First off — thank you for shard-first v2 and the per-namespace locking work; both made today's diagnosis and repair far more tractable than it would have been a month ago.

We hit a recurring index corruption (twice: 2026-06-27 and 2026-07-08) and root-caused it today from the bundle source. Sharing the full evidence chain in case it's useful.

What we observed

The datastore index accumulated hundreds of thousands of double-prefixed entries (agentic-tooling/agentic-tooling/…) that looked like phantoms. On the 2nd occurrence: 371,989 of 631,410 entries (index grown to 287MB), which pushed shards over the upload ceiling and wedged every write. A sibling repo's namespace (watchers) turned out to be the live carrier with another 270,106 doubled keys (index 175MB → 45MB after strip).

Root cause (traced in the s3.js bundle, v2026.07.02.1)

The keys weren't phantoms — they were real objects with real index entries. The local cache root (~/.swamp/repos/<id>/) had accumulated a nested directory named after the namespace itself (with its own .namespace.json, data/_catalog.db, etc. — a full second store copy, 1.9GB/389,959 files) plus sibling copies of five foreign namespaces. Because push's walk covers the whole cache root and (with an empty S3 prefix) uploads paths verbatim, the nested tree gets uploaded and indexed as doubled keys. Index-add timestamps cluster in batches of exactly 25 (DEFAULT_PUSH_CONCURRENCY), consistent with normal pushChanged walks. We found the same pollution on all four cache roots across two machines.

Two adjacent findings from the repair, offered as data points:

One sidecar had dirtyPathsOverflowed: true with doubled paths queued — as far as we could tell, a post-cleanup push would then have issued mass deletes for the doubled keys. We purged the cache dirs and reset the sidecars before any push, so this is unconfirmed-but-plausible rather than observed.
After index repair we also found the reverse skew: 215 objects present in S3 but absent from the index (recoverable via a listing walk, but there's no self-heal in that direction today).

We haven't identified what created the nested copies (the oldest carries a registration date of 2026-06-18) — possibly a mis-scoped invocation on our side. We've added a local watchdog tripwire that alerts on the signature, so we'll capture the creator if it recurs and will happily follow up here.

Ask (gently)

Would it be feasible for push to constrain its walk to the store's own expected layout, or to validate/refuse keys that resolve outside the configured namespace before indexing/uploading them? Either would turn this failure mode from silent index growth into a visible error. We have full artifacts (pre-purge cache listings, stripped-key manifests, timestamp histograms) and are glad to share any of it, or to test a candidate fix against our setup.

Environment

CLI 20260708.012813 / @swamp/s3-datastore 2026.07.02.1 / MinIO (forcePathStyle), two writer machines (macOS + Linux), namespace at bucket root plus prefixed sibling namespaces.

Upstream repository: https://github.com/systeminit/swamp-extensions

Environment

Extension: @swamp/s3-datastore@2026.07.02.1
swamp: 20260708.012813.0-sha.0f7166fe
OS: darwin (aarch64)
Deno: 2.8.3
Shell: /bin/zsh

02Bog Flow

Shipped

7/8/2026, 8:12:22 PM

Click a lifecycle step above to view its details.

03Sludge Pulse

stack72 assigned stack727/8/2026, 6:33:40 PM

stack72 commented 7/8/2026, 8:12:36 PM

Thanks @mgreten for reporting this! The fix has been merged and a release is on its way. We appreciate your contribution to swamp.

Bundled extension model cannot deserialize AWS SDK S3 XML that works in direct Deno

gcp/sheets spreadsheets-values: batch_get crashes with "Request with GET/HEAD method cannot have body"

Token creation API and UI scope selector for collective tokens (phase 4 of #960)

Make Cloudflare CRUD methods factory-aware so list-discovered resources can be managed in-place

[sc] quest-grant Mongo ops (recentHeartbeat/signalCounts/countIssueShipped) regressed to ~5s timeouts — missing index

Add @swamp/kubernetes/statefulset model type

serve: workflow.approvals and workflow.run.search still full-scan the workflow-run store after the #1284 data.query pushdown

Sync guard skips bound namespace directory during push — 0 files pushed after namespace migrate

Flaky #1052 preparePush test: mtime-granularity race on Linux tmpfs

command/shell renders JSON-object stdout lines as the literal string 'undefined'

Issue redactor treats the MongoDB authSource parameter as a secret, corrupting reproduction steps

namespace unset --migrate silently deletes artifact history via catalog-rebuild GC (~484k files lost)

Docs: document repo root resolution from subdirectories

Expose the swamp.club feed as an RSS feed

Email invitation is not linked to account created from invite URL

Add serve, datastore, vault, and worker wildcard scopes to token registry (phase 4 prep, #960)

datastore setup extension drops 'namespace' from --config, making extensions that require it impossible to configure

Docs: update manual to document --yes/-y as the standard confirmation-skip flag

Docs: update data.query and CEL expressions references for SQL pushdown

Docs: document collective token CLI usage in API key authentication pages

Add a --release-notes option to `swamp update` (or even make it the default)

Docs: manual should state workflow unknown keys are rejected and placement is step-only (follow-up to #1240)

Make GCP CRUD methods factory-aware so list-discovered resources can be managed in-place

serve: read ops (data.query, workflow.approvals, workflow run.search) linearly scan the whole datastore — ~350–600ms floor regardless of match count

Running "swamp workflow run ..." in a subdir of the swamp repo fails

Guide agents to inspect reports after failed model and workflow runs

serve: main JS thread holds ~40–70% of a core while idle (no clients, no runs, --no-schedule)

Search box collapses to near-zero width on mobile on the Extensions page

Extensions page: Search bar missing on mobile

spreadsheets-values: every method fails with 'Missing required path parameter: spreadsheetId' — implementation reads a global argument the schema does not declare

namespace migrate/set workflow broken for extension datastores: migrate skipped by sync guard, set not idempotent, migrate-index empties shards

ExtensionInfo.latestVersion typed 'string' but registry returns null; harden pull/resolve for prerelease-only extensions

workflow.run: relay a custom model method's output as method_output (command steps already do)

promptChoice infinite-loops on non-interactive stdin

gcp _lib: unconditional x-goog-user-project header causes USER_PROJECT_DENIED (403) for service accounts without the serviceUsageConsumer role

extension push: issue-verdict review is indistinguishable from missing/stale

swamp extension version crashes with 'Invalid CalVer version: "null"' when only a prerelease (beta/rc) version is published

Pushing an extension after creating a public git repo that is the same as other previously pushed extensions doesn't update the "Verified public repository"

`datastore setup filesystem` never transfers content when leaving a sync-based (remote) datastore — copies only the catalog and hardcodes `filesPulled: 0`

Scaleway models: create methods should not require the resource ID they generate (ipId/gatewayId/serverId/...)

gcp/calendar events.list reads wrong array field ('defaultReminders' not 'items') → always returns 0 events

provisioner.provision fails with 'Cannot read properties of undefined (reading trim)' at method_execution_service.ts:852

method_execution_service.ts:852 crashes on undefined.trim() when writeResource receives undefined instance name

CLI support for collective API tokens (phase 3 of #960)

@swamp/gcp/calendar: service-account & ADC auth mint cloud-platform-scoped tokens → 403 on Calendar API

Scope enforcement and whoami introspection for collective tokens (phase 2 of #960)

fix(software-factory): DeclaredSchemaSchema.enum rejects boolean literals

Surface cloud-provider labels (scaleway, digitalocean, oracle/oci) as first-class search/filter facets on the Extensions page

Profile link in genesis subdirectory does not navigate back to main profile

Profile stats: per-user daily rollup table + MV in ClickHouse

Reconcile collective-token CLI activity to members (attribute → bonus → stamp → even-split)

Telemetry stats consumer: unsanitized subcommand in Mongo update path drops user metrics silently

Telemetry stats consumer: insert_id dedup commits before the write it guards — any upsert failure is silently swallowed on retry

Job-level labels: block is silently ignored, not rejected

Issue submission redactor silently corrupts report bodies

CLI telemetry: fold third-level verbs into subcommand + canonicalize aliases

Docs: update extension info reference to document extensions (type grafts) in content metadata

Docs: swamp skill repo-structure reference still documents repo-local .swamp/telemetry/ as live

Add safe adoption and lifecycle methods for existing Hetzner resources

Docs: update manual telemetry reference for user-global spool and telemetry.yaml opt-out

export const extension can graft methods onto foreign-collective types, undisclosed in extension info / contentMetadata

Add lookup support to the Cloudflare DNS records model

Expand token scope registry to full taxonomy (phase 1 addendum, #960)

Docs: document user-global telemetry (telemetry.yaml opt-out, spool location, telemetry stats behavior)

score_daily rollup strands stale grains when a day-less grant re-stamps onto a later day

swamp-club: global rate limiter runs pre-auth and collapses all direct clients into one per-IP bucket — 429s authenticated CLI traffic

Collective API token scopes: domain, persistence, and auth threading (phase 1 of #960)

Telemetry migration from ~/.swamp/telemetry should auto-detect home-as-repo

feed_post_score_events queue grows unboundedly (no TTL, no delete-on-delivery)

Issue-filing skill should require ownership classification before submitting a bug

Interactive finder silently drops the preview pane on narrower terminals, and layout selection is unstable

@-mention autocomplete for users when writing lab ripples

Remove comments from /feed and redesign the page

Interactive finder preview corrupts: overlapping/stale lines when preview content overflows the pane

Score-refresh consumer: harden coalescing against a diverse backlog (bistability)

Remove feed comments — consolidate discussion in Discord

Add a 'my collective extensions' filter to extensions page

Relax the profanity filter

Extension search returns edit-distance noise for short queries ("asdl" → "AWS DEADLINE")

Telemetry recoverOrphaned startup race with multiple replicas (created_at-based)

Grant revoke is never applied to the running server — under `--grant-reload auto` and `manual` alike, a revoked grant stays in force until an explicit `access reload`