fix(run_agent): isolate background review fork from external memory plugins

[teknium1]

Summary

Background memory/skill review forks no longer leak harness prompts into external memory plugin namespaces (honcho, mem0, supermemory, etc.).

Reported by @utku — the review harness prompts (Review the conversation above and update the skill library...) were getting ingested by memory plugins as if they were real user conversation.

Root cause

_spawn_background_review() creates a fresh AIAgent to do the review. That __init__ rebuilds its own _memory_manager from memory.provider in config, scoped to the parent's session_id. Then run_conversation() triggers three ingestion sites against the user's real memory namespace:

Site	What gets written
on_turn_start(turn_count, prompt)	harness prompt as the turn's user message; advances cadence
prefetch_all(prompt)	harness prompt as a recall query
sync_all(prompt, review_output, session_id)	harness prompt + review output as a (user, assistant) pair

Changes

• run_agent.py _spawn_background_review() — pass skip_memory=True to the fork's AIAgent(...) constructor
• tests/run_agent/test_background_review.py — regression test asserting the kwarg is set

Why this works

skip_memory=True short-circuits both _memory_store and _memory_manager construction in __init__. The existing fork-setup block at L4310-4314 explicitly rebinds _memory_store, _memory_enabled, _user_profile_enabled, and the nudge intervals from the parent right after construction — so:

• Built-in MEMORY.md / USER.md writes via the memory tool still land on disk ✓
• Skill writes still work (skills don't go through MemoryManager at all) ✓
• All three ingestion sites short-circuit on if self._memory_manager: ✓
• shutdown_memory_provider() on the fork becomes a no-op (nothing to shut down) ✓
• Cached system prompt is still inherited verbatim from parent (prefix cache parity preserved) ✓

Validation

	Before	After
Fork _memory_manager (with honcho configured)	non-None, providers=['honcho']	None
Fork _memory_store after rebinding	parent's store object	parent's store object (identity preserved)
Fork _memory_enabled after rebinding	True	True
on_turn_start / prefetch_all / sync_all gates	all evaluated	all short-circuit
Targeted background-review tests	28/28 pass	28/28 pass

E2E verified with a stubbed honcho provider in an isolated HERMES_HOME: with skip_memory=True the fork constructs with _memory_manager=None even though memory.provider=honcho is set in config; without it, the fork builds a manager wired to the stub.

Test plan

scripts/run_tests.sh tests/run_agent/test_background_review.py tests/run_agent/test_background_review_cache_parity.py tests/run_agent/test_background_review_toolset_restriction.py tests/run_agent/test_review_prompt_class_first.py — 28 passed.

[github-actions]

🔎 Lint report: hermes/hermes-48113ebb vs origin/main

ruff

Total: 1 on HEAD, 1 on base (➖ 0)

🆕 New issues (1):

Rule	Count
PLW1514	1

First entries

hermes_cli/send_cmd.py:61: [PLW1514] `pathlib.Path(...).read_text` without explicit `encoding` argument

✅ Fixed issues (1):

Rule	Count
PLW1514	1

First entries

../../../../../tmp/lint-base/hermes_cli/send_cmd.py:61: [PLW1514] `pathlib.Path(...).read_text` without explicit `encoding` argument

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 8350 on HEAD, 8350 on base (➖ 0)

🆕 New issues (3):

Rule	Count
invalid-argument-type	3

First entries

run_agent.py:14073: [invalid-argument-type] invalid-argument-type: Argument to function `_is_oauth_token` is incorrect: Expected `str`, found `str | dict[Unknown | str, Unknown | str | dict[str, str]] | Any | ... omitted 3 union elements`
run_agent.py:14076: [invalid-argument-type] invalid-argument-type: Argument to function `len` is incorrect: Expected `Sized`, found `(str & ~AlwaysFalsy) | (dict[Unknown | str, Unknown | str | dict[str, str]] & ~AlwaysFalsy) | (Any & ~AlwaysFalsy) | ... omitted 3 union elements`
run_agent.py:7770: [invalid-argument-type] invalid-argument-type: Argument to function `build_anthropic_client` is incorrect: Expected `str`, found `str | dict[Unknown | str, Unknown | str | dict[str, str]] | Any | ... omitted 3 union elements`

✅ Fixed issues (3):

Rule	Count
invalid-argument-type	3

First entries

run_agent.py:7754: [invalid-argument-type] invalid-argument-type: Argument to function `build_anthropic_client` is incorrect: Expected `str`, found `str | dict[Unknown, Unknown] | Any | ... omitted 3 union elements`
run_agent.py:14057: [invalid-argument-type] invalid-argument-type: Argument to function `_is_oauth_token` is incorrect: Expected `str`, found `str | dict[Unknown, Unknown] | Any | ... omitted 3 union elements`
run_agent.py:14060: [invalid-argument-type] invalid-argument-type: Argument to function `len` is incorrect: Expected `Sized`, found `(str & ~AlwaysFalsy) | (dict[Unknown, Unknown] & ~AlwaysFalsy) | (Any & ~AlwaysFalsy) | ... omitted 3 union elements`

Unchanged: 4364 pre-existing issues carried over.

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

[alt-glitch]

Competing fix: this PR and #24281 (reopen of #21511) both address #21510 — background review fork leaking harness prompts into external memory plugins. This PR takes a cleaner approach (skip_memory=True with detailed comments + regression test) while #24281 also adds defense-in-depth in the Honcho plugin itself.