feat(knowledge): correct auth-field docs + pagination tests for enterprise connectors (#10538)

[mrveiss]

Thinking Path

Audited #10538 before writing anything, per instructions. The Slack/Confluence/Jira KB
ingestion connectors already exist, fully implemented, in
autobot-backend/knowledge/connectors/{slack,confluence,jira}.py, delivered across five
already-merged PRs: #12189 (scaffold), #12199 (offline mock/replay connector + fixtures),
#12219 (operator setup guide), #12225 (auth-schema/credential-key alignment, closing #12221),
plus later cross-cutting refactors #12659 (shared _load_ts/_store_ts extraction) and
#12979/#12999 (pooled HTTP client migration). All three connectors implement the exact same
AbstractConnector contract as every other connector in the family (base.py) — auth_schema(),
output_schema(), {"status_code","body"} HTTP-result shape, _load_ts/_store_ts checkpointing,
get_http_client().tracked_request() for outbound HTTP (no raw aiohttp.ClientSession) — and are
reachable end-to-end: knowledge/connectors/__init__.py imports them behind the
kb_enterprise_connectors feature flag, api/knowledge_connectors.py's _SUPPORTED_TYPES lists
slack/confluence/jira, and GET /knowledge_base/connector_types /
POST /knowledge_base/connectors operate on whatever ConnectorRegistry has registered. 69
existing offline tests (mocked HTTP, real assertions on content/metadata, auth-failure paths, and
the feature-flag gating) were green before this PR and remain green after.

So there was no connector left to build. Digging into why the issue was still open (owner's own
comments on the issue thread), the remaining ask was live-credential validation against real Slack/
Confluence/Jira tenants — which the owner separately clarified (2026-07-24 comment) is a per-install
operator task using the secrets manager, not something dev/CI does. That leaves the issue's code-side
acceptance criteria fully met, but two real gaps surfaced during the audit that a straight "close, no
changes" would have left behind:

1. The operator setup guide had gone stale. PR fix(kb): align connector credential reads with declared auth_schema (#12221) #12225 (auth-key alignment) landed after PR
   docs(kb): operator setup guide for Slack/Confluence/Jira connectors (#10538) #12219 (the setup guide) and renamed the connector's config-read keys (bot_token->token,
   email/api_token->username/password) to match the declared auth_schema() — but the guide
   was never updated. It still told operators to send the old, now-rejected field names, still
   described the mismatch as unresolved, and still said "Issue feat(knowledge): Slack / Confluence / Jira KB ingestion connectors #10538 stays open for the
   live-credential wiring" — which is exactly backwards now that fix(kb): align connector credential reads with declared auth_schema (#12221) #12225 fixed it. An operator
   following the guide today would get 422s.
2. No pagination test. All three APIs paginate (response_metadata.next_cursor for Slack,
   start offset for Confluence, startAt/total for Jira) and the connector code handles it, but
   the offline test suite only ever mocked a single page.

Both are small, directly in-scope fixes for #10538's own deliverable, so I fixed them here rather
than filing follow-ups, per the "fix pre-existing issues discovered along the way" rule.

What Changed

• docs/knowledge-base/enterprise-connectors-setup.md: corrected every config-key table, curl
  example, and source-line citation to the current, post-fix(kb): align connector credential reads with declared auth_schema (#12221) #12225/tech-debt(backend): extract knowledge-connector base — _load_ts/detect_changes duplicated across connectors #12659 code (token /
  username+password instead of bot_token / email+api_token; Redis checkpoints now cited to
  the shared base.py helpers instead of removed per-connector copies). Replaced the "Known
  limitation: auth-field mismatch" section with an "Auth-field alignment" section documenting that
  fix(kb): align connector credential reads with declared auth_schema (#12221) #12225 resolved it.
• autobot-backend/knowledge/connectors/{slack,confluence,jira}_test.py: added one multi-page
  pagination test per connector, each driving discover_sources() through a real multi-call
  sequence (2-3 mocked pages) and asserting every page's sources were collected — not merely that no
  exception was raised.
• No connector code changes — nothing to build; see the audit above.
• Filed fix(http): raw aiohttp.ClientSession ceiling regressed 12→13 (undocumented offenders, #12992) #13041 for an unrelated, pre-existing regression discovered while running the verification
  suite: test_raw_client_session_ceiling_12992 is currently red on Dev_new_gui (13 raw
  aiohttp.ClientSession sites vs. the recorded ceiling of 12) due to 3 undocumented new raw
  sessions in agent_loop/search/config_declared_provider.py, orchestration/dag_executor.py, and
  services/slm_client.py — none touched by this PR, filed separately per convention.

Verification

Full offline connector suite (69 tests, unaffected + 3 new pagination tests = 72 net after adding,
47 shown for the 3 touched files):

knowledge/connectors/slack_test.py       .............. (14 incl. new pagination test)
knowledge/connectors/confluence_test.py  ............... (15 incl. new pagination test)
knowledge/connectors/jira_test.py        .................. (18 incl. new pagination test)
============================== 47 passed in 1.13s ==============================

Full connector-family suite (unaffected files) still green:

knowledge/connectors/{slack,confluence,jira}_test.py
knowledge/connectors/mock_test.py
knowledge/connectors/connectors_init_gating_test.py
============================= 69 passed in 10.25s ==============================

flake8 --max-line-length=120 and black --check --line-length=120 on the 3 touched test files: clean.

Registration/reachability proof (no code path added — confirming the existing one):

• autobot-backend/api/knowledge_connectors.py:84-86 lists slack/confluence/jira in
  _SUPPORTED_TYPES.
• autobot-backend/knowledge/connectors/__init__.py:85-88 imports (and therefore
  @ConnectorRegistry.registers) all three behind kb_enterprise_connectors.
• GET /knowledge_base/connector_types (knowledge_connectors.py:346-369) and
  POST /knowledge_base/connectors (:401-482) operate on ConnectorRegistry.registered_types() /
  .create() dynamically — no hardcoded connector list to update.

Baseline diff: test_raw_client_session_ceiling_12992 fails identically before and after this PR
(13 vs. ceiling 12, same offender list — none in knowledge/connectors/, confirming this PR didn't
touch the ratchet). Pre-existing, filed as #13041.

Startup import smoke: unaffected — this PR touches only test files and docs, no importable modules
changed.

Model Used

Claude Sonnet 5

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No hardcoded values detected that have SSOT config equivalents!