Activity

1. dezren39

   added

   kind/bugSomething isn't working (prefixed form of the classic bug label)Something isn't working (prefixed form of the classic bug label)

   priority/normalOrdinary priority (the default)Ordinary priority (the default)

   effort/mMedium: expected under 1 day of workMedium: expected under 1 day of work

   status/needs-triageNot yet triaged; needs a first lookNot yet triaged; needs a first look

   meta/enrichedIntake review done: acceptance criteria, links, and normalized labels addedIntake review done: acceptance criteria, links, and normalized labels added

   on Nov 4, 2026
2. dezren39 commented on Nov 4, 2026

   

   dezren39

   on Nov 4, 2026

   Author

   More actions

   Intake review

   • Related issues/PRs: merged PRs with this issue number in the title: none found in a merged-PR title search. No merged PR explicitly resolves this issue, so it stays open.
   • Notes: Enriched by the fleet monitor (Duty 2, protocol §5): acceptance criteria appended to the body, labels normalized (kind/bug, priority/normal, effort/m).
   • Risks / suggested approach: Generate the served-route list from the server and diff it against the spec; disposition each facade route (implement, or drop from spec). Risk: client codegen may already depend on spec-only routes.
3. dezren39

   added a commit that references this issue on Nov 4, 2026

   Merge pull request #1638 from developing-today-forks/fix/1627-facade-…

   ...

   63c50ee
4. dezren39

   closed this as completedin #1638on Nov 4, 2026
5. dezren39 commented on Nov 4, 2026

   

   dezren39

   on Nov 4, 2026

   Author

   More actions

   Fixed — root cause, fix, and scratch-instance proof

   Root cause. The façade handlers and routes were never the problem. opencode serve composes its listener as WebUi.handler() wrapped around the API router (the transform argument of ServerProcess.start, built in packages/cli/src/server-process.ts). The transform in packages/cli/src/services/web-ui.ts passed only /api, /api/*, /auth/* and /openapi.json through to the router and answered every other path from the embedded web shell: GETs fell back to the SPA index.html (200 text/html), and non-GET/HEAD requests got the shell's bare 405. The completions group is the only API group outside /api — all eight façade paths live under /v1 — so the serve listener advertised the doors in its OpenAPI spec (generated from the Api definition) but routed none of them. The fetch-handler entry (ServerFetch.make, used by tests and the workerd/embedded entries) applies no transform, which is why the façade answered there while serve did not. This matches the issue's suspicion exactly: serve-entry router composition, not route registration.

   Fix. PR #1638 (merged, 63c50eec0d27139b6ebdc57ed85d4d3ace35b5b3): the WebUi transform now passes /v1 and /v1/* through to the API router alongside /api. One condition added, no other behavior change — the shell still owns every non-API path, and unknown /v1 paths now 404 through the router pass-through instead of serving the SPA. Regression coverage in packages/cli/test/web-ui.test.ts (the spec-vs-served check for this composition): the live ServerProcess test now asserts unauthenticated /v1/* meets the API auth gate, authenticated façade requests return façade JSON (models list, chat 400 invalid_request_error, systemone {error} envelope, embeddings JSON), and the stub-API tests cover /v1 pass-through with the UI disabled and /v1/missing → 404. 4/4 tests pass (116 assertions); tsgo -b typecheck in packages/cli passes; oxlint clean.

   Scratch-instance proof (fixed build from source, own port 13499, own data dir/DB, the deployed server's config overlay; plugin list emptied in the scratch config copy only, because a cold plugin cache cannot finish activation inside awaitPlugins' 5-second window — the deployed server's warm cache is unaffected):

   • GET / → 200 text/html (SPA shell intact)
   • GET /v1/models without auth → 401 application/json (the API auth gate, not the shell)
   • GET /v1/models with auth → 200 JSON: the full route catalog (deepseek-flash-plan, haiku-plan, terra-plan, sonnet-plan, opus-plan, … owned_by opencode-route) plus the embeddings entry local-embed/snowflake-arctic-embed-m-v2.0
   • GET /v1/models/deepseek-flash-plan → 200 JSON single-model object
   • POST /v1/chat/completions with a real route (deepseek-flash-plan, one-line prompt) → 502 JSON server_error — the request went through route resolution and failover and was ledgered: 12 route_attempt rows in the scratch DB with session_id NULL (the façade convention), including attempts on opendesign and antigravity-direct. It did not complete because the scratch instance has an empty credential store (and the antigravity client secret is separately broken today, antigravity: client-secret extraction (first GOCSPX match) stale for agy 1.4.11 — refresh fails invalid_client #1633) — a well-formed routing error from inside Generate, not a 405 and not the SPA.
   • POST /v1/responses (unknown model) → 400 JSON {"error":{"message":"Model unavailable: opencode-route/no-such-model","type":"invalid_request_error"}}
   • POST /v1/messages (unknown model) → 400 JSON in the Anthropic error shape, same message
   • POST /v1/systemone with the recorded clef classifier probe (the exact body transcribed in packages/core/test/facade-parity.test.ts) → 401 with the door's upstream-failure envelope, and a route_attempt row (provider cloudflare-ai-gateway, model @cf/cloudflare/clef-flash). Parity check: the same probe sent to the still-running standalone door at 127.0.0.1:18742/v1/systemone returned the byte-identical response (same envelope, same Cloudflare code 10000 "Authentication error", same 401). The upstream authentication failure is a pre-existing credential condition that currently affects the door and the façade identically; the façade's systemone is behaviorally indistinguishable from the door for this request.
   • POST /v1/embeddings → 502 JSON {"error":{"message":"Transport error (POST http://127.0.0.1:18752/v1/embeddings)","type":"server_error"}}, with route_attempt rows for local-embed. State of the embeddings door: the route is registered and served by the façade; its backend is still the HTTP sidecar registration (nothing listens on :18752, per step 5's gate), so requests answer a well-formed transport error. The Embeddings: bake the local backend into the fork (native in-process), retire the :18752 sidecar #1554 native-onnx flip is untouched by this change, as instructed.

   Deployment note. A plain restart of the running server does not suffice: the deployed binary is the nix build of the currently pinned fork rev, which predates this fix. The running server carries the fix only after the code repo's flake pin for the fork advances to a rev containing 63c50eec and the server is redeployed/restarted on the new build (a switch round, or a targeted redeploy of the server package plus a restart of the login-session server process). The switch round in flight at the time of this fix still uses the old pin.

   Unblocks (follow-ons not performed here, per the sequence's ownership):

   • Door front migration: clef-proxy (18742) callers repoint to the façade #1410 — the clef repoint (classifier-endpoint.json url → http://127.0.0.1:13420/v1/systemone) can proceed once the deployed server carries the fix; the parity check is unblocked (expect the current upstream Cloudflare auth failure to appear identically on both sides until that credential is repaired).
   • Retire bin/opencode-route-proxy.mjs (developing-today/code) behind the façade #1153 — the OpenClaw rewire (baseURL → 127.0.0.1:13420/v1) four-part proof can be re-attempted on the redeployed server; route-proxy retirement stays behind that proof, as designed.
   • Door front migration: muse-proxy (18747) callers repoint to the façade #1155 — the muse attestation probe's façade leg is cleared (its Meta-quota gate, resetting Sun Nov 8 18:00 CST, is unchanged).
6. developing-today commented on Nov 4, 2026

   

   developing-todayBot

   on Nov 4, 2026 – with developing.today

   More actions

   Live verification on the deployed server — the redeploy this issue was held open for

   The fix is now live on the standing server. Path: code PR anomalyco#5567 advanced the fork pin to 63c50ee (and, via inputs.opencode.follows, ended the cascade float — code#5519), merged as 4d2eeecb; a switch on 2026-11-04 (~10:36 CST, SWITCH_EXIT=0) cycled opencode.service onto the pinned build. Verified live, in order:

   • Cascade check (pre-switch): with a fresh XDG_CACHE_HOME, the evaluated server package reads opencode-2.0.30+63c50ee-wrapped and flake.lock is byte-identical before/after evaluation (sha256 ebafe50b…) — the float no longer re-resolves anything.
   • Running build: the unit's ExecStart is /nix/store/m00436…-opencode-2.0.30+63c50ee-wrapped/bin/opencode serve --service --hostname 127.0.0.1 --port 13420; the process exe resolves to /nix/store/4r20ria3…-opencode-2.0.30+63c50ee/bin/.opencode-wrapped — exactly the outPath the PR's scratch verification built.
   • Façade battery on :13420 (deployed auth):
     • GET / → 200 text/html (SPA intact)
     • GET /v1/models without auth → 401 JSON UnauthorizedError
     • GET /v1/models with auth → 200 JSON, 28 models, including local-embed/snowflake-arctic-embed-m-v2.0
     • POST /v1/systemone (recorded clef probe) → 401 in the door envelope — byte-identical to the still-running :18742 door's response in the same run (Cloudflare code 10000; that credential is now tracked separately as fix(providers): cloudflare-ai-gateway credential fails upstream auth (code 10000), blocking clef classifications #1651)
     • POST /v1/chat/completions (terra-plan) → 200, a real completion through the façade
     • POST /v1/embeddings → 502 well-formed transport-error envelope (the :18752 sidecar remains down behind the Embeddings: bake the local backend into the fork (native in-process), retire the :18752 sidecar #1554 gate, as designed)
     • Ledger: the battery's POSTs wrote route_attempt rows with session_id NULL (façade convention)

   The issue was closed by @dezren39 earlier today; this comment supplies the post-redeploy live evidence the close was waiting on. The unblocks recorded here have since executed: #1410 (repoint + door retirement) and the #1146 step-3 OpenClaw rewire behind #1153 — evidence on those issues.