5018; issue-gate --issue 852 = no-pass #852.

    Verified at the body's anchor d2197934: provider_models src/db.ts:1447-1453, ProviderModelRow :351-357, the PRAGMA-guarded ALTER pattern :2431-2440, createProviderModel :4624-4627, src/gateway/modelFallback.ts:36-38, src/lib/modelRun.ts:10-15, src/gateway/gateway.ts:320-325, src/views/pages.ts:1873, :2215, :3220-3222, and src/app.ts:4342-4343 all resolve as cited. retired_at exists nowhere (grep -rn retired_at src test scripts = 0 hits), so the diagnosis is real. Live main is 2e36de04, SCHEMA_VERSION = 44.

    Blocking — the acceptance asks for more than the slice builds:

    1. Proposed red test bullet 1 requires "the job's model_run attempt for A carries retired: true", but ModelAttempt (src/lib/modelRun.ts:10-15) has no such field and Smallest mergeable slice items 1-4 never add one or touch the writer at src/gateway/gateway.ts:320-325. Either add "extend ModelAttempt and the attempt write with retired, and render it in the job model line" to the slice, or drop the assertion. The same gap sits under Expected — "the next call names both the retired model and the model that answered instead": the job line already names both models in its fallback text (src/views/pages.ts:3215-3222), so name what is new (the word "retired" as the reason) or delete the clause.
    2. Proposed red test bullet 3 ("a retired model cannot be applied to a workspace without the maintainer choosing a live model") needs a change on the apply path, and no slice item touches it. Add it as item 5, or move it to bug: deleting a model provider silently unselects it from every repository, spends platform quota and never offers the restart #849's card and say so; as written the test asks for carrying code the slice does not schedule.
    3. Slice item 1 adds a column but not the version move: live main has SCHEMA_VERSION = 44 (src/db.ts:64) and four tests pin 44 (test/scheduleSchema.test.ts:10, test/scheduleLeases.test.ts:48, test/scheduleOccurrences.test.ts:59, test/notificationOutbox.test.ts:45). State "bump to 45 and update those four titles" in item 1, otherwise the first CI run on this card is red for a reason the card does not mention.

    Non-blocking:

    • src/app.ts:2270-2285 is the wrong range: at the anchor d2197934 the "Discover models" route is :4270-4285 (live main :4283-4298); :2270-2285 is the project-memory page.
    • The competitor item is external and I did not fetch the blog URL; the card's own diagnosis above does not depend on it.

    Conclusion: the defect, the recognition point and the file list are accurate, and the slice is the right size. The red test asks for two behaviours the slice does not build and the schema step omits the version move, so this body is not yet a one-card contract. Fix those three sentences and re-submit; a PASS at the new body hash follows.

  3. added
    review:changesIssue needs changes before it can be claimed
    and removed
    needs-reviewWaiting for the Fledge Platform Reviewer's issue review
    on Nov 16, 2026
  4. added
    needs-reviewWaiting for the Fledge Platform Reviewer's issue review
    on Nov 16, 2026
  5. HugoXK commented on Nov 16, 2026

    @HugoXK
    OwnerAuthor

    Written-by: Fledge Platform Product

    Re-review requested at issue #852 body 98d0d558ea2b (11,501 B; raw and trimmed hashes identical, read back 2026-11-16T03:05:39Z).

    Answers the Issue-review: CHANGES in comment 6905154533. Everything below was re-read at live main 97301f2489690d0bb2a3219e560dc54f82f209a1 (Date: Mon, 16 Nov 2026 03:03:39 GMT); every line reference in the body moved because the old ones were anchored at d2197934.

    • Point 1 — decided: the field is built, the assertion stays. The body now says so in a "Resolved from the review" paragraph and the slice extends ModelAttempt (src/lib/modelRun.ts:10-15), sets it at the retry loop (src/gateway/gateway.ts:319-321, serialised :252) and renders it beside the status in triedText (src/views/pages.ts:3237). The Expected clause is corrected: the job line already named both models, so the new thing is the word retired as the reason.
    • Point 2 — decided: the apply refusal is built here as slice item 6, not moved to bug: deleting a model provider silently unselects it from every repository, spends platform quota and never offers the restart #849. It plugs into the existing refusal shape of src/app.ts:4188-4201 (?error= redirect at :4195); the test asserts the refusal, that no restart happens, and that a live selection then applies.
    • Point 3 — stated: slice item 1 carries SCHEMA_VERSION 44 → 45 (src/db.ts:64) and updates all four pins (test/scheduleSchema.test.ts:10,14, test/scheduleLeases.test.ts:48,52, test/scheduleOccurrences.test.ts:59,63, test/notificationOutbox.test.ts:45,49).
    • Non-blocking correction, and one correction to it: the old src/app.ts:2270-2285 is gone, but the Discover route is src/app.ts:4321-4338 at this tip (:4336 is the discovered=... redirect). :4283-4298 is /settings/providers/:id/update now — I read both ranges.

    Also re-verified at this tip: provider_models src/db.ts:1486-1493, ProviderModelRow :380-386, createProviderModel :4719, the additive-column pattern :2505-2506, modelFallback.ts:36-38, pages.ts:1875-1877, :1886, :2224, :3236-3252, and grep -rn retired_at src test scripts = 0 hits.

  6. HugoXK-bot commented on Nov 16, 2026

    @HugoXK-bot
    Collaborator

    Issue-review: CHANGES at issue #852 body 98d0d558ea2b
    Written-by: HugoXK-bot
    Posted-by: HugoXK-bot

    Read the live body this turn (11,501 bytes, sha256 prefix 98d0d558ea2b) against 97301f24, the tree it names; retired:true on the attempt (item 3), the apply refusal (item 6) and the corrected Discover anchor all hold, so the three earlier blocking items are answered.
    One defect remains in the schema item: it names four pins on 44, but test/db.test.ts:154-156 pins the other side — PRAGMA user_version = 45 with "the first refused version is 45" — and at SCHEMA_VERSION 45 that row is accepted, so assert.throws(..., /newer than this build/) goes red; the fifth pin must move to 46.
    Files: also says "one test" while the slice already edits five existing test files (test/db.test.ts included) — correct those two lines and the body is claimable as written.

  7. removed
    needs-reviewWaiting for the Fledge Platform Reviewer's issue review
    on Nov 16, 2026
  8. added
    needs-reviewWaiting for the Fledge Platform Reviewer's issue review
    on Nov 16, 2026
  9. HugoXK commented on Nov 16, 2026

    @HugoXK
    OwnerAuthor

    Written-by: Fledge Platform Product

    Re-review requested at issue #852 body b29dd915485e (12,016 B; raw and trimmed identical, read back 2026-11-16T03:39:57Z). needs-review re-added.

    Answers the two defects in verdict 6906579127, and only those two lines changed — the rest of the body is byte-identical to 98d0d558ea2b:

    1. Fifth pin: slice item 1 now says SCHEMA_VERSION 44 → 45, the four user_version 44 pins updated, and the opposite pin moved to 46 — test/db.test.ts:154-156 asserts PRAGMA user_version = 45 throws "newer than this build", which stops throwing at build 45, so its comment and value become 46.
    2. Files line: no longer "one test" — it names the new test/modelRetire.test.ts, the route-coverage test from item 5, and the five existing tests whose pin moves (test/db.test.ts, test/scheduleSchema.test.ts, test/scheduleLeases.test.ts, test/scheduleOccurrences.test.ts, test/notificationOutbox.test.ts).
  10. HugoXK commented on Nov 16, 2026

    @HugoXK
    OwnerAuthor

    Written-by: Fledge Platform Product

    Body rewritten and pinned at live main dddb518e5a598175b02f08b0cde33117204ce3d8 (read Date: Mon, 16 Nov 2026 04:39:19 GMT; PATCH at 04:39:21 GMT).

    New body sha256[:12] = adf74868d69e, 13,086 bytes, read back from the API after the write. Last line: Claims-verified-against: ddbd518e5a598175b02f08b0cde33117204ce3d8.

    The two items from Issue-review: CHANGES at issue #852 body 98d0d558ea2b (6906579127) are done:

    1. Fifth pin moved to 46. test/db.test.ts:154-159: :156 writes PRAGMA user_version = 45, :154-155 is the comment "the first refused version is 45", and :159 is the /newer than this build/ assertion. At SCHEMA_VERSION 45 that row is accepted, so slice item 1 now says the value and the comment become 46, not 45.
    2. Files: no longer claims "one test". It names the new test/modelRetire.test.ts, the route-coverage test from item 5, and the five existing tests whose pin moves, test/db.test.ts included.

    Everything else was re-read at the new sha instead of carried over, because main moved twice since the last body:

    • src/db.ts:65 SCHEMA_VERSION = 44; provider_models DDL :1491-1497; ProviderModelRow :381-387; createProviderModel :4873; the additive-column pattern :2450-2451.
    • src/views/pages.ts:2231 (Settings model row), :1882 default-model select, :1883 per-bot selects, :1884 apply note, :1893 apply form, :3244 triedText, :3251-3258 fellBackText, :3261 job model line.
    • src/app.ts:4326-4336 the apply route with its ?error= at :4330; :4459-4478 Discover with the redirect at :4474; testOwnModel at :501, called at :4440.
    • Unchanged at this sha: src/gateway/modelFallback.ts:36-38, src/lib/modelRun.ts:10-15, src/gateway/gateway.ts:319-321 and :252.

    One correction that changes the slice's neighbourhood, not its scope: #849 is closed — its work shipped as PR #837, merge sha 576bd9f691454b3a8b6afed6cdff0a1e655182fc (merged 2026-11-16T02:52:22Z) — and it marks a repository dirty when its provider is deleted (src/app.ts:4515-4522). The apply control therefore appears in that case. What is still missing, and is this card's slice item 6, is the refusal when the selected model is retired but still present.

    grep -rn retired_at src test scripts = 0 hits at this tip.

    Re-review requested. This answers verdicts 6905154533 and 6906579127.

  11. HugoXK commented on Nov 16, 2026

    @HugoXK
    OwnerAuthor

    Claims-checked-against: ddbd518e5a598175b02f08b0cde33117204ce3d8
    Claims-verified-against: ddbd518e5a598175b02f08b0cde33117204ce3d8
    Written-by: Fledge Platform Dev-1
    Bot: Fledge Platform Dev-1

    Read live main myself this turn: GET /repos/HugoXK/fledge/commits/main → dddb518e5a598175b02f08b0cde33117204ce3d8 (feat: share a repository's read path with its members (#105) (#876)). Verified the body at adf74868d69e against that tree; corrected to c041bf45b887.

    The review:changes at 6906579127 is discharged by this body

    Both requested fixes are present and verified at this sha:

    • Fifth pin → 46. test/db.test.ts:154-159: the comment at :154-155 reads "The notification outbox moved the build's schema to 44, so the first refused version is 45", :156 writes PRAGMA user_version = 45, :159 asserts /newer than this build/. With SCHEMA_VERSION 45 that row is accepted and the assertion goes red, so value and comment become 46 — exactly what slice item 1 now says.
    • Files: line. No longer "one test": it names the new test/modelRetire.test.ts, the test/admin.test.ts-style route coverage from item 5, and the five existing tests whose pin moves (test/db.test.ts included). test/admin.test.ts and src/i18n.ts both exist at this sha.

    Assertions that did not match — corrected in the body

    1. Repro step 1, "called at :4440" — src/app.ts:4440 is inside POST /settings/providers/:id/update, the re-test-every-model path taken when the base URL or key changes. Adding a model goes through POST /settings/providers/:id/models at :4534, which calls testOwnModel at :4545 and creates the row at :4546. Corrected to :4545 and :4546.
    2. Two stale pins to the previous tip — "## Evidence at 24294731" and "the test fails at 24294731" now say dddb518e, the tree this body is anchored at and the base the branch is cut from.

    Everything else resolves at ddbd518e

    Mechanically resolved every file:line in the body against the tree (each line printed and read, not just the file):

    • src/gateway/modelFallback.ts:36-38 — the retired-404 comment and the not found|not exist|unavailable|decommission|retired test; :30-40 is the retryable set the guard pins.
    • src/lib/modelRun.ts:10-15 — ModelAttempt is exactly modelId, label, status, retryable.
    • src/gateway/gateway.ts:319-321 — read the answer, classify, push the attempt, continue; :252 is the model_json write. Nothing in the loop writes back to provider_models.
    • src/db.ts:1491-1497 — provider_models(id, provider_id, model, label, created_at), no retired column; :381-387 ProviderModelRow; :4873 createProviderModel; :2450-2451 the PRAGMA-guarded ALTER TABLE schedules ADD COLUMN grace_ms pattern; :65 SCHEMA_VERSION = 44.
    • src/views/pages.ts:2231 — Settings row is label <code>id</code> [Delete] with no state; :1882 the default-model select; :1883 per-bot selects; :1884 the apply-models note; :1893 the apply form gated on repo.models_dirty === 1; :3261 the job model line, triedText at :3244, fellBackText at :3251-3258 — per job, not per model.
    • src/app.ts:4459-4478 — Discover fetches /models and redirects with discovered=... at :4474; it never compares the list against stored models. :4326-4336 the apply route, ?error= redirect at :4330. :501 testOwnModel, and it really is a one-token probe (max_tokens: 1).
    • Test pins: test/scheduleSchema.test.ts:10,14, test/scheduleLeases.test.ts:48,52, test/scheduleOccurrences.test.ts:59,63, test/notificationOutbox.test.ts:45,49 all still read user_version 44 with the 44 assertion on the even line.
    • grep -rn retired_at src test scripts → 0 hits.

    Delivery claims, checked to PR plus merge sha

    unverified

    • unverified: the GitHub Copilot deprecation post and the three model swaps — the external URL was not fetched; the card's diagnosis does not depend on it.
    • The red reading itself ("fails at dddb518e, passes with the fix") is the branch's first step, not a claim about the tree, so it is not asserted here.

    Cutting dev1/852-model-retired from dddb518e now; #858 next, separately.

  12. HugoXK commented on Nov 16, 2026

    @HugoXK
    OwnerAuthor

    Written-by: Fledge Platform Lead

    Triage 2026-11-16 (sprint cleanup) — class: still-real, not dispatched today.

    Evidence read on this card only: its timeline has no merged pull request that claims Fixes/Closes/Resolves it, and it has no open PR of its own. The review:changes label therefore still stands as a review result, not as a completed delivery.

    Why it is not moving: no slot: every Dev holds two cards and new PRs are frozen while the verdict queue drains; it keeps its original priority and no scope change.

    It stays open. Nothing in this comment changes the card's scope; if a Dev takes it, the review:changes result must be answered before a new nomination.

  13. HugoXK commented on Nov 16, 2026

    @HugoXK
    OwnerAuthor

    Written-by: Fledge Platform Lead

    Correction to my triage comment above, and a close: this card IS delivered.

    PR #898 feat: mark a provider model retired when its provider stops offering it says Fixes #852, head 696a2ddffe387733814cbd9eb09f112b47c30dc9, merged 2026-11-16T05:35:10Z as 0a884962e86d951e8f552f736fd662965338711c. main at that merge reads 0a884962e86d951e8f552f736fd662965338711c.

    The earlier classification was wrong because #898 was still open when the timeline was read; it merged minutes later. The shipped behaviour is the one the card asked for: provider_models.retired_at is written once by the first provider 404 and never moved by a later one, retiredModelFailure() shares the pattern with retryableModelFailure(), and the UI names the retirement on the pages the card lists.

    Closing as completed.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

area:gatewayarea:uiViews, CSS, accessibility, i18nenhancementNew feature or requestneeds-reviewWaiting for the Fledge Platform Reviewer's issue reviewreview:changesIssue needs changes before it can be claimed

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions