Skip to content

BUGS.md

Latest commit

 

History

History
284 lines (184 loc) · 16.9 KB

BUGS.md

File metadata and controls

284 lines (184 loc) · 16.9 KB
Raw

AgentContextOS — Bug Tracker

How to use: Log every bug you hit while testing here. Copy the Bug entry template into the Open section, give it the next BUG-NNN id, and fill what you know — a half-filled entry beats an un-logged one. When a bug is fixed, flip its Status, add the PR/commit, and move it to Resolved. Keep the Summary counts and the Index row in sync.

Companion to TRACKER.md (build status). This file is testing findings, not roadmap.

Summary

Severity Open In progress Resolved Total
🔴 Critical 0 0 0 0
🟠 High 2 0 1 3
🟡 Medium 1 0 3 4
🟢 Low 0 0 3 3
All 3 0 7 10

This session: BUG-002, BUG-003, BUG-004, BUG-007, BUG-008, BUG-009 fixed + verified (code + tests + browser). BUG-005 / BUG-006 investigated → structural backend changes, deferred with root cause below. BUG-010 environmental (couldn't reproduce). The two High "Open" are BUG-005 (audit) and BUG-010 (availability, can't-reproduce).

Index

ID Title Severity Status Area Found
BUG-001 New corpus does not persist (vanishes on reload) 🟠 High ✅ Resolved admin-ui 2026-06-11
BUG-002 Edit corpus menu item does nothing 🟡 Medium ✅ Resolved admin-ui 2026-06-11
BUG-003 Edit webhook subscription menu item does nothing 🟡 Medium ✅ Resolved admin-ui + gateway 2026-06-11
BUG-004 Live Status shows seed/demo data as "Live" with no demo badge 🟡 Medium ✅ Resolved admin-ui 2026-06-11
BUG-005 Immutable audit log returns 0 events for real governance actions 🟠 High 🔵 Open gateway 2026-06-11
BUG-006 Ingest reports success but corpus document_count stays 0 🟡 Medium 🔵 Open gateway 2026-06-11
BUG-007 Metrics "Captured" timestamp triggers a React hydration mismatch 🟢 Low ✅ Resolved admin-ui 2026-06-11
BUG-008 "Lands with Phase 6" preview copy is stale (project is GA) 🟢 Low ✅ Resolved admin-ui 2026-06-11
BUG-009 Query Playground shows "Demo data" badge before the first run 🟢 Low ✅ Resolved admin-ui 2026-06-11
BUG-010 Gateway process exited mid-session (availability) 🟠 High 🔁 Can't reproduce gateway 2026-06-11

Open

BUG-002 — Edit corpus menu item does nothing

  • Severity: 🟡 Medium
  • Status: ✅ Resolved
  • Area: admin-ui (Corpora page) — the update ("U" of CRUD) path is unimplemented across UI + gateway + store
  • Found: 2026-06-11 · Version/commit: main @ 6cbab40 + uncommitted step-3.10 work on build/phase-3/step-3.10-corpus-admin-crud
  • Environment: admin-ui console (Next.js), demo or live mode.

Steps to reproduce

  1. Open Corpora → on any existing row, open the ⋯ row-actions menu.
  2. Click Edit.

Expected: An edit dialog opens to change the display name / description (or other mutable fields), then saves and persists.

Actual: Nothing happens — the menu closes, no dialog, no request.

Evidence / root cause: The Edit menu item is rendered with no onClick — unlike its siblings "View details" (onClick={() => setSelected(c)}) and "Delete" (onClick={() => setDeleting(c)}) (corpora/page.tsx:130). The entire update path is missing:

  1. UI: no editing state, no EditCorpusDialog — Edit is a placeholder item (corpora/page.tsx:130).
  2. Client: no updateCorpus in lib/api.ts — only fetchCorpora / createCorpus / deleteCorpus.
  3. Gateway: no PUT/PATCH /v1/corpora/{id} — the router has GET list + POST + GET {id} + DELETE only (corpora.py). No CorpusStore.update on the SPI.

Notes / suspected cause: The BUG-001 fix implemented Create + Delete but skipped Update, leaving Edit as a dead menu item. This is in scope for the current step-3.10-corpus-admin-crud branch. Full fix spans: gateway PATCH /v1/corpora/{id} + CorpusStore.update (noop + pg) → updateCorpus in api.ts → an edit dialog + onClick in the page. Partial workaround: delete + recreate (loses the corpus id and any ingested docs).

Fix: Implemented the full update path — gateway PATCH /v1/corpora/{id} + CorpusStore.update (SPI + noop + pg, tenant-scoped, embedding model/dim immutable) + updateCorpus in api.ts + an Edit dialog wired to the row menu. Contract + gateway-route tests added; OpenAPI re-exported; verified live (edit → save → persists to gateway).

BUG-003 — Edit webhook subscription menu item does nothing

  • Severity: 🟡 Medium · Status: ✅ Resolved
  • Area: admin-ui (Webhooks page) + gateway — update ("U") path unimplemented, mirroring BUG-002
  • Found: 2026-06-11 · Version/commit: build/phase-3/step-3.10-corpus-admin-crud @ 6e0c2c4

Steps to reproduce

  1. Open Webhooks → create a subscription (or use an existing row).
  2. Open the ⋯ row-actions menu → click Edit.

Expected: An edit dialog opens to change URL / event types / description, then saves.

Actual: Nothing happens — menu closes, no dialog, no request. (Verified twice in isolation.)

Evidence / root cause: The webhook row Edit item has no working handler; there is no updateWebhook in lib/api.ts and no PATCH /v1/webhooks/subscriptions/{id} on the gateway. Same class of defect as BUG-002, for webhooks.

Fix: Mirrored the BUG-002 fix for webhooks — gateway PATCH /v1/webhooks/subscriptions/{id} (signing secret stays masked) + SubscriptionStore.update (SPI + noop) + updateWebhook in api.ts + an Edit dialog; the row's active toggle now also persists in live mode. Gateway-route tests added; OpenAPI re-exported; verified live (edit → save → persists).

BUG-004 — Live Status shows seed/demo data as "Live" with no demo badge

  • Severity: 🟡 Medium · Status: ✅ Resolved
  • Area: admin-ui (Live Status page)
  • Found: 2026-06-11 · Environment: live mode, gateway with no breakers/quotas/etc. registered.

Steps to reproduce

  1. Run the gateway with noop backends (no circuit breakers / quota data).
  2. Open Live Status.

Expected: Sub-sections that fall back to seed data are visibly marked (a demo badge), as Playground/Audit/API-Keys do.

Actual: Header shows "Live"; the Circuit breakers / Quotas / Cost / Drift / Experiments cards render realistic seed numbers (e.g. quotas "4,120,000 / 5,000,000 tokens", "$96.40 / $250.00") with no demo indicator — yet /v1/status/breakers returns {"breakers":[],"total":0}. Fabricated numbers read as live telemetry.

Evidence: per-section useLive(... source) exists in status/page.tsx (breakers L128, quotas L221, drift L325) but source !== "live" isn't surfaced. Genuinely-live data (overall health, components, API-surface telemetry) is correct.

Fix: Each seed-fallback card (breakers / quotas / drift / feedback / cost / experiments) now renders a SourceBadge driven by its own useLive source — "Demo data" when the gateway has no data for it, "Live" when it does. The page header stays "Live" for the genuinely-live health stream, so live telemetry and seed numbers are no longer conflated. Verified live (header "Live", all six cards "Demo data"). (status/page.tsx)

BUG-005 — Immutable audit log returns 0 events for real governance actions

  • Severity: 🟠 High · Status: 🔵 Open (needs gateway triage — may be noop-store limitation)
  • Area: gateway (audit) / admin-ui (Audit Log falls back to demo)
  • Found: 2026-06-11 · Environment: live mode, noop AuditStore.

Steps to reproduce

  1. Against a fresh gateway, create + delete a corpus, create + delete a webhook, ingest a document (all via the UI).
  2. curl -H "X-Tenant-Id: acme" "http://127.0.0.1:8000/v1/audit?limit=200".

Expected: Governance-relevant actions appear in the immutable audit log.

Actual: {"total":0}. The Audit Log page therefore shows demo data (honestly badged). The structured logs (/v1/status/logs) DID capture corpus.created/corpus.deleted, so the events occur — but the audit store/endpoint surfaces nothing.

Notes: Investigated (this session) — root cause confirmed: /v1/audit reads the same store the AuditWriter writes to (app.state.audit_store = audit_writer.store), but the corpus / webhook handlers only emit structured logs (_log.info("corpus.created")) and never call AuditWriter.write — so governance actions are logged but never audited. Deferred (not fixed this session): wiring AuditWriter consistently across every governance handler (corpus / webhook / ingest / …) with correct AuditEvent shapes is a governance design task + product decision (which actions are auditable), not a one-line fix.

Fix: TBD — deferred, see Notes (structural)

BUG-006 — Ingest reports success but corpus document_count stays 0

  • Severity: 🟡 Medium · Status: 🔵 Open (needs gateway triage — may be noop-store limitation)
  • Area: gateway (ingest → corpus store)
  • Found: 2026-06-11 · Environment: live mode, noop stores.

Steps to reproduce

  1. Create a corpus; ingest a document into it via Ingest (POST /v1/ingest/document → 200, "1 chunk / 1 embedding").
  2. Reload Corpora (or GET /v1/corpora).

Expected: Corpus document_count reflects ingested document(s).

Actual: document_count stays 0, so the UI shows "ingested" yet "0 documents" — confusing in a demo.

Notes: Investigated (this session) — root cause confirmed: no document_count increment exists anywhere in the gateway; the ingest pipeline indexes chunks/embeddings but never touches the corpus counter (the create handler's "maintained by the ingest path" comment is aspirational). Deferred (not fixed this session): a correct fix needs a new CorpusStore surface (e.g. increment_document_count / recompute) wired from the ingest path — feature work, not a one-line fix.

Fix: TBD — deferred, see Notes (structural)

BUG-007 — Metrics "Captured" timestamp triggers a React hydration mismatch

  • Severity: 🟢 Low · Status: ✅ Resolved
  • Area: admin-ui (Metrics page)
  • Found: 2026-06-11 · Environment: full page load / refresh of /metrics (dev shows console errors; prod silently re-renders).

Steps to reproduce

  1. Hard-load /metrics (full SSR, not client nav).

Expected: No hydration warnings.

Actual: 8× Text content did not match. Server: "7:05:28 PM" Client: "19:05:29" — a 12h/24h locale mismatch.

Evidence / root cause: metrics/page.tsx:170 new Date(metrics.captured_at).toLocaleTimeString() is locale/timezone-dependent → SSR (server locale) ≠ client. Fires only on full load, not client-side nav.

Fix: Added suppressHydrationWarning to the Proc value node (metrics/page.tsx:170) — the React-sanctioned escape hatch for locale/timezone-dependent timestamps. Verified: tsc + all admin-ui unit tests green.

BUG-008 — "Lands with Phase 6" preview copy is stale (project is GA)

  • Severity: 🟢 Low · Status: ✅ Resolved
  • Area: admin-ui (API Keys, Tenants — copy)
  • Found: 2026-06-11

Actual: API Keys (api-keys/page.tsx:55) and Tenants (tenants/page.tsx:53) banners say management "lands with Phase 6 governance" — but the platform is at v1.0.0 GA (Phase 6 long complete). Misleading copy.

Fix: Reworded both preview banners to drop the stale "lands with Phase 6 governance" claim → "the data shown here is local and not yet backed by the gateway." (api-keys/page.tsx, tenants/page.tsx)

BUG-009 — Query Playground shows "Demo data" badge before the first run

  • Severity: 🟢 Low · Status: ✅ Resolved
  • Area: admin-ui (Query Playground)
  • Found: 2026-06-11

Actual: With a live gateway configured, the Playground header shows a "Demo data" badge on initial load (result source defaults to "mock"); it clears after the first live query. Slightly misleading even though the corpus selector is already live. (playground/page.tsx:183,355)

Fix: Gated the badge on phase !== "idle" so "Demo data" only appears after a run that actually used mock data, not on initial load of a live-configured page. (playground/page.tsx)

BUG-010 — Gateway process exited mid-session (availability)

  • Severity: 🟠 High · Status: 🔁 Can't reproduce (needs info)
  • Area: gateway (availability)
  • Found: 2026-06-11 · Environment: local uvicorn gateway on :8000.

Actual: During UI testing the gateway process (PID 35523) vanished — POST /v1/corpora returned ERR_CONNECTION_REFUSED seconds after a successful GET. No 5xx (which a request-triggered crash would produce), so most likely stopped externally (parallel terminal/session). Restarted under preview control; could not attribute to a UI action.

Notes: If reproducible, capture gateway stderr at exit. Likely environmental, not a code defect.

Fix: n/a (environmental) — logged for visibility.

In progress

Nothing being worked on yet.

Resolved

BUG-001 — New corpus does not persist (vanishes on reload)

  • Severity: 🟠 High · Status: ✅ Resolved
  • Area: admin-ui (Corpora page) — root cause spanned gateway + backend
  • Found / Resolved: 2026-06-11 / 2026-06-11 · Where: working tree on build/phase-3/step-3.10-corpus-admin-crud (uncommitted at time of fix)

Symptom: Create corpus showed a "Corpus created" toast but the row vanished on reload — the UI never called a backend, and /v1/corpora was read-only (no create endpoint; default store in-memory).

Fix: Built the corpus write path end-to-end — POST /v1/corpora + DELETE /v1/corpora/{id} on the gateway, persistent CorpusStore create/delete (noop + pg_corpus_store), createCorpus/deleteCorpus in lib/api.ts, and the Corpora page now calls them when live (corpora/page.tsx:195, :246); demo mode still simulates locally. ⚠️ Update/Edit was not included → BUG-002.

Bug entry template

Copy this block, paste it under Open, bump the id, and fill it in. Leave a field blank (or delete the line) if it doesn't apply — don't let a missing detail stop you from logging.

### BUG-NNN — <one-line title>

- **Severity:** 🔴 Critical | 🟠 High | 🟡 Medium | 🟢 Low
- **Status:** 🔵 Open
- **Area:** <gateway / admin-ui / ragctl / ingest / retrieval / query / policy / backends / docs / …>
- **Found:** <YYYY-MM-DD> · **Version/commit:** <v1.0.1 / git sha>
- **Environment:** <OS · how it was running — docker compose / local `uv run` / deployed>

**Steps to reproduce**
1.
2.
3.

**Expected:** <what should happen>

**Actual:** <what actually happened>

**Evidence:** <error text, log lines, screenshot path, request id / trace id>

**Notes / suspected cause:** <optional — leave for triage>

**Fix:** <PR link / commit once resolved>

Legend

Severity

Meaning
🔴 Critical Data loss, security/tenant-isolation breach, or the platform is unusable. Drop everything.
🟠 High A core surface is broken (query/ingest/retrieve fails, wrong results). No safe workaround.
🟡 Medium Real bug with a workaround, or an edge case off the happy path.
🟢 Low Cosmetic, copy, or minor polish.

Status

Meaning
🔵 Open Logged, not yet being worked.
🟣 In progress Someone is actively fixing it.
✅ Resolved Fixed and verified — has a PR/commit.
⚪ Won't fix Intentional, out of scope, or by design — say why.
🔁 Can't reproduce Couldn't reproduce; needs more detail to reopen.