Repository navigation
fix: make redirect_uris the source of truth for OAuth2 redirect URIs - #29440
Conversation
Editing an app's callback added a new accepted URI without removing the old one, and the admin API only showed the callback (PLAT-619, SEC-870). - Backfill every row so the callback is the first redirect URI. The migration handles empty, missing, and out-of-order lists and is safe to rerun. The callback column is kept, marked deprecated. - Read redirect URIs from the list only. A callback that is not in the list is no longer accepted. - Write both columns together from one shared function. A callback-only update now replaces the previous primary and keeps the rest. - Validate every entry the same way as the callback. - Cap the list at 32 entries of 2048 bytes each, including at registration. No existing request or response changes shape.
|
/coder-agents-review |
|
Chat: Review in progress (17/17 reviewers complete) | View chat deep-review v0.9.0 | Round 4 | Last posted: Round 4, 28 findings (1 P2, 11 P3, 9 Nit, 7 Note), COMMENT. Review Finding inventoryFinding inventory - PR #29440Findings
Contested and acknowledgedCRF-5 (P3, modelmethods.go:666) - Rolling-upgrade read/write skew drops an admin callback edit
CRF-23 (Note, registration.go:255) - RFC 7592 responses echo raw app.RedirectUris
Round logRound 1Netero-only first pass: 3 Notes, mechanical floor clean (P3-and-below gate). Panel of 17 (kurapika, razor, knov, knuckle, ryosuke, ging-go, chopper, komugi, mafuuu, bisky, hisoka, mafu-san, pariston, gon, leorio + wildcards killua, luffy). Convergence: 6 reviewers on CRF-4 (admin lockout), 2 at P2. kurapika, ging-go, komugi, killua: no findings. 1 P2, 4 P3, 2 Notes new, 5 Nits. CRF-1 superseded, CRF-3 folded. Reviewed against 6967259..96bbec0. Round 2Churn guard PROCEED: 12 addressed, 1 acknowledged (CRF-5, documented), 0 silent. Author pushed 718b1f9, 58d2036, f9bed52, 69bee1c, a8067c9, f59171f, da9563c; branch rebased (base d671150). Panel re-verifies the fixes against d671150..4c16899. Round 2 (post-panel)Netero: no findings; all 12 fixes verified mechanically. Panel of 16 (kurapika, razor, knov, ryosuke, meruem, mafuuu, hisoka, bisky, komugi, mafu-san, pariston, gon, leorio, chopper + wildcards takumi, luffy). All 12 fixes confirmed at root. New: CRF-16 P3 (3-way convergence Meruem/Mafuuu/Luffy: CRF-4 fix missed the RFC 7592 writer, verified in code), CRF-17 P3 (false doc invariant + admin/DCR validation divergence), CRF-20 P3 (CRF-5 doc understates consequence), plus 3 Nits and 2 Notes. CRF-5 held open at P3: whole panel agrees documented mitigation is the correct engineering response (no in-PR code fix; the read-flip and self-heal invariants conflict) but a silent non-self-healing drop cannot be agent-accepted; needs explicit human accept or a linked PLAT-619 ticket. CRF-6 closure (tests+doc for build-ahead) accepted as documented stacked design. Reviewed against d671150..4c16899. Event COMMENT (no P0/P1). Round 3BLOCKED by churn guard. Author addressed 5 (CRF-17 6f7112a, CRF-19/CRF-21 a4c5940, CRF-20 05348bf, CRF-22 c995ef8) and re-resolved CRF-5 without a new response (acknowledged, PROCEED-side). Three findings silent: CRF-16 (P3, verified registration.go unchanged), CRF-18 (Nit), CRF-23 (Note). These three were auto-folded into the round-2 review body (their lines were off the diff), which likely explains why they were missed while every inline thread was resolved. No panel, no Netero. Block review posted naming the silent items and the CRF-5 human-decision request. Branch rebased (base c29116d), squashed to single commit d6e9e78. Round 4Churn guard PROCEED: 2 addressed (CRF-16 fix 1e38088, CRF-18), 2 acknowledged (CRF-23 leave-as-is, CRF-5 explicit human accept), 0 silent. Block cleared. Panel re-verifies the CRF-16 fix (new stored-aware DCR update validator) plus the previously-unreviewed round-3 fixes (CRF-17 unification 6f7112a, CRF-22 struct-tag removal c995ef8), since round 3's panel was blocked. Reviewed against c29116d..255b2f5. Round 4 (post-panel)Netero: no findings. Panel of 15 (kurapika, razor, knov, meruem, ryosuke, chopper, bisky, hisoka, komugi, mafuuu, mafu-san, pariston, gon, leorio + wildcard killua). All four target fixes (CRF-16/17/18/22) verified at root, no cap bypass (stored is server-derived). But the fixes introduced/exposed new issues: CRF-24 P3 (admin confidential http-transport divergence, PR's "same rules" claim overstated), CRF-25 P3 (CRF-16 raw-vs-dedup count re-locks legacy over-cap DCR apps with duplicate URIs, verified in code), CRF-26 P3 (false doc invariant reintroduced by the stored-skip), CRF-29 P3 (validateScheme messages stutter after unification), plus CRF-28/30 Notes and CRF-27 Nit. Event COMMENT (no P0/P1). Reviewed against c29116d..255b2f5. About deep-reviewCRF = Coder Review Finding (P0-P4, Nit, Note)
|
There was a problem hiding this comment.
This is a clean, well-tested change. Collapsing every consumer onto RegisteredRedirectURIs() gives one source of truth with a primary-first invariant, the migration is idempotent and lossless (WHERE redirect_uris[1] IS DISTINCT FROM callback_url with array_prepend/array_remove, verified by a re-run assertion), and the tests read the database directly instead of trusting response echoes. Bisky put it well: "this suite is lovely."
One P2 stands out. Six reviewers independently converged on CRF-4: UpdateApp re-validates the app's carried-over stored redirect URIs against the two caps this PR introduces, so a legacy DCR app whose list predates the caps becomes uneditable through the admin API, reintroducing the exact PLAT-619 lockout this PR exists to remove. Two reviewers rate it P2. The fix is to validate only the entries the request actually supplies and let already-stored entries pass through until the list-editing API lands.
Severity count: 1 P2, 4 P3, 2 Notes, 5 Nits. The P3s cluster around real seams: a silent lost callback edit during a rolling upgrade (CRF-5), resolveRedirectURIs shipping untested branches ahead of its caller (CRF-6), and error messages that name a field and index the admin never supplied (CRF-7). CRF-4 and CRF-5 both hinge on deployment or data states that no reviewer can bound from the diff, and both need a human decision rather than silent acceptance.
Netero's first-pass CRF-1 (field mislabel) is superseded by CRF-7: what read as cosmetic becomes the operator-visible symptom once the caps make the input reachable.
🤖 This review was automatically generated with Coder Agents.
… app update - Admin updates re-checked the whole stored list against the new caps. - Apps registered before the caps with over 32 URIs, or a URI over 2048 bytes, could not be edited at all. - Only URIs the request adds are checked now. The count cap applies only when the list grows.
… redirect URIs - An old replica writes only callback_url on an admin update. - New replicas read redirect_uris first, so that edit is not seen until the app is saved again. - Drain old replicas before upgrading to avoid it.
…de window - redirect_uris is now the source of truth, with the primary first. - During a rolling upgrade, a callback edit on an old replica is not seen by new replicas until the app is saved again. - An app may register at most 32 redirect URIs of at most 2048 bytes each. Older apps beyond that keep working.
Docs previewCheck off each page once it's been reviewed. If a page changes in a later push, its checkbox clears automatically so it gets a fresh look. Pages not yet wired into the docs navigation aren't listed here. |
- Covers all four documented input shapes, plus dedup and callback ordering. - The doc comment notes that the list-with-callback and neither shapes get their caller when the admin API gains redirect_uris in the next PR.
…he callback field - A failure on the request callback_url now says "callback URL ..." with no list index. - List entries keep the indexed wording for registration callers.
…lidation - The shape check now rejects fragments, so admin and registration store the same rules. - Rename ValidateOAuth2CallbackURL to ValidateOAuth2RedirectURIShape and the local validator to validateAppRedirectURIFields. - The column comment says "first entry of redirect_uris" instead of a 1-based index. - RegisteredRedirectURIs documents that it is never empty. - Test that registration stores a deduplicated list. Use plain numbered test URIs.
|
/coder-agents-review |
There was a problem hiding this comment.
Round 2. The author addressed all 15 round-1 findings and the panel verified each fix at the source, not just the response. CRF-4 (the P2 admin lockout) is fixed at root: validateAppRedirectURIFields skips stored URIs and gates the count cap on growth, so a legacy over-cap app stays editable, covered by a new 33-URI test. The renames, the fragment unification, the min-length invariant doc, the error-message rework, and the generated-comment fix all landed. Hisoka put it well: "Boring code gets silence. This wasn't boring, and it still walked away."
One fix was incomplete. Three reviewers (Meruem, Mafuuu, Luffy) independently found that the CRF-4 relaxation was applied only to the admin writer; the RFC 7592 / DCR self-service update path still runs the new caps over the client's full resent list with no stored exemption, so a legacy over-cap DCR app cannot update its own registration without first dropping URIs. Verified: UpdateClientConfiguration gates on req.Validate() -> ValidateRedirectURIs, which has no stored context. Same lockout class, on the writer the fix missed (CRF-16). The rule that one correction implies every instance applies here: thread stored URIs through both writers.
CRF-5 (rolling-upgrade skew) stays open at P3. The whole panel agrees the documented mitigation is the correct engineering response, because the read-flip that fixes PLAT-619 and a self-heal from callback_url are mutually exclusive, so no in-PR code change closes it. But a silent, non-self-healing drop whose failure mode reopens the exact stale-redirect-URI exposure PLAT-619 exists to end is not something a reviewer or the author-agent can accept as permanent. It needs an explicit human accept or a linked PLAT-619 tracking item, not documentation alone. Relatedly, CRF-20: the mitigation doc and migration comment read like a display lag and undersell that the stale URI stays a live redirect target during the window.
New this round: 3 P3 (CRF-16, CRF-17, CRF-20), plus CRF-5 held open at P3, 3 Nits, 2 Notes. CRF-6's build-ahead branches were closed with a table test and a disclosing comment rather than removal; the panel accepts that as a documented stacked-PR design.
coderd/oauth2provider/registration.go:306
P3 [CRF-16] The CRF-4 fix was applied only to the admin writer; the RFC 7592 / DCR update path still applies the new 32-count / 2048-byte caps to a legacy over-cap client's full resent list, so that client is locked out of its own config updates. (Meruem P3, Mafuuu P3, Luffy P3)
UpdateClientConfigurationgates onreq.Validate(), which callscodersdk.ValidateRedirectURIs... Unlike the admin path, this path passesnilforstoredand has no stored-skip: every entry the client resends is re-judged against the new caps.
Three reviewers converged. Verified: req.Validate() -> ValidateRedirectURIs enforces the hard caps with no stored context, and RFC 7592 update is a full replace, so a pre-cap DCR app with >32 URIs or a >2048-byte URI cannot edit any field (name, logo, scope) without first trimming its list, while GET still hands it that same list. Narrower than the admin case because the client owns its list, but it is the same class the fix set out to remove. Apply the same stored-aware relaxation on the update path (scope to update; create should still enforce the caps), or route both writers through one (candidate, stored) validator.
🤖
codersdk/oauth2_validation.go:87
Nit [CRF-18] The ValidateRedirectURIScheme doc comment is duplicated verbatim (the same 7 lines appear twice). (Knov)
Pre-existing at the base SHA, not introduced here, but this PR rewrites the surrounding file and it is a central file to the change, so it is cheap to fix in the same pass. Delete the second copy.
🤖
coderd/oauth2provider/registration.go:255
Note [CRF-23] The RFC 7592 responses echo the raw app.RedirectUris (also lines 192, 412) while every enforcement and admin path reads through the single owner RegisteredRedirectURIs(). (Kurapika, Razor)
Benign today because a DCR app always has a non-empty, deduplicated list, so the raw read equals the owner's output. Worth knowing: it is a second read of the same column that will diverge from the owner if the empty-list fallback or the dedup semantics ever change. Not worth changing now.
🤖
🤖 This review was automatically generated with Coder Agents.
- ValidateRedirectURI now rejects hostless http(s) and opaque URIs for every client type. - Move the shared rules into a helper so both paths use one copy. - Test that ValidateRedirectURI rejects everything the shape check rejects.
…thub.com/coder/coder into plat-619-1-redirect-uris-source-of-truth
…upgrade - The doc bullet and migration comment said the edit was "not visible". - They now say it is silently discarded and the app keeps accepting the previous redirect URIs until saved again.
- Drop the forward reference to the next PR from the resolveRedirectURIs doc. - Let the public client rejection message stand on its own instead of repeating "public client".
- Remove the oauth2_callback_url struct tag. It ran the shape check first and failed with a generic tag message. - The handler already runs the same check and names the reason against callback_url. - Test that a callback with a fragment reports that reason.
|
/coder-agents-review |
There was a problem hiding this comment.
Round 3 is blocked before the panel. The author addressed five of the round-2 findings, verified by the churn guard: CRF-17 (DCR shape validation unified in 6f7112a), CRF-19 and CRF-21 (a4c5940), CRF-20 (mitigation doc reworded in 05348bf), and CRF-22 (struct tag removed so the callback reason is reported, c995ef8). Good, targeted fixes.
But three findings from the round-2 review are still open with no reply and no code change, so the panel will not re-review this round. All three were auto-folded into the round-2 review body (their lines were not on that round's diff), which is the likely reason they were missed while every inline thread was resolved. They are in the round-2 review body above; please scroll to the end of it.
-
CRF-16 (P3) coderd/oauth2provider/registration.go:306. The CRF-4 fix (validate only what the request introduces, skip stored entries, cap only on growth) was applied to the admin writer only. The RFC 7592 / DCR update path still runs the caps over the client's full resent list with no stored context, so a legacy over-cap DCR app is locked out of its own config updates. Verified again this round: registration.go is unchanged and UpdateClientConfiguration still gates on req.Validate() -> ValidateRedirectURIs with no stored argument. Same lockout class as CRF-4, on the writer the fix missed. Needed: apply the same stored-aware relaxation to the RFC 7592 update path (scope to update; create should still enforce the caps), or route both writers through one (candidate, stored) validator. If you intend to accept this gap instead, say so and why.
-
CRF-18 (Nit) codersdk/oauth2_validation.go:87. The ValidateRedirectURIScheme doc comment is still duplicated verbatim (the same seven lines appear twice). Pre-existing, but this PR rewrites the file. Delete the second copy, or reply that it is out of scope.
-
CRF-23 (Note) coderd/oauth2provider/registration.go:255. The RFC 7592 responses still echo raw app.RedirectUris (also lines 192, 412) rather than the single read owner RegisteredRedirectURIs(). Benign today; a one-line reply acknowledging it is enough to clear it.
Separately, CRF-5 (P3, rolling-upgrade skew) was re-resolved without a response to the reopened thread. As stated there, the documented mitigation is the right engineering answer and no in-PR code change can close the window, but accepting a silent, non-self-healing drop as permanent is a human decision. It needs an explicit human accept-the-risk or a linked PLAT-619 tracking item, not silent re-resolution.
Once you push fixes or respond to CRF-16, CRF-18, CRF-23, and CRF-5, the full panel will re-review.
🤖 This review was automatically generated with Coder Agents.
…pdate itself - The RFC 7592 update path validated the full resent list with no stored context, so a legacy over-cap client could not save any change. - It now skips redirect URIs the client already has and applies the count cap only when the list grew, matching the admin path. - Create still enforces the caps on every entry. - Remove the duplicated ValidateRedirectURIScheme doc comment.
…thub.com/coder/coder into plat-619-1-redirect-uris-source-of-truth
|
Responses to the round-2 review-body items: CRF-16: fixed in 1e38088. The RFC 7592 update handler now validates against the redirect URIs the client already has, skipping stored entries and applying the count cap only when the list grew. This matches the admin path. Create is unchanged and still enforces the caps on every entry. Added tests at the codersdk level and at the handler level with a seeded over-cap client. CRF-18: removed the duplicated doc comment in the same commit. CRF-23: acknowledged. The RFC 7592 responses return the stored list directly. Every dynamically registered client has a non-empty list, so this matches what |
|
/coder-agents-review |
There was a problem hiding this comment.
Round 4 clears the block. The author responded to all four outstanding round-2/3 items: CRF-16 fixed in code (RFC 7592 update now uses a stored-aware validator), CRF-18 removed, CRF-23 acknowledged, and CRF-5 given an explicit human accept-the-risk. Both accepts are recorded and closed. The panel verified the four target fixes (CRF-16, CRF-17, CRF-22, CRF-18) at the source: the CRF-16 stored-skip opens no cap bypass because stored is server-derived from the existing row, create still enforces the caps on every entry, and the two writers share the grow-only rule. Bisky: "this one actually sparkles."
But round 3's panel was blocked, so this is the first panel pass over the round-3/4 validation rework, and that rework introduced or exposed four P3s, none blocking (no P0/P1):
- CRF-25: the CRF-16 fix compares the client's raw resent list against a deduplicated
storedbaseline. A legacy DCR app with duplicate redirect URIs (pre-cap DCR stored them un-deduped, and the migration only removes callback duplicates) gets its unchanged, GET-echoed list rejected on PUT because raw count > unique count. Hisoka traced it and I confirmed the arithmetic in code: it re-locks the exact PLAT-619 class the fix set out to remove, for that subset. The admin path does not have this hole because it compares deduped against deduped. - CRF-24: the admin confidential path still accepts
http://to non-loopback hosts that the DCR path rejects, so the PR's "admin and registration enforce the same rules" claim is overstated. Razor rates it P3 (plaintext auth-code transport), Knov and Ryosuke read it as intentional admin-trust looseness. It is admin-gated and pre-existing, so it needs a human call: enforceValidateRedirectURIfor confidential admin entries, or document the deliberate split with its rationale. - CRF-26 and CRF-29 are the CRF-17/CRF-22 unification's loose ends: a doc invariant that the CRF-16 stored-skip made false again, and scheme error messages that now stutter ("callback URL redirect URI must have a scheme").
Severity count: 4 P3, 2 Notes, 1 Nit, all new. The code is close; these are refinements to the validation rework, not the core design, which the panel continues to find sound.
🤖 This review was automatically generated with Coder Agents.
- Legacy clients may store duplicates, and the stored baseline is deduplicated. - Resending the list as returned no longer reads as growth.
…rowth" This reverts commit 7ba41c6.
…I caps update itself" This reverts commit 1e38088.
… URIs on app update" This reverts commit 718b1f9.
- Stored redirect URIs are checked again on every update, with no exception for apps that predate the caps. - An app over the limits must be deleted and created or registered again. - Remove the duplicated ValidateRedirectURIScheme doc comment again.
|
Response to CRF-16. Decision: the stored-skip logic is removed. Commits 6ff2f47, bb15f69 and 6e118d8 revert it, and a2918ae documents the recovery path. The caps stay. Reasoning:
The docs Limitations section and the validator comments now say this. |
…hority - Registration accepted the bare scheme:value form before this change. - Add test cases for the three private-use forms.
- Rename ValidateOAuth2RedirectURIShape to ValidateRedirectURIShape to match its siblings. - Scheme errors no longer start with "redirect URI", since every caller adds its own subject.
Renumber the redirect URIs migration from 000596 to 000598, since main took 000596 and 000597 in the meantime.
dylanhuff-at-coder
left a comment
There was a problem hiding this comment.
I left a few comments for docs and regression tests. There is now a duplicate 000598 migration which needs to be changed before merging, but otherwise lgtm.
The PR description still says stored URIs are skipped during validation and lists the removed 33-URI rename-success test. That no longer matches the implementation. Please update both to reflect that the full resolved list is validated, including on a rename.
Main merged 000598_mcp_server_signing_secret, so the redirect_uris migration collided on the same version number. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Saving the app again on the new version does not restore the edit, since the form shows the old callback from redirect_uris. The admin must enter the intended value again once every replica runs the new version.
…ect URIs Replace the later-release promise with the current behavior: editing the callback replaces the primary and preserves the other registered URIs, and only the primary is displayed.
…t URIs Cover an update that keeps the callback but retains a stored list of 33 URIs or a 2049-byte alternate. Both must fail with a redirect_uris error and leave the row unchanged.
TL;DR
First of three in PLAT-619. Since #29014 the authorize and token endpoints accept every URI in
callback_urlandredirect_uris. Editing an app's callback through the admin API added a URI and removed nothing, and the API showed only the callback. An admin could not move an app off a stale redirect URI. This PR makesredirect_uristhe only source of truth, with the first entry as the primary, and keepscallback_urlas a deprecated copy of that entry until a later PR drops it.redirect_urison the admin API. Markcallback_urldeprecated.Implementation Details
The fix
000596rewrites every row soredirect_urisstarts withcallback_url. It handles NULL lists, empty lists, and lists where the callback is missing or not first, in one statement that is safe to rerun. The table holds one row per app.RegisteredRedirectURIson the app model is the one read path. The authorize and token endpoints, the corrupt-callback log line, and the API response all read through it. Acallback_urlthat is not in the list is no longer accepted; only a binary older than the migration could write such a row.callback_urlto its first entry in the same statement.callback_urlnow replaces the previous primary and keeps the other registered URIs. Before, it added the new callback and kept the old one.Validation
oauth2_callback_urltag, and the registration rules for public apps. The shape rule moved out of the struct tag intoValidateOAuth2RedirectURIShapeso it can run per entry. It now also rejects fragments, so admin and registration store the same rules.ValidateRedirectURIs, so registration is bounded too. It previously accepted anything under the 4 MiB request body limit.callback_urlis reported against that field as "callback URL ...", with no list index. List entries keep the indexed wording for registration callers.Unchanged
redirect_urison the admin API is the next PR.NOT NULLand a cardinality check onredirect_uriswait for the column drop, so an old replica cannot fail an insert during a rolling upgrade.callback_url. New replicas read the list first and do not see it. Saving the app again on new code fixes the row. Drain old replicas before the rollout to avoid it. The migration comment and the OAuth2 provider doc say the same.Docs
docs/admin/integrations/oauth2-provider.mdLimitations: the list is the source of truth, the upgrade window, and the 32 entry and 2048 byte caps.Tests
TestMigration000596OAuth2RedirectURIsPrimaryseeds each row shape, migrates, and checks the callback is first and nothing was lost. It runs the statement again and checks nothing changed.TestOAuth2ProviderAppRedirectURIs: a callback-only update moves the primary and keeps the rest; choosing an existing alternate as the callback makes it the only entry; a stalecallback_urlis ignored on read; create stores the callback as the list; an oversized callback is refused on create and update; an app with 33 stored URIs, one oversized, can still be renamed and keeps its list; registration stores a deduplicated list.TestResolveRedirectURIs: a table over the four documented input shapes, plus dedup and callback ordering.TestOAuth2AuthorizeFollowsCallbackEdit: after a callback edit the old primary is refused on Coder, the other URI still works, an omittedredirect_urigoes to the new primary, and a code issued to the old URI fails exchange withinvalid_request.TestOAuth2AuthorizeIgnoresCallbackColumn: a seeded row with a stale column is refused for that URI and accepted for the listed one.PublicDCRCallbackURLPolicyandBasicOperationsnow also assert the stored list after each update.TestRegisteredRedirectURIs: the column is ignored, duplicates are dropped.TestValidateOAuth2RedirectURIShapeandTestValidateRedirectURIsSizeincodersdk: accepted and rejected shapes including a fragment, 32 accepted and 33 refused, 2048 accepted and 2049 refused, an oversized value is refused before parsing.Manual Tests
Scenario summary, all 33
redirect_urislist, not NULL000596: every row now starts with its callback, and nothing was lost, across all five row shapesUPDATE 0, rows unchangedcallback_url(deprecated) andredirect_uriscallback_urlthrough the list's first entryPUTnow replaces the primary and keeps the rest of the list, instead of adding to itinvalid_request)callback_urlcallback_urlcolumn (only possible pre-migration) is ignored by every read path; a normalPUTrepairs ithttps, and bare private-use scheme are each refusedcallback_urland enforcement both followNo PR bugs found. One bug found and fixed in the test runbook itself
(scenario 18's helper originally posted to the wrong token endpoint path);
see the note under section 6.
Shell helpers used throughout
Run against the dev server through a
coder port-forward ... --tcp 3000:3000tunnel; the session token and embedded Postgres both live inside the
workspace, so they're read over
sshrather than exported once.1. Register apps on the old binary (parent commit)
2. Reproduce the PLAT-619 bug
The admin's edit added
MOVEDand removed nothing; there was no way throughthe admin API to revoke
CB.3-4. Snapshot, upgrade, and the migration's output
Snapshotted all five rows and
schema_migrations = 595, then checked outthe PR branch and restarted the dev server on the same database (no
--db-reset).5. The fix: a callback-only `PUT` actually replaces the primary
Starting from D1's three-entry post-migration list, the admin walked it
down to one correct entry using nothing but repeated
callback_urledits-- the PLAT-619 headline, achievable here in steps; the next PR's list
field makes it one request.
6. Backward compatibility and the stale column
7. Validation and caps
8. The RFC 7592 client update writes both columns
Run before cleanup, since it needs the untouched DCR app from section 1.
9. Cleanup
Stack: this PR, then #29502 (admin API), then #29603 (UI).