canonical(secrets): close the final two #15276 tail sites in the vault-resolution sweep

[mrveiss]

Thinking Path

#15276 listed 29 direct (non-vault) credential read sites as the tail of the
#15267/#15268 sweep. A ground-truth check against the tree found that a prior
PR (#15291, merged as c13c92882 on Dev_new_gui) had already migrated 27 of
the 29 and left the remaining two as TRACKED_GAP entries in
repo_tests/credential_vault_resolution_allowlist.py, ratcheted at
TRACKED_GAP_CEILING = 2 — both blocked on the same obstacle: their file is a
KNOWN_LARGE-grandfathered file already at its frozen line-count ceiling
(scripts/python_file_size_known_large.py / repo_tests/python_file_size_ratchet_baseline.py,
both a CHECKED PAIR that fail if the recorded count and the real count
diverge), and a bare from services.provider_key_vault import resolve_provider_key
plus its call site adds lines that ceiling may not grow to absorb (#14236's
"the exemption freezes" rule).

The two remaining sites:

• services/execution/claude_code_backend.py:317 — anthropic_api_key
• services/notification_service.py:450 — smtp_password

Each needed a genuinely line-neutral edit rather than a raised ceiling. For
claude_code_backend.py, restructuring _resolve_api_key's early-return into
a single if not self._api_key: block freed exactly the line the local
import needs. black (a repo pre-commit hook, confirmed already clean on both
files pre-edit) then inserts its mandatory blank line after any function-local
import, which would have pushed the file 1 line over ceiling — offset by
merging self._mcp_host = mcp_host / self._mcp_port = mcp_port in
__init__ into one tuple assignment. notification_service.py got the same
treatment: the local import took the slot the blank line after the docstring
used to occupy, and merging smtp_host/smtp_port into one assignment
absorbed black's inserted blank after the import. Both files verified
black --check clean and at their exact recorded ceiling (661 / 634 lines)
after the edit.

AUTOBOT_SMTP_PASSWORD also needed to join SERVICE_CREDENTIAL_KEY_NAMES (and
_CREDENTIAL_CONSUMERS) in provider_key_vault.py — anthropic_api_key
already had a resolved-key entry via LLM_PROVIDER_KEY_NAMES, but nothing
previously hydrated the SMTP password from the vault, since #15291 explicitly
removed it from that set rather than claim false coverage.

No new guard was needed for regression protection: #15269/#15278/#15280's
existing guard (repo_tests/credential_vault_resolution_guard_test.py) already
scans every git-tracked production .py file (not a fixed site list) for any
direct read of a credential-shaped ssot_config field not routed through
resolve_provider_key, already carries a vacuity floor
(test_the_ssot_config_field_scan_finds_known_credential_fields asserts
len(fields) >= 20), and test_allowlist_entries_still_correspond_to_a_real_match
would fail if either newly-migrated site's stale ALLOWLIST entry were left in
place. Removing both entries and dropping TRACKED_GAP_CEILING from 2 to 0 in
credential_vault_resolution_ratchet_test.py is therefore the complete
regression-guard extension: a reintroduced bare read at either site is
automatically unlisted and fails the guard.

What Changed

• autobot-backend/services/provider_key_vault.py — AUTOBOT_SMTP_PASSWORD
  added to SERVICE_CREDENTIAL_KEY_NAMES and _CREDENTIAL_CONSUMERS; module
  docstring updated to describe the line-neutral technique used at both
  remaining sites instead of the "deliberately not here" note.
• autobot-backend/services/execution/claude_code_backend.py —
  _resolve_api_key now resolves ANTHROPIC_API_KEY through
  resolve_provider_key (env wins, then vault) instead of reading
  config.anthropic_api_key directly. Line-count-neutral (661 lines, matches
  KNOWN_LARGE/ratchet-baseline ceiling exactly).
• autobot-backend/services/notification_service.py — _send_email now
  resolves AUTOBOT_SMTP_PASSWORD through resolve_provider_key instead of
  reading config.smtp_password directly. Line-count-neutral (634 lines,
  matches ceiling exactly).
• repo_tests/credential_vault_resolution_allowlist.py — both TRACKED_GAP #15276 entries removed (now real, vault-routed reads, not gaps).
• repo_tests/credential_vault_resolution_ratchet_test.py —
  TRACKED_GAP_CEILING lowered from 2 to 0 to match.

Ground truth vs. the issue body (all sites verified against the tree, not
assumed): every one of the other 27 named sites (llm_shared/adapters/*,
llm_shared/providers/*, services/provider_health/providers.py,
agent_loop/slack_hook.py, security/threat_intelligence.py,
initialization/lifespan.py, integrations/capability_registry.py,
knowledge/base.py) was already migrated to resolve_provider_key by #15291
— confirmed with a fresh git grep and a direct run of the guard's own
production_credential_reads() scanner (see Verification), not by sampling.
No premise in the issue body was found false; the "third-party/service
credential" classification and the exact resolve_provider_key(env_var, config_value) call shape from #15267/#15268 were followed as instructed, with
no parallel resolution path, no new helper duplicating the seam, and no
config-direct fallback introduced.

Verification

• python3 -m pytest repo_tests/credential_vault_resolution_guard_test.py repo_tests/credential_vault_resolution_ratchet_test.py repo_tests/credential_vault_prose_strip_test.py repo_tests/python_file_size_ratchet_test.py -q — 63 passed.
• Direct scan: production_credential_reads() (the guard's own scanner) run
  against the full tree returns 0 unlisted direct reads and 0 stale
  allowlist entries (allowlist size 24, all AUTH_BOOTSTRAP/NOT_A_READ,
  zero TRACKED_GAP).
• python3 -m pytest autobot-backend/tests/services/notification_service_test.py autobot-backend/chat_workflow/delegation_test.py autobot-backend/integrations/capability_registry_test.py autobot-backend/agent_loop/search/registry_test.py autobot-backend/voice_processing/realtime/openai_provider_test.py -q — 126
  passed (all consumers of the touched modules).
• python3 -m flake8 and python3 -m black --check — clean on all three
  edited source files.
• function_length_checker.py (the pre-commit-function-length hook's
  engine) — no violations on either edited file.
• wc -l on both edited KNOWN_LARGE files matches their recorded ceiling
  exactly (661 / 634) — the CHECKED PAIR (scripts/python_file_size_known_large.py
  / repo_tests/python_file_size_ratchet_baseline.py) needed no change.
• git push ran the local pre-push hook, which selected
  repo_tests/credential_vault_resolution_ratchet_test.py as the relevant
  test and reported pytest: all relevant tests pass.
• Not run locally: the full repo_tests/ and autobot-backend/tests/ suites
  (very large; a full run was started in the background but did not finish
  producing output within the working session — CI is the source of truth for
  the full suite). Local interpreter is Python 3.10.12 against a
  requirements.txt declaring floors as high as 3.12-targeted packages (90 of
  206 declared versions unmet) — every test invocation above carries pytest's
  own "environment is BELOW the declared dependency floors" notice, so a local
  pass does not stand in for a CI pass, only for "this specific surface is not
  broken."
• No credential value is logged or placed in an exception message on any path
  touched: resolve_provider_key is a pure env_var or cache.get(...)
  lookup with no logging; neither edited call site adds a logger.* or
  raise referencing the resolved value.

Model Used

Claude Sonnet 5

Closes #15276

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No new hardcoded values of either class — ssot and other both block.

Known backlog in pipeline-scripts/hardcoded_values_baseline.txt is suppressed and tracked in #14371.

[mrveiss]

Full local run of `repo_tests/` completed (15m36s, ran in background after the PR body was written): 1843 passed, 7 failed, 10 skipped, 2 xfailed. All 7 failures are pre-existing local-environment artifacts, unrelated to this change (none touch `provider_key_vault.py`, `claude_code_backend.py`, or `notification_service.py`), confirmed by re-running each in isolation:

• `pip_ignore_scope_test.py` (4 failures) — `ModuleNotFoundError: No module named 'tomllib'`; `tomllib` is 3.11+ stdlib, this session runs Python 3.10.12.
• `sdk_response_model_contract_test.py::test_every_sdk_response_model_is_paired_or_explicitly_unpaired` (1 failure) — passes standalone (`1 passed`); a test-isolation/ordering artifact when run inside the full suite, not a real regression.
• `validate_access_control_reporting_test.py::test_a_probe_exits_the_code_the_shell_exported` (2 failures, params 20/37) — the probe subprocess shells out to `/usr/bin/python3` directly and gets `ModuleNotFoundError: No module named 'autobot_shared'`; a local PYTHONPATH/venv artifact, not the interpreter this repo's own tests run under.

This supersedes the "not run locally" note in the PR body's Verification section — the full suite was run and the only failures are these three known-unrelated clusters.