fix(agent): honor prompt_caching.cache_ttl=off on MoA and fallback stub paths

[686f6c61]

What does this PR do?

prompt_caching.cache_ttl: off sets _cache_disabled on the real AIAgent in init_agent. Main-loop requests honor that flag. MoA advisor/aggregator decoration and the shared plan_cache_sections_for_destination() helper do not: they resolve policy against a blank SimpleNamespace stub, so getattr(stub, "_cache_disabled", False) is always false and cache_control markers still get injected.

That breaks the documented contract from #33555 that the config disable is global. Operators who turn caching off (providers that mishandle markers, OAuth extra-usage concerns, debugging) still saw markers on:

• MoA aggregator requests
• MoA advisor requests on cache-honoring slots
• Auxiliary / MoA destinations planned through plan_cache_sections_for_destination

Fix

1. Add prompt_caching_disabled_from_config() with the same TTL disable detection as init_agent.
2. Stamp _cache_disabled onto the policy stub in:
   • plan_cache_sections_for_destination() (optional cache_disabled param; defaults to config)
   • _maybe_apply_moa_cache_control() (same)
3. MoA acting-aggregator path passes getattr(self._agent, "_cache_disabled", False) so a live agent disable wins when available.

Caching-enabled routes are unchanged: with cache_disabled=False / default TTL, native Anthropic destinations still receive breakpoints.

Related Issue

Fixes #76085

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• Tests (adding or improving test coverage)

Changes Made

• agent/agent_runtime_helpers.py — config helper + stub flag in plan_cache_sections_for_destination
• agent/moa_loop.py — stub flag in _maybe_apply_moa_cache_control; pass agent flag from prepared aggregator
• tests/agent/test_cache_disabled_on_stubs.py — config detection, plan-path strip, MoA decoration strip, enabled path still marks

How to Test

pytest tests/agent/test_cache_disabled_on_stubs.py -v

Docker (ghcr.io/astral-sh/uv:python3.12-bookworm-slim + uv pip install -e ".[dev]"): 7/7 passed.

Manual:

1. Set in config:

   prompt_caching:
     cache_ttl: false

2. Run an MoA turn or a path that hits auxiliary fallback replan against a Claude destination
3. Confirm request payloads no longer carry cache_control markers

Checklist

Code

• Contributing Guide
• Conventional Commits (fix(agent): …)
• Searched for open PRs on prompt_caching.cache_ttl disable is bypassed on stub-resolved cache-plan paths (MoA aggregator/advisors, auxiliary fallbacks) #76085 — none
• Focused change
• Tests added and passing
• Linux / Docker

Documentation & Housekeeping

• Docs — N/A (restores existing documented disable)
• Config example — N/A
• Cross-platform — N/A
• Tool schemas — N/A

Notes

I defaulted the stub paths to read config when the caller does not pass an explicit flag so auxiliary fallback does not need a full agent object. When MoA has self._agent, the agent flag is still threaded for consistency with runtime state after init.

[686f6c61]

Rebased onto current main and force-pushed.

Follow-up from review hygiene elsewhere:

• dropped an unused pytest import (F401 risk)
• tests call production helpers (plan_cache_sections_for_destination, _maybe_apply_moa_cache_control) with config/stubs — no source-text pins

Tip is e16781f. Happy for review.

[686f6c61]

Note for maintainers: #76121 opened later against the same issue (#76085) with a similar approach (stamp _cache_disabled onto the blank policy stubs for MoA + fallback replan).

This PR was first (opened ~11:07 UTC; #76121 ~11:35 UTC). Happy to fold any extra coverage from #76121 (e.g. auxiliary client tests) into this branch if useful, or to stand down if you prefer consolidating there — just say which to keep.

[teknium1]

Thanks for tracing the blank-stub policy paths; the current-main premise is real: agent/agent_runtime_helpers.py:1889 and agent/moa_loop.py:409 both pass a stub without _cache_disabled to the policy.

Problems

• agent/moa_loop.py:1698 accesses self._agent directly and coerces an unavailable agent to False. The existing prepared-aggregator test builds the facade via MoAChatCompletions.__new__ without setting _agent (tests/agent/test_moa_aggregator_cache_control.py:133-153), so this raises inside the planner try, skips planning, and breaks its expected tool marker. An explicitly None agent would also suppress the new planner's config fallback.

Suggested changes

• Preserve a tri-state value here: use getattr(self, "_agent", None) and pass None when no live _cache_disabled snapshot exists, allowing plan_cache_sections_for_destination() to read config as intended.
• Add a regression for the prepared-aggregator no-agent/config-off path.

Automated hermes-sweeper review.

[686f6c61]

Thanks — addressed on the latest tip.

• Prepared-aggregator path now uses getattr(self, "_agent", None) and passes a tri-state cache_disabled (None when no live agent / no _cache_disabled snapshot) so plan_cache_sections_for_destination() can still read config instead of raising or forcing False.
• Added a regression: MoAChatCompletions.__new__ facade with no _agent + prompt_caching.cache_ttl=off must not inject cache_control.

Docker: test_cache_disabled_on_stubs.py + test_moa_aggregator_cache_control.py 11/11 passed.

[686f6c61]

@JoaoMarcos44 thanks for the careful side-by-side — that is exactly the coordination I was hoping for.

I ported the pieces from #76121 that were not yet here onto this tip:

1. blank_cache_policy_stub() — single factory used by both plan_cache_sections_for_destination and _maybe_apply_moa_cache_control, so _cache_disabled cannot be left off a hand-rolled stub again.
2. Advisor + one-shot synthesis threading — _run_references_parallel / _run_reference and aggregate_moa_context now pin the live agent disable onto decoration instead of only re-reading config.
3. Kept our earlier tri-state no-agent prepared-aggregator fix and its regression.
4. Added factory / synthesis / advisor-runtime regressions in tests/agent/test_cache_disabled_on_stubs.py.

Docker: test_cache_disabled_on_stubs.py + test_moa_aggregator_cache_control.py 15/15 passed.

If maintainers are happy merging this branch as the single path, you can close #76121 as superseded; if they prefer yours as the merge target I am fine either way. Thanks again for the comparison.

[JoaoMarcos44]

Closed #76121 as superseded — verified your ported changes match (pulled the branch, diffed, reran the tests locally, 15/15 + 193 passed on the broader suite). Appreciate you folding those in cleanly.

One ask: since 15599816 absorbs the blank_cache_policy_stub factory and the advisor/one-shot-synthesis threading from #76121, could you add a Co-authored-by: trailer to that commit (or a follow-up commit) crediting the source?

Co-authored-by: JoaoMarcos44 <87440198+JoaoMarcos44@users.noreply.github.com>

That's the standard way GitHub attributes a merged commit to both authors. Thanks again for coordinating on this.

[686f6c61]

@JoaoMarcos44 done — amended the consolidate commit with your trailer and force-pushed.

Tip is now ebf54b1 (same tree as 15599816, with):

Co-authored-by: JoaoMarcos44 <87440198+JoaoMarcos44@users.noreply.github.com>

Thanks again for closing #76121 as superseded and for the careful verification.

[kshitijk4poor]

Merged via #76622 — your four commits were cherry-picked onto current main with authorship preserved (rebase merge, no squash), including the Co-authored-by credit for @JoaoMarcos44 from the #76121 consolidation.

Commits on main: 8a29703, 67db870, 8e1a351, 8ee5174 (+ 88a629b, a small follow-up from review: shared cache_ttl_means_disabled() predicate so agent_init and the stub paths can't drift, consistent injection guard in aggregate_moa_context, and a strengthened test assertion).

Review verification before merge: caching-enabled output verified byte-identical to main across native Anthropic / OpenRouter Claude / non-caching routes (the disable now strips markers on all stub paths), plus a mutation check proving the new tests fail when the stub flag is defeated. Solid fix and excellent coordination with the duplicate PR — thanks!

[686f6c61]

Thanks @kshitijk4poor — glad the four commits landed via #76622 with authorship preserved, and the follow-up (cache_ttl_means_disabled() shared predicate + stronger tests) is a solid hardening of the disable path. Appreciate the careful verification.