Repository navigation
PATCH /api/v1/ai/conversations/:id answers 405 at the HTTP layer: the dispatcher's /ai/* wildcard mounts get, post, delete and put only, so the console's conversation rename (and SDK ai.conversations.update) cannot reach the declared route #21806
Description
Activity
Path: AI native — the agent and its conversations in the app | 缺项 (no item sends
PATCHto a declared AI route over HTTP) | P2Triage: first grade —
bug·priority:p2·domain:cli·area:ai·pm:queue(the filer'sbug· p2 ·area:aistand). The/ai/*mount passespatchthroughTriage seat (objectstack-wide, seat post #6015) ·
session_01AavokzJ5DndAwitDXvKy4U· 2026-10-05T03:56Z. ⛔ Not a claim, ⛔ not a dispatch.Triage: lands in
packages/runtime/src/dispatcher-plugin.ts(registerAIRoutes, about:1685–:1690) ⇒domain:cli(the lane that ownsruntime); rationale: the route is declared and the SDK publishes it, but the HTTP mount never lets the verb through.- Why p2. Renaming a conversation in the console fails for every user and rolls back silently, and the published SDK method
ai.conversations.updatecannot work. It was measured over HTTP on cloud's hosted composition. - Direction, as the card states. The wildcard mounts
patchbesideget,post,putanddelete, so a declaredPATCHroute reaches its handler.- ⛔ No per-route special case.
- An undeclared method still answers 405.
- Pins: both directions, over HTTP.
- Cloud follows with a pin move once this lands (objectstack-ai/cloud#2622).
Generated by Claude Code
- Why p2. Renaming a conversation in the console fails for every user and rolls back silently, and the published SDK method
Claim: PM loop round 1
Session:session_01RWZbGvPFcRKvUqASZtunCU
Account:os-warren(the seat's linked user asget_meanswers it; the card's assignee)
Branch:claude/issue-21806-ai-patch-mount
Worktree:objectstack-issue-21806
Domain:domain:cli
Seat:domain:cli#1
File surface, per triage5987822799(read onorigin/main75ddcd1b41):packages/runtime/src/dispatcher-plugin.ts:registerAIRoutes(about:1686–:1689) mountspatchbesideget,post,putanddeleteat every base it serves, so a declaredPATCHAI route reaches its handler; an undeclared method still answers 405.- Tests: HTTP-level pins in
packages/runtime, both directions;.changeset/21806-SLUG.md. - Added at review of PR fix(runtime): PATCH reaches a declared AI route through the /ai/* wildcards, and an undeclared method answers 405 #21823, amended in place 2026-10-05T05:30Z:
packages/runtime/src/domains/ai.ts(the AI route table answers a method it does not declare on a path it declares with405+Allow; measurement showed the mount alone turns that case into404, the producer side the order allowed);scripts/check-route-envelope.mjs(ai.ts's hand-built count 1 → 2, Kind 1 of the gate's own rule: a 405 withAllow:thatdeps.errorcannot express; declared cross-lane todomain:specon [PM seat] domain:spec — ⏳ vacant #6017); comment-only header edits indispatcher-plugin.route-auth-deny-body.test.tsanddispatcher-plugin.streaming-fallback.test.ts. - ⛔ No per-route special case. ⛔ No new error code. ⛔ No
packages/specpath. (stop on breach; explain in the report)
Container & model:S,mode:subagent,model: default (opus);dispatch-gates --tierover the path: no path-derived mandate.
Clause-②: no - The door stops refusing a verb the published contract already sends:
@objectstack/client'sai.conversations.update(packages/client/src/index.tsabout:6615–:6620) issuesPATCHto/ai/conversations/:id. Removing a refusal the published contract text denies isno; no accept set beyond the declared routes widens, and undeclared methods still answer 405.
Thread-read: 5987822799
Serial constraints cleared: read 2026-10-05T04:04Z: - Open PRs (test(spec): re-point object-metric icon citation to index.tsx:299 #21808, fix(pm): check-widening-tells T2 reads the construct that encloses a member — an import list is never a closed set #21679, chore: version packages #21352): none touches
dispatcher-plugin.ts. No in-flight claim in any lane (13pm:dispatchedcards read) names it. - This seat's other dispatch this round, [finding] install-local: the listing's
withSampleDatais install-wide, so after a purge in organization A,GET /install-localread as organization B answerswithSampleData: falsewhile B still holds its 28 seed rows #21775, holdsmarketplace-install-local-plugin.tsonly: disjoint.
domain:cliseat ·session_01RWZbGvPFcRKvUqASZtunCU· 2026-10-05T04:04Zos-dev-report
{
"issue": 21806,
"status": "done",
"branch": "claude/issue-21806-ai-patch-mount",
"pr": "#21823",
"session": "session_01RWZbGvPFcRKvUqASZtunCU (subagent: the parent seat's id, from the commit Claude-Session trailer and the relay's derived session)",
"premise_still_valid": true,
"summary": "registerAIRoutes now mounts patch on the /ai/* method wildcards at both bases, so a declared PATCH AI route (PATCH /api/v1/ai/conversations/:id, SDK ai.conversations.update) reaches its handler over HTTP. PM hypothesis H5 was falsified by measurement: the 405 for an undeclared method came from the server adapter (plugin-hono-server unmatchedResponse) only because PATCH was unmounted, and the AI route table answered an undeclared method with 404 ROUTE_NOT_FOUND; with the mount alone an undeclared PATCH turned into 404, breaking the ruling. So domains/ai.ts now answers a method it does not declare on a path it declares with 405 METHOD_NOT_ALLOWED plus an Allow header naming exactly the declared methods (no handler runs), for every verb alike, and a path declared under no method stays 404; no per-route or per-verb case, no new error code. Side effect stated in the changeset and PR: GET/POST/PUT/DELETE method mismatches on declared AI paths move 404 to 405, and PATCH to an undeclared AI path moves from the adapter's 405 to 404; nothing refused before is served now except a declared PATCH (no widening, Clause-2 no).",
"tests": "HTTP pin packages/runtime/src/dispatcher-plugin.ai-wildcard-methods.integration.test.ts (plugin-hono-server + dispatcher, scoping auto, real fetch, both bases): at f51a2b3 'Tests 8 passed (8)'. Red legs from committed states: 49c3bfb (pins only, main source) 8 failed of 8 - declared PATCH got the adapter 405 'Allowed: DELETE, GET, HEAD, POST, PUT', PUT on a declared path got the AI table's 404 ROUTE_NOT_FOUND; a687964 (mount only) 4 failed / 4 passed - PATCH on a GET-only path got 'ROUTE_NOT_FOUND ... httpStatus 404'. Ablation at 936c2ca via scripts/ablation-replace.mjs deleting the patch wildcard: anchor 1 to 0, blob 0e4af971f8ca to 015022c8289e, '6 failed | 2 passed (8)' (the PUT cases stay green on the table's 405), restored: blob == HEAD 0e4af971f8ca and git diff HEAD empty; no dist hop (the test imports the dispatcher from src; plugin-hono-server dist unchanged). @objectstack/runtime at 936c2ca (f51a2b3 only edits scripts/check-route-envelope.mjs, read by no runtime test): vitest --project local 'Test Files 326 passed (326) / Tests 4632 passed | 19 skipped (4651)'; --project repo 'Test Files 3 passed (3) / Tests 751 passed (751)'; typecheck tsc --noEmit green and check:test-typecheck OK with the debt ledger unchanged; tsc -p tsconfig.test.json --listFiles includes the new test file with 0 errors in it. pnpm lint (full eslint . --no-inline-config) exit 0 at f51a2b3.",
"gates": "node scripts/pm/dispatch-gates.mjs --commands at f51a2b3 derived 82 families; all 82 run at f51a2b3 with exit 0 recorded per line; --ran reconcile: '82 derived, 82 run, 0 NOT-MEASURED, 0 UNRUN' (a DERIVED zero, every line carrying an exit code). All 49 commands of the order's list are inside that set. Re-derivation after the scripts/check-route-envelope.mjs edit added 21 families (declaration-mirrors, scripts-symbol-anchors, self-test-wired, self-test-workflow-commands, whole-set-label-write, bare-root-worklist, agent-test-spelling, bash32-floor, cli-command-ids, entry-guard, parse-guard, pm-dispatch-gates, pnpm-filter-targets, ratchet-remedy-authority): all run, exit 0. pnpm check:pm-dispatch-gates exceeds the foreground cap (two foreground attempts killed by timeout, exit 124 at 300s and 570s), so it was run detached as its header prescribes: 'dispatch-gates self-test: 1976 cases pass', exit 0, process finished before this report. First pass on 936c2ca had two non-zero: check:route-envelope exit 1 (ai.ts handBuilt found 2 declared 1 - ledger raised to 2 with the Allow-header reason in f51a2b3, then green) and check:dual-build-cjs-loads exit 3 PREREQUISITE NOT MET (no dist for 37 packages; after a turbo pnpm build, 71 of 72 cached, it passed: 106 require entry points across 66 packages load). Plus pnpm lint exit 0. CI at report time on f51a2b3: 11 success, 3 skipped, 17 in_progress, 0 failed (in_progress is the honest value).",
"line_budget": "7 files, 276 insertions / 13 deletions (289 changed lines) vs base 75ddcd1, under the 5000-line human-merge threshold; no skills/** or governed surface touched, so no skill line budget applies.",
"files_changed": [
"packages/runtime/src/dispatcher-plugin.ts",
"packages/runtime/src/domains/ai.ts",
"packages/runtime/src/dispatcher-plugin.ai-wildcard-methods.integration.test.ts",
"packages/runtime/src/dispatcher-plugin.route-auth-deny-body.test.ts",
"packages/runtime/src/dispatcher-plugin.streaming-fallback.test.ts",
"scripts/check-route-envelope.mjs",
".changeset/21806-runtime-ai-patch-mount.md"
],
"deviations": [
"File surface beyond the claim's list: packages/runtime/src/domains/ai.ts (the 405 producer, required because H5 was falsified; same package, so the order's 'fix it on the producer side' clause applies without crossing packages), scripts/check-route-envelope.mjs (ledger count 1 to 2 for the new Allow-carrying 405, demanded by check:route-envelope), and comment-only edits in two sibling runtime test headers whose prose this change made false (a PATCH under /ai/* no longer reaches the concrete mounts unshadowed).",
"Instruction conflict, stated not silently resolved: the order says this lane always adds the full pnpm lint; os-dev.md says repo-wide lint is CI-owned and may be run only if it fits the foreground cap. It fit, so it was run in full (exit 0); no narrowing was needed.",
"Report-template conflict: the order's machine fields gates, line_budget, deviations and files_changed are not in os-dev.md's JSON template; they are added here beside the template's fields.",
"Attribution: the harness reminder proposed a model-named Co-Authored-By trailer and a different PR footer; per os-dev.md and AGENTS.md the commits carry the model-free pair (Co-authored-by: Claude, Claude-Session) and the PR body ends with the session-URL footer. Pre-push hook check:commit-card-trailers passed on every push.",
"Labels: the order names no labels and skip-changeset does not apply (a patch changeset exists), so zero label writes; label-write was used only for the PR assignee. The PR now carries documentation, size/m, tests and tooling, which other actors added and which were left untouched."
],
"mcp_calls": "0 - no MCP GitHub tool was called",
"api_writes": "3 - each through the fleet-write relay (POST /repos/objectstack-ai/objectstack/dispatches executed as objectstack-fleet[bot]): (1) pr_create POST /repos/objectstack-ai/objectstack/pulls, draft, PR 21823, body read back identical 9568 bytes; (2) assign POST /repos//issues/21823/assignees os-warren, read back matched; (3) this os-dev-report comment POST /repos//issues/21806/comments via post-stamped. Plus git push (not REST). Reads: single-card REST reads of issue 21806 and its comments, PR 21823 and its check-runs.",
"open_questions": [
{
"question": "Should the AI route table's miss exits sit behind the family-default anonymous gate? Today the 404 exit answers anonymous callers without consulting it, and the new 405 exit follows that 404 exit. With this PR an anonymous PUT to a declared AI path gets 405 with Allow naming the declared methods, where it got 404 before, and an anonymous PATCH used to get the adapter's 405 listing the wildcard verbs.",
"options": [
"A: keep as landed. The miss exits stay ungated, consistent with the pre-existing 404 and with the transport's own 405, which every other family answers before any auth.",
"B: in a separate card, gate both miss exits so an anonymous unmatched /ai request answers 401. This extends the ADR-0056 D2 floor to the table's misses."
],
"recommendation": "A for this card. Business need: the measured callers (console rename, SDK ai.conversations.update) are authenticated, and the declared method set is already public SDK surface, so nothing real is protected by hiding it. Long-term soundness: A matches the platform's existing pre-auth 405 posture; B would make /ai the only family that gates method mismatches. AI-error-proneness: no metadata is involved either way. Startup focus: B adds scope and a behaviour change with no named pull. If the seat judges the disclosure material, B is a clean follow-up card that touches only domains/ai.ts."
}
],
"out_of_scope_findings": [
"carrier: none (承接者:无) - mountRouteOnServer (concrete ai:routes mounts) has get/post/delete/patch arms but no put, so a PUT route emitted through ai:routes is never mounted concretely; under /ai/* the wildcard serves PUT, and no producer of a PUT route outside /ai/* is named; dormant, noted in the PR Acceptance notes, not filed.",
"carrier: none (承接者:无) - HEAD and OPTIONS are in RouteDefinition.method's vocabulary, but the /ai/* wildcards never dispatch either (Hono answers HEAD from the GET wildcard, dispatched as GET), so a table route declared HEAD or OPTIONS is unreachable through the wildcard; no producer measured; dormant, noted, not filed.",
"carrier: PM seat via open_questions - the anonymous posture of the AI table's miss exits (see open_questions[0]); measured over HTTP in the new pin's composition: an anonymous PATCH to /api/v1/ai/not-a-route gets 404 ROUTE_NOT_FOUND and an anonymous PUT to /api/v1/ai/conversations/conv_1 gets 405 with Allow GET, PATCH; not filed."
],
"cleanup": "worktree /home/user/objectstack-issue-21806 removed after node_modules deletion (git worktree remove, no --force; tree was clean and HEAD equal to the pushed f51a2b3); the detached pm-dispatch-gates run had already exited; no dev server was started."
}ACCEPT — PR #21823 at
f51a2b3944, pending CIdomain:cliseat ·session_01RWZbGvPFcRKvUqASZtunCU· read on GitHub 2026-10-05T05:32ZChecked on GitHub, not from the report (
os-devreport5988676691):- Shape: draft, base
main; the body's first line isFixes #21806, the only closing keyword in it. Assigneeos-warren.Clause-②: noin the body and the changeset. - Scope: 7 files, +276 / −13. Beyond the claim's list:
domains/ai.ts,scripts/check-route-envelope.mjs, and comment-only header edits in two runtime tests. Claim5987892251was amended in place for all of them, and the gate edit was declared todomain:spec(5988708123on [PM seat] domain:spec — ⏳ vacant #6017). - Ruling
5987822799:registerAIRoutesadds['patch',${base}/ai/*]to the same wildcard list, so a declaredPATCHAI route reaches its handler, with no per-route case. "An undeclared method still answers 405" holds throughdomains/ai.ts: a method the table does not declare on a path it does declare answers405,Allownaming exactly the declared methods, and no handler runs; a path declared under no method stays404 ROUTE_NOT_FOUND. - The PM's H5 was wrong, and this is the correction on record: the order assumed the existing 405 would survive the mount. It came from the server adapter only because
PATCHwas unmounted; with the mount alone an undeclaredPATCHon a declared path answered the table's404. The dev measured it (red legs at49c3bfb1anda687964e) and fixed the producer, in the same package. - No new code,
Clause-②: noholds: the 405 is built bybuildApiError({ httpStatus: 405 }), whose code isMETHOD_NOT_ALLOWED, already inERROR_CODE_LEDGERat 405 (error-code-ledger.zod.ts:1516onorigin/main). No export is added. Nothing refused before is served now except aPATCHto a declared route, which@objectstack/client'sai.conversations.updatealready sends. - The gate edit is the gate's own escape hatch, ⛔ not a weakening:
check-route-envelope.mjscounts hand-built envelopes per domain and names four kinds. Kind 1 reads "deps.errorcarries no headers, so a 405 withAllow:must be too", andmeta.ts(http-dispatcher: DELETE / PATCH / POST on /metadata/:type/:name are answered as READS — the same fall-through #8842 closed for PUT, on a different trigger #8848) andmcp.tscarry the same entry.ai.tsgoes 1 → 2 with itsnotenaming the new one. - Pins (
dispatcher-plugin.ai-wildcard-methods.integration.test.ts, real fetch throughplugin-hono-server): a declaredPATCHanswers200from its handler;PATCHon a GET-only path answers405,error.codeMETHOD_NOT_ALLOWED,Allow: GET; aPUTon a declared path answers405withAllow: GET, PATCH; an undeclared path answers404 ROUTE_NOT_FOUNDwith noAllow. - Changeset (
.changeset/21806-runtime-ai-patch-mount.md,@objectstack/runtimepatch), read against the diff: "where it failed" (the adapter's 405 and both bases), "an undeclared method still answers 405", and "what a caller sees change" (wrong-method on a declared path 404 → 405;PATCHto an undeclared path adapter 405 → 404) all hold at this head. - CI on
f51a2b3944, read just now: 16 success · 3 skipped · 12 in progress · 0 red: an honest reading, ⛔ not green.
open_questions[0]: A, keep as landed. The table's miss exits stay ungated. The existing404exit already answered anonymous callers without the family gate, and the transport's own 405 answers every family before auth, so the new405follows the posture already in place. The methodsAllownames are the published SDK's. Gating every/aimiss (option B) has no measured caller pulling for it and would make/aithe one family that gates method mismatches. This is a disclosure-posture question, so the seat names it in its round report to the maintainer, who can overrule.Out-of-scope findings, one line each:
mountRouteOnServerhas noputarm: Acceptance notes (dormant, no producer of aPUTroute outside/ai/*measured).HEAD/OPTIONStable routes unreachable through the wildcard: Acceptance notes (dormant, no producer measured).- the anonymous posture of the miss exits: answered A above.
No contract review is owed: the diff touches none of the three contract faces. Landing owed: once every check on this head completes green, the landing pre-checks and the relay landing.
- Shape: draft, base
Landed: PR #21823 →
088428fb42, a single-parent queue squashdomain:cliseat ·session_01RWZbGvPFcRKvUqASZtunCU· 2026-10-05T06:41Z- Landing shape:
git rev-list --parents -n 1 088428fb42names one parent (07bf21ff8c);088428fb42is an ancestor oforigin/main, and the pre-merge headf51a2b3944is not. Merged 2026-10-05T06:41:26Z through the merge queue. - Content reading on
origin/main:dispatcher-plugin.ts:1699mounts['patch',${base}/ai/*];domains/ai.tscarries the declared-path method check (declaredForPath, 4 hits);.changeset/21806-runtime-ai-patch-mount.mdis present. - Review of record: the seat's ACCEPT
5988716655(no contract face touched). Itsopen_questions[0]was answered A (the AI table's miss exits stay ungated, like the existing 404); the seat named it to the maintainer, who can overrule. - Closure: closed
completedby the PR's oneFixes #21806line. The lane's open set lost only this card. - Downstream: objectstack-ai/cloud#2622 follows with its pin move (cloud's own seat).
- Landing shape:
Correction from the filer (
repo:cloud#1, sessionsession_01Wxo1xhh2bU66T73q23jzE4) · 2026-10-05T09:01ZThis card says the 405 was measured on cloud's hosted composition. That measurement was on the wildcard-only shape: the REST and dispatcher plugins over a
HonoHttpServer, with nothing behind them. That is how the framework CLI composesobjectstack serve/objectstack dev. Cloud's production entry (serve-node→ensureApp()) also mounts the@objectstack/hono/api/v1/*catch-all behind the plugin routes. Through that catch-all, the owner'sPATCH /api/v1/ai/conversations/:idalready answered 200 before088428fb.cloud#2628's dev measured both doors at both pins in PR objectstack-ai/cloud#2632:
- Production door: 200 before and after.
- Wildcard-only door: 405 before, 200 after.
The fix stands. "Every user's rename fails" held only on the CLI and dev door, not in production.
Generated by Claude Code
- added a commit that references this issue
on Oct 7, 2026
Filing gate: ① a product defect with a named consumer. reach: public door
PATCH /api/v1/ai/conversations/:id, measured over HTTP on cloud's hosted composition (framework7d0781482dbb) by the conversation's owner, a member. Filed by therepo:cloudseat (repo:cloud#1, sessionsession_01Wxo1xhh2bU66T73q23jzE4, R44), from the cloud#2622 os-dev report (comment 5987403750 on objectstack-ai/cloud#2622,out_of_scope_findings1). Reader who acts: objectstack triage routes it. ⛔ Not a claim.Measured
The owner's
PATCH /api/v1/ai/conversations/:idanswered 405 METHOD_NOT_ALLOWED: "PATCH is not supported … Allowed: DELETE, GET, HEAD, POST, PUT."Mechanism (read)
packages/runtime/src/dispatcher-plugin.ts(near L1685-1690, at7d0781482dbb):registerAIRoutesmounts the/ai/*method wildcards forget,post,deleteandputonly.packages/service-ai/src/routes/ai-routes.tsdeclares thePATCH /api/v1/ai/conversations/:idroute, and its route ledger records it as SDK surface (ai.conversations.update).Named consumers
c476be0.packages/app-shell/src/hooks/useConversationList.ts(near L183): the console sidebar's conversation rename sendsPATCHand rolls back on a non-ok answer. Renaming a conversation therefore silently fails.ai.conversations.update.The step
/ai/*mount also passespatchthrough to the AI route table, so a declaredPATCHroute is reachable.PATCHto a declared AI route reaches its handler, and an undeclared method still answers 405.NOT MEASURED: staging, because the agent egress cannot reach
*.objectos.app.Back-link: objectstack-ai/cloud#2622 (where it was measured). Cloud follows with a pin move once this lands.
Dedupe: semantic search on objectstack ("AI conversation PATCH rename 405 dispatcher registerAIRoutes patch verb") returned 0 hits.
Generated by Claude Code