fix(cron): layer enabled MCP servers onto per-job enabled_toolsets (#23997)

[kshitijk4poor]

Problem

A cron job that set enabled_toolsets to native toolsets (e.g. ["web", "terminal"]) silently got zero MCP tools, while a job with no per-job list got every globally-enabled MCP server. _resolve_cron_enabled_toolsets returned the per-job list verbatim, bypassing the MCP-merge that the platform-fallback branch does via _get_platform_tools. discover_mcp_tools() registered the MCP tools into the registry, but get_tool_definitions(enabled_toolsets=...) kept only the named native toolsets — the agent rejected every mcp_* call as "Unknown tool". (R2 of #23997)

Fix

_merge_mcp_into_per_job_toolsets layers MCP membership onto a per-job allowlist with the same semantics as _get_platform_tools:

• no_mcp sentinel → no MCP servers
• one or more MCP server names already listed → treat as an allowlist (add nothing further)
• otherwise → union in every globally-enabled MCP server

To avoid duplicating the "which MCP servers are enabled" computation (the original PR re-implemented it), this extracts a shared enabled_mcp_server_names(config) helper in hermes_cli.tools_config and has BOTH the gateway/CLI platform resolver and the cron per-job resolver call it — so every path agrees on MCP membership (extend, don't duplicate).

Scope — what was already fixed on main

The issue's headline (bare MCP server names rejected, registry never includes them in cron) was already fixed on main before the issue was filed (commits c10fea8d2 server-alias + 04918345e discover-MCP-before-cron-agent). This PR closes the remaining cron-specific gap (R2). Two smaller residuals are tracked separately, not in scope here:

• R1: server:* / mcp:server alias notation still rejected (→ fix(tools): log unknown toolsets in quiet mode #24104).
• R3: silent drop under quiet_mode (→ fix(tools): log unknown toolsets in quiet mode #24104's warning).

Salvage / credit

Salvaged from #32788 by @sherman-yang (authorship preserved). Reworked to reuse the shared enabled_mcp_server_names helper instead of re-implementing the MCP membership set in cron/scheduler.py (the DRY concern flagged in review).

Tests

tests/cron/test_scheduler.py — per-job native allowlist unions in enabled MCP; no_mcp opts out; explicit MCP name acts as allowlist. Full tests/cron/: 517 passed; tests/hermes_cli/test_tools_config.py: 97 passed (the gateway path still resolves MCP identically after the helper extraction).

Closes

Fixes #23997

---

Review-driven changes (/hermes-pr-review, 3 reviewers — verdict: Approve/ship-ready, no Critical)

• Strengthened the fall-through test: test_resolver_empty_per_job_falls_through_to_platform now stubs _get_platform_tools and asserts the platform branch is actually taken with platform="cron" and its result returned — was previously a near-tautology (result is None or isinstance(result, list)).
• Documented the enabled-flag semantics: the shared enabled_mcp_server_names docstring now notes a server is enabled unless explicitly falsey (false/0/no/off via _parse_enabled_flag); a missing/unrecognized flag is treated as enabled — matching prior gateway behavior.

Reviewers confirmed: the enabled_mcp_server_names extraction is behavior-identical for the gateway (verbatim move of the inline computation, same _parse_enabled_flag); the 3 merge branches are correct with proper allowlist exclusion of unnamed enabled servers; merge runs on the per-job branch only (no double-merge with the platform fallback); the lazy import poses no circular-import risk (cron already imports from hermes_cli widely); disabled-toolset subtraction reliably runs last so a disabled MCP server can't sneak back. Both files pass in isolation (cron 517, tools_config 97).

Note: tests assert the resolved toolset-name list (the contract of these pure resolver functions), not end-to-end tool registration — consistent with the existing resolver tests.

[github-actions]

🔎 Lint report: salvage/f5-cron-mcp-per-job vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 11352 on HEAD, 11348 on base (🆕 +4)

🆕 New issues (3):

Rule	Count
unresolved-attribute	2
invalid-argument-type	1

First entries

tests/cron/test_scheduler.py:76: [invalid-argument-type] invalid-argument-type: Argument to `set.__init__` is incorrect: Expected `Iterable[str]`, found `list[str] | None`
tests/run_agent/test_credits_notices_toggle.py:76: [unresolved-attribute] unresolved-attribute: Unresolved attribute `_credits_session_start_micros` on type `AIAgent`
run_agent.py:2984: [unresolved-attribute] unresolved-attribute: Object of type `Self@get_credits_spent_micros` has no attribute `_credits_session_start_micros`

✅ Fixed issues (1):

Rule	Count
invalid-assignment	1

First entries

tests/run_agent/test_credits_notices_toggle.py:76: [invalid-assignment] invalid-assignment: Object of type `None` is not assignable to attribute `_credits_session_start_micros` of type `int`

Unchanged: 5964 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.

[alt-glitch]

Duplicate of #32788 — both fix #23997 with the same mechanism: a _merge_mcp_into_per_job_toolsets() helper layering globally-enabled MCP servers onto a per-job enabled_toolsets allowlist (no_mcp sentinel -> none; explicit MCP names -> allowlist; otherwise -> union all enabled). #32788 (open, 2026-05-26) is the earlier of the two. This PR additionally extracts a shared enabled_mcp_server_names() helper into hermes_cli/tools_config.py, but the core cron/scheduler.py fix is the same approach. Linking for reviewer dedup; not a knock on the work.