chore(ci): remove obsolete self-hosted Supabase health-gate - #15
Open
yamancan wants to merge 413 commits into
Open
chore(ci): remove obsolete self-hosted Supabase health-gate#15yamancan wants to merge 413 commits into
yamancan wants to merge 413 commits into
Conversation
…estHash to helpers.ts
… trionlabs8004 bindings
…eAgentUri and generateRequestHash
Addresses the one real issue from adversarial review: a non-transient foreign-key violation (e.g. a genuinely dropped/unparseable parent Registered event) made the batch defer forever, silently stalling the whole reputation or validation stream behind one bad event. - Track consecutive defers per contract in indexer_state.defer_attempts (migration 039); reset to 0 on a completed scan, increment when a batch is left un-advanced. - After MAX_DEFER_ATTEMPTS (5) consecutive defers, treat the failure as permanent: log at error level (so monitoring can surface the dropped event) and skip it so the checkpoint can advance and unblock the stream. Adds getDeferAttempts and a defer_attempts arg to updateCheckpoint, plus a test for the escape-hatch path. Regenerated the edge copy. https://claude.ai/code/session_018NmC15oDtEFsBE7tUzDW1B
- Add a toText() helper: scValToNative returns a Uint8Array for non-UTF-8
Soroban String fields, and String(uint8array) yields a comma-joined byte
list ("104,105") rather than text. Route String/Symbol/URI/tag fields
through toText so non-UTF-8 bytes decode correctly; also replaces the
scattered String(data.x ?? '') pattern.
- Drop the redundant throw in the validation_response parser (it was caught by
the wrapper and returned null anyway, identical to the out-of-range path).
- Correct the scValToNative comment in reputation.ts (u64/i128 always return
bigint; only u32/i32 return number).
Verified against the emitting Rust contracts: event names, topic ordering,
and value ranges match; the metadata_set body is always an ScMap (single-field
contractevent default), so bytesToUtf8(data.value) is never passed undefined.
https://claude.ai/code/session_018NmC15oDtEFsBE7tUzDW1B
… cleanup (S2 review) - Cold-start retention clamp (HIGH): on a fresh checkpoint, startLedger was the deploy ledger, which ages out of the RPC retention window and makes getEvents fail with "start is before oldest ledger" every run (permanent wedge). Probe the oldest retained ledger and clamp the scan start to it; older events are recoverable only via the backfill script. - Request limit:10000 (max) instead of the default 100 -> ~100x fewer pagination round-trips on busy contracts / cold-start catch-up. Cursor pagination still walks the full window, so correctness is unchanged (Soroban getEvents: empty page = end-of-stream, verified against the spec). - Soft time budget: runIndexer accepts a deadline; the loop stops cleanly between pages/contracts and the edge function passes a 110s budget (< its 120s hard timeout) so the lock is released and the checkpoint stays consistent instead of the isolate being killed mid-write. - retry.ts: clamp the Retry-After delay by maxDelayMs (a `Retry-After: 120` would otherwise sleep 120s, overriding maxDelayMs and the time budget). - Collapse the three per-contract checkpoint reads into one getCheckpointState round-trip. - Remove the dead last_cursor column (migration 040): it was written every run but never read (resume is always from last_ledger+1). Refresh the generated indexer_state types (also adds the missing defer_attempts). - Surface escape-hatch drops as result.skippedEvents for observability. https://claude.ai/code/session_018NmC15oDtEFsBE7tUzDW1B
- Treat transient PostgreSQL SQLSTATEs as retryable, not just FK violations: add 55P03 (lock_not_available), 40001 (serialization_failure), 40P01 (deadlock_detected). The lock_timeout on insert_feedback_response (migration 026) raises 55P03 under contention "for the caller to retry" — but the caller dropped it as a permanent error. Now they defer and retry next run. - Tighten release_indexer_lock (migration 041): migration 038's `(p_owner IS NULL OR ...)` made a NULL-owner release an unconditional delete that could wipe a live, differently-owned lock. Now release only deletes the row it owns (NULL matches NULL). Give the backfill script its own owner token so its lock release is fenced too. - Switch insert_feedback_response idempotency from tx_hash to the globally unique Soroban event id (migration 041 + event_id column and partial unique index). tx_hash is nullable (NULL = NULL is never true, silently disabling idempotency) and not unique per event (a tx emits many events / two responses in one tx would collapse). The parser now carries event.id; backfill builds a deterministic txHash:index id. - Refresh the stale generated RPC signatures in types.ts (p_owner, p_event_id). - Document UriUpdated's destructive-but-replay-safe UPDATE. https://claude.ai/code/session_018NmC15oDtEFsBE7tUzDW1B
…4 review)
- Make the drift --check airtight: it now walks the edge copy and fails on
orphan files (a renamed/deleted source file or stale committed file would
otherwise linger in _shared and ship to Deno while --check stayed green).
- Add `deno check` of the real edge entrypoint to CI. tsc uses Node module
resolution and structurally cannot catch Deno-only failures (a specifier
that resolves under Node but 404s under Deno, npm: specifiers, Deno.* APIs);
the drift check only proves the copy matches source, not that it loads under
Deno. Verified the check passes against the current entrypoint.
- Guard sdk-config.ts (vendored from packages/sdk) with a drift test: the
contract addresses / deploy ledger drive which chain state the indexer reads,
so a stale copy after a redeploy would silently index the wrong contracts.
- env.ts now reads Deno first (the edge copy runs under Deno; a partial process
polyfill could shadow the real Deno.env value) and treats empty string as
unset so a blank in one runtime can't mask a real value in the other.
- Generator transform also rewrites dynamic import('./x.js') specifiers, not
just static `from` imports.
- Pin the generated edge copy to LF via .gitattributes so a CRLF checkout
can't produce phantom drift failures; the --check now also normalizes EOL.
- Correct the docker functions-init comment: it is a local-dev regenerate that
writes the tree, not CI's read-only --check.
https://claude.ai/code/session_018NmC15oDtEFsBE7tUzDW1B
… lock-release blip mask a successful run The retention clamp previously guarded only cold start (lastLedger === 0). If the indexer is stopped longer than the RPC event-retention window (as little as ~17h), the saved checkpoint ages out: startLedger = lastLedger+1 falls before oldestLedger, every getEvents hard-fails with "start is before oldest ledger", the checkpoint never advances, and the contract is wedged permanently. Probe the oldest retained ledger and clamp forward on ANY start, not just cold start, without assuming retention is long enough to make the resume case impossible. The forced-forward jump is surfaced via the existing gap detector; the skipped range remains recoverable through the backfill script. Also harden runIndexer's finally: a transient release_indexer_lock transport failure no longer replaces a successful run's result with a throw (the lock self-heals via acquire's 180s stale-sweep). It is logged instead. Tests: +1 resume-after-downtime clamp case; probe-call-count and clamp-message assertions updated for the now-unconditional probe. 67 passing, tsc clean, edge copy regenerated and drift-checked. https://claude.ai/code/session_018NmC15oDtEFsBE7tUzDW1B
…ll contracts
The retention-clamp probe (getOldestLedger) was awaited unguarded, so a
transient RPC failure on it threw out of runIndexerLoop and aborted the whole
run — including contracts fully in-window that needed no clamp at all. On a
flaky/rate-limited RPC that meant zero forward progress on every affected run.
Wrap the probe per-contract: on failure, log and skip the clamp for that
contract and proceed with the un-clamped start (if it really has aged out of
retention, the getEvents below fails in the existing per-contract catch and
only that contract defers), while the others still advance. A failed probe is
not memoized, so the next contract re-attempts it.
Also:
- Decouple the defer-counter decision from the run-global result.timedOut via a
per-contract deadlineStopped flag, so a future refactor that clears/reorders
the shared flag can't mis-count a deadline stop as a defer.
- parsers/identity.ts: route the metadata `key` topic through toText, not
String(), so a non-UTF-8 key is decoded as text instead of being persisted as
a comma-joined byte list ("104,105").
Tests: probe-failure-does-not-abort-run (all contracts still advance);
Retry-After clamped to maxDelayMs; toBigInt integer/type rejection. Regenerated
the edge _shared copy; drift check + tsc clean. 73 passing.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…igrations) The deploy SSHs into the VPS and ran `bash scripts/build-edge-function.sh`, which (after the indexer pipeline consolidation) calls `node sync-indexer-to-shared.js`. The VPS deploy shell has no `node` on PATH, so the step exited 127 and the job died BEFORE applying migrations — every deploy failed at the edge-rebuild step. The rebuild is redundant: supabase/functions/_shared/indexer/ is committed and CI enforces zero drift from packages/indexer/src (`sync-indexer-to-shared.js --check`), so `git reset --hard origin/main` already places the correct, current edge artifact. Remove the host-side rebuild so the deploy needs no `node` and proceeds to migrations + edge restart. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Prod has the `mpp_enabled` column but schema_migrations does not record
037_mpp_column.sql as applied (it was applied via a different path), so the
deploy's migration loop re-runs 037 and dies on `ADD COLUMN` /
`CREATE INDEX` ("column already exists"), never reaching 038-041.
Guard both with IF NOT EXISTS (the function/grant are already CREATE OR REPLACE).
Schema outcome is identical; the migration now re-runs as a no-op, gets recorded,
and the runner proceeds to the new indexer migrations. Matches the idempotent
pattern 038-041 already follow.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Mark 037_mpp_column.sql as applied (ON CONFLICT DO NOTHING) before the migration
loop so the runner skips re-running it. Its effects (mpp_enabled column, index,
search function) are already present on prod from an earlier apply path that
never recorded it here; re-running collided ("column already exists", then
"cannot change return type"). This unblocks the genuinely-new 038-041 indexer
migrations. Reconcile line is temporary and will be removed once recorded.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
037_mpp_column.sql is now recorded in prod's schema_migrations (the reconcile ran and the deploy confirmed "037 - already applied"), so the temporary reconcile INSERT is no longer needed. Future deploys skip 037 normally. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- register: add "Are you an agent?" quick-add card at the top of /register (copyable npx skills add ... --skill 8004stellar, view-skill link, agnostic copy) - rebrand "Claude Code Skills" -> "Agent Skills" and use agent-agnostic wording across intro/one/developers - standardize skill ids on the ...stellar convention: 8004s -> 8004stellar; rename skills/x402s -> skills/x402stellar (frontmatter, cross-refs in both SKILL.md files, README, TECHNICAL, and webapp install commands)
… leaderboard hardening Verified follow-ups from the multi-agent indexer review: - SSRF: resolve agent_uri hostnames (A/AAAA) and block any that point at private/internal space (cloud metadata 169.254.169.254, internal docker services); also catch IPv4-mapped IPv6. Fails open on restricted runtimes. - Index the OZ `transfer` event so agents.owner follows NFT transfers, and delete orphaned agent_metadata on transfer (contract's clear_all_metadata emits no per-key event). Applied to both the live indexer and the backfill. - Stop silently dropping events on a non-retryable write failure: defer the batch (checkpoint un-advanced) so a transient blip is retried; widen the retryable PG-code set; wrap getLatestLedger in withRetry. - Migration 042: widen leaderboard avg/total_score cast to numeric(40,2) (a single 1e20 feedback otherwise overflows numeric(10,2) and permanently freezes the refresh), and throttle refresh_leaderboard() with a force variant for the backfill's final rebuild. - API: sanitize+cap service fields in normalizeServices; scope the feedback_responses query to the current page. - Document the registry-id data-wipe step in the redeploy runbook.
The indexer wrote one PostgREST round-trip per event, so a 100-event page = 100 sequential writes and a 5000-event backfill = 5000 calls, overflowing the 120s function budget. A page whose events ALL parse to the same pure-upsert type — the dominant backfill shape (long runs of NewFeedback / ValidationRequest / Registered / MetadataSet) — now collapses into ONE multi-row upsert. Mixed pages, single events, and the RMW types (UriUpdated, AgentWallet*, AgentTransferred, FeedbackRevoked, ResponseAppended, ValidationResponse) keep the per-event path. Correctness: - Parse-once: the page is parsed a single time and feeds both the fast path and the per-event fallback (no double-parse; preserves parser call counts). - The checkpoint (maxLedger) advances only on a SUCCESSFUL bulk write, so a deferred page is never checkpoint-skipped. - On ANY bulk failure the page falls back to the per-event loop, preserving the FK-defer / skip-after-N handling; re-apply is safe (idempotent upserts, and a failed PostgREST statement rolls back wholesale, so nothing partial leaks). - In-batch dedupe keeps the last row per ON CONFLICT key (Postgres forbids a repeated conflict key in one upsert array); last-wins mirrors per-event replay. Tests: 89 pass (+8) — bulk mappers/dedupe (db.test.ts) and fast-path / fallback / mixed-page integration (indexer.test.ts). Dual-runtime synced; tsc, deno check, and sync --check all clean.
Chunk each bulk-upsert writer at BULK_CHUNK_SIZE (500) rows so a single homogeneous event page (up to EVENTS_PAGE_LIMIT events, each up to the on-chain 4KB metadata/URI cap) can't balloon into one multi-MB INSERT ... ON CONFLICT that trips statement-size / memory / time limits. Chunks upsert sequentially; a later-chunk failure throws and the indexer falls back to the idempotent per-event path, so an already-applied earlier chunk is harmless. Routes all four bulk writers through one upsertChunked helper (also collapses the four near-identical upsert+assert blocks). Addresses the sole Low finding from the multi-lens security review of the bulk-upsert branch; the two Info items were verified non-issues (the processed counter matches per-event semantics; the warn-log error.message matches the file's established ops-logging pattern).
…e bind-mount (#10) docker restart reuses the existing container, which can keep a stale handle to the bind-mounted supabase/functions/ after the deploy's git reset --hard swaps files — leaving edge serving deleted/old code or wedged entirely (observed in prod outage). Switch to docker compose up -d --force-recreate so the container is rebuilt and the mount + loaded functions are fresh. -p pins the existing Dokploy project (verified via container labels) so no duplicate stack is spawned; --no-deps keeps it edge-only so the DB is never touched by a code-only deploy. Validated with compose --dry-run against the live VPS: targets trionlabs8004-edge-functions Recreate, nothing else.
…ransfer, honest liveness, dead-letter) (#9) * fix(indexer): 8004 faithful-mirror hardening (ledger guards, atomic transfer, honest liveness, dead-letter) Grounds five indexer fixes in the 8004 faithful-mirror principle (the DB must monotonically + idempotently reflect chain truth and never silently lose data): G4 monotonic ledger guards: UriUpdated / FeedbackRevoked / ValidationResponse blind UPDATEs now carry a PostgREST `.or(<col>.is.null,<col>.lte.N)` guard so a replayed/out-of-order older event can no longer overwrite newer mirrored state. Feedback stays append-only (revocation is a flag, never a delete). (migration 046) G4 atomic transfer: AgentTransferred is one SECURITY DEFINER apply_agent_transfer RPC (owner change + metadata clear in a single, ledger-guarded transaction), so the mirror is never observed mid-transfer with the new owner but the prior owner's metadata attached. (migration 047) G3 honest liveness: updateCheckpoint stamps last_advanced_at only on real forward progress; both health endpoints read it, so a wedged-but-still-ticking contract now reads stale instead of fresh. (migration 044) G3 dead-letter: the MAX_DEFER_ATTEMPTS skip and the RPC retention clamp now record the dropped event/range in indexer_dead_letter (durable, replayable) instead of an ephemeral log; a silent drop is no longer indistinguishable from censorship. Best-effort write that can never re-wedge the loop. (migration 045) G9 bulk/per-event equivalence test: pins all four bulk-eligible writers to byte-identical rows vs the per-event path, so the mirror is provably path-independent. Also fixes a latent upsertChunked deno-check failure (TS2208/TS2345) from the prior bulk-upsert change. Migrations 044-047 are additive and safe to apply ahead of the code. Local: tsc clean, 102/102 vitest, sync in-sync, deno indexer+api gates clean. Verified by a 5-agent adversarial pass (PostgREST/.or semantics, plpgsql atomicity, merge-duplicates upsert-omit, dead-letter wiring). * fix(backfill): canonicalize writes to match the live indexer (G8) The backfill script is the README-mandated rebuild path after every redeploy and the only recovery for dropped/clamped events, but it had drifted from the live indexer in ways that corrupt the rebuilt mirror. Align it with the (now verified) canonical writers so a backfilled mirror == a live-indexed one: - value: i128 feedback value was Number(d.value) — lossy past 2^53 and a throw on a bigint via JSON.stringify. Serialize as a decimal string into the numeric(78,0) column, matching db.ts value.toString(). - idempotency: the feedback INSERT had no onConflict, so a re-run / retry threw duplicate-key. Upsert on (agent_id,client_address,feedback_index). - validation parity: drop new_feedback with feedbackIndex < 1 or valueDecimals outside 0..18, matching parsers/reputation.ts. - atomic transfer: AgentTransferred did a non-transactional PATCH + DELETE; route it through the same apply_agent_transfer RPC (migration 047) the live path uses. - monotonic guards: set uri_updated_ledger / revoked_ledger / response_ledger (migration 046) so a later live event compares correctly after a backfill. backfill is lock-exclusive with the live indexer and replays in ledger order, so these make the two writers produce identical state. Script-only, reversible. * feat(web): surface validation pillar, per-record chain links, and indexer freshness (G5/G6/G3) Make three 8004 faithfulness properties VISIBLE in the explorer (additive UI only; svelte-check clean): G5 validation tab: the Validation Registry (third, independent trust axis) was fully indexed but invisible — the agent page had only Metadata + Reputation tabs and never queried the validations table. Add a read-only Validation tab querying validations, presenting each attestation by its VALIDATOR address (distinct provenance from client feedback), response score / pending, request+response URIs and hashes, and per-record chain links. Feedback and validation are shown as distinct trust signals, never blended. G6 per-record verifiability: tx_hash/created_ledger are stored but were never surfaced. Add explorerAccountUrl/explorerLedgerUrl helpers; link the owner, identity registry, and each feedback + validation record to stellar.expert; show the on-chain feedback hash so a consumer can independently re-verify. G3 freshness badge: a faithful mirror must expose its own freshness. Add a root layout read of indexer_state and a small "synced to ledger N" badge (amber "syncing…" when the laggard contract's last_advanced_at is >5min stale), pairing with the honest-liveness backend change. Degrades to nothing if unavailable.
Dokploy already auto-deploys this stack on push (git-watch): it pulls main, runs docker compose up which applies migrations via the in-compose migrate service and recreates the edge runtime. The SSH workflow was a SECOND deploy path racing Dokploy — it git reset --hard + recreated the edge container, then Dokploy's own deploy (~8 min poll lag) ran compose up which swapped the bind-mounted functions/ inode under the still-running edge container without recreating it, orphaning the mount (worker boot error: could not find an appropriate entrypoint) and taking edge down. Confirmed: migrate service ran on Dokploy's deploy (exit 0, migrations applied); webapp app was rebuilt by Dokploy on the same push. Stop deploying from CI. Instead wait for Dokploy to deploy the pushed commit, then assert the stack came up healthy (migrate exit 0, every migration file applied, edge not orphaned, indexer advancing, no fresh 500s), failing the run loudly if not. This is the safety net that was missing when prod went down silently twice.
Add registry_address (NOT NULL DEFAULT = live mainnet identity registry) to agents + agent_metadata + feedback + feedback_responses + validations, plus network on agents, plus parallel composite UNIQUE indexes alongside the existing PKs/uniques. Purely additive: no PK/FK/unique-constraint change, no read-query or leaderboard change. The constant DEFAULT backfills every existing row and covers indexer writes that do not yet supply the column, so the current indexer and all reads keep working unchanged. The composite indexes exist so Phase B can switch onConflict targets to them with no breakage window. Phase A of 5 toward composite (registry_address, id) keying so an identity-registry redeploy no longer requires TRUNCATEing reputation history (README 'After redeploying'). A-C are non-destructive; D promotes the composite PK and drops legacy constraints. Staged against a fresh prod schema dump on local postgres 17: clean apply + idempotent re-apply + post-state verified (cols, 5 unique indexes, DEFAULT backfill).
normalizeRecord() camelCased every key as it walked the response tree, including the opaque user-controlled `metadata` map. A key like `social_links` came back as `socialLinks`, so SDK reads could not faithfully reproduce on-chain state. Pass the `metadata` value through verbatim and add a round-trip regression test. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Rate limiter failed OPEN: a non-inet IP key ('unknown', a spoofed/empty
X-Forwarded-For) made check_rate_limit's `p_ip::inet` cast throw, which
the caller swallowed and allowed the request. Now isPlausibleIp() funnels
bad keys to a shared '0.0.0.0' sentinel, and migration 049 wraps the cast
in a BEGIN/EXCEPTION (also fail-closed) and revokes the function's EXECUTE
grant from anon/authenticated/PUBLIC (it was an unauthenticated INSERT
primitive), granting only service_role.
minScore was filtered per-page in JS, so the reported total and hasMore
were wrong and pages could under-return. Resolve qualifying agent ids from
leaderboard_scores (bounded, highest-scored first) and constrain both the
count and data queries at the DB, keeping pagination consistent.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- mpp-client: drive the RPC from stellarConfig.rpcUrl instead of the hardcoded 'soroban-mainnet.stellar.org', which is not a usable public endpoint and broke every live mainnet MPP charge. The network-mismatch guard above guarantees the configured RPC matches the challenge network. - mpp-client: apply the computed validUntilLedger to the sponsored auth entry before signing (previously dead code), bounding the signature to the challenge expiry; drop the now-unused authorizeEntry import. - x402: add the primary network guard at the call site (TryAgentPanel checks paymentRequired.accepts[].network before signing), with a defense-in-depth guard in the signer for parity with MPP. - api proxy: positive-allowlist the endpoint and reject any '.'/'..' path segment so a crafted path can't traverse to another edge function. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…e hatch A transient getEvents failure set scanCompleted=false, which incremented defer_attempts identically to a genuine write defer. Five consecutive RPC blips could therefore trip the skip-after-N escape hatch and drop a real event on the next run — undermining the "never silently drops" guarantee. Track fetch failures separately (fetchFailed) and preserve the defer counter on a fetch failure (like a deadline stop) instead of incrementing; only a retryable WRITE defer increments it. Update the existing test to the corrected expectation and add a regression test. Edge copy re-synced. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The reproducible-build claim (byte-identical mainnet/testnet WASMs) was never verified by CI. Add a committed digest manifest (contracts/wasm.sha256) and a `make verify-wasm` target, gate it in CI, verify Cargo.lock is up to date before the build (cargo metadata --locked), assert the build does not mutate the lock, and run `make test` with --locked. Also add a least-privilege `permissions: contents: read` block. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The shipped 8004stellar skill misstated on-chain behavior that would lead
integrators to ship broken or wrong-trust code:
- append_response is permissionless (any authenticated caller; responder is
recorded in the event), not agent-owner-only.
- validation response is a 0-100 score (values >100 revert), not 0/1.
- generateRequestNonce is async and must be awaited.
- Identity/Reputation/Validation error-code lists now match errors.rs.
The x402 skill's stellar-sdk compatible-version range is made consistent.
README notes the agent identifier's `{network}` uses friendly labels
(`mainnet`) rather than the CAIP-2 chain id (`stellar:pubnet`).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The allowlist guard rejected only literal '..'/'.'/'' segments, but fetch()'s WHATWG URL parser percent-decodes each segment once before normalizing. So `agents/%2e%2e/%2e%2e/indexer` passed the raw check yet collapsed to `.../api/indexer`, escaping the api/v1 namespace to reach a different edge function on the same gateway. Decode each segment once (matching that single decode), reject dot-segments and embedded separators on the decoded value, then rebuild the upstream path from re-encoded segments so the URL handed to fetch() is already normalized and no traversal can survive parsing. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The trionlabs8004 Supabase stack was migrated from the self-hosted Dokploy docker-compose deployment to Supabase Cloud (2026-07-17) and the self-hosted stack was decommissioned. deploy-webapp-vps.yml waited for Dokploy to deploy the pushed commit to the self-hosted compose stack, then health-checked containers (trionlabs8004-migrate / -db / -edge-functions) that no longer exist. On any push touching webapp/supabase/migrations|functions, packages/indexer, config.ts or docker-compose.supabase.yml it would hang 15 min waiting for a deploy that never happens, then fail. Cloud migrations/functions now ship via `supabase db push` / `supabase functions deploy` against the migration-managed repo. A cloud-side health-gate can be added later as a separate workflow. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
Important Review skippedReview was skipped as selected files did not have any reviewable changes. 💤 Files selected but had no reviewable changes (1)
⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (1)
You can disable this status message by setting the Use the checkbox below for a quick retry:
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
yamancan
marked this pull request as ready for review
July 18, 2026 00:51
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Ne
.github/workflows/deploy-webapp-vps.yml(self-hosted Supabase deploy health-gate) kaldırıldı.Neden
trionlabs8004Supabase stack'i self-hosted Dokploy docker-compose'dan Supabase Cloud'a taşındı (2026-07-17) ve self-hosted stack decommission edildi.Bu workflow, push edilen commit'i Dokploy'un self-hosted compose stack'ine deploy etmesini bekleyip ardından artık var olmayan container'ları (
trionlabs8004-migrate/-db/-edge-functions) health-check ediyordu. Şu path'lerden birine push olduğunda 15 dk hiç gerçekleşmeyecek bir deploy'u bekleyip fail ediyordu:webapp/supabase/migrations/**,webapp/supabase/functions/**webapp/packages/indexer/**,webapp/packages/sdk/src/core/config.tswebapp/docker/docker-compose.supabase.ymlEtki
supabase db push/supabase functions deployile (migration-managed repo) gidiyor.Sonraki (opsiyonel, ayrı PR)
İstenirse cloud tarafı için yeni bir health-gate eklenebilir: push'ta
supabase db push+ Supabase Management API ile migration/edge/indexer-lag doğrulaması (cloud creds GH secrets'a eklenerek).🤖 Generated with Claude Code