fix(agent): pop persisted marker on repair mutations

[aniruddhaadak80]

What

Alternation repair merges/prunes flushed message dicts in place but never popped the _db_persisted marker, so the next flush skipped the rewritten rows — the DB kept pre-merge bytes while the live transcript (resume, prompt cache) moved on. The merge paths already drop the api_content sidecar on the same condition; the persisted marker (documented contract: any in-place content mutation must pop it) now follows.

Related Issue

Code-scan find (no open issue covers repair persistence invalidation).

Type of Change

• Bug fix (non-breaking change that fixes an issue)

Changes Made

• agent/agent_runtime_helpers.py: pop _DB_PERSISTED_MARKER on actual mutation in _merge_assistant_into (tool_calls/content/reasoning changes only, mirroring the sidecar discipline), _prune_unanswered_tool_calls, and _merge_consecutive_users; invalidate agent._db_flush_scan_prefix when repair_message_sequence repairs (guarded for partial agents).
• tests/run_agent/test_repair_persistence_invalidation.py: user/assistant merges pop the marker and clear the prefix; clean runs leave both alone; prefix-less agents still repair.

How to Test

1. Unit: scripts/run_tests.sh tests/run_agent/test_repair_persistence_invalidation.py -q (4 passed; merge tests fail pre-fix).
2. ruff check clean on both touched files.

Checklist

Code

• Read the Contributing Guide
• Commit messages follow Conventional Commits (fix(scope):)
• Searched existing PRs for duplicates
• PR contains only changes related to this fix
• Tests added (behavioral contracts, not snapshots; red-green verified)

Documentation & Housekeeping

• No doc updates needed
• No config keys added/changed; no new env vars
• Cross-platform: pure dict logic, no OS surface

Screenshots / Logs

N/A.

[Enough1122]

AI code review — automated review for reference; please use your judgment.

Summary

Alternation repair (_merge_assistant_into, _prune_unanswered_tool_calls, _merge_consecutive_users) now pops the _DB_PERSISTED_MARKER whenever it mutates a flushed row, and repair_message_sequence resets the flush-scan prefix — so the next flush rewrites repaired rows instead of skipping them as "already persisted" (which left the DB holding pre-merge bytes).

Findings

• The mutated flag threading in _merge_assistant_into (agent/agent_runtime_helpers.py:9) is precise: pure no-op merges (e.g. empty tool_calls normalization that changes nothing) keep the marker and avoid a needless DB rewrite. The elif prev_calls: prev["tool_calls"] = prev_calls copy-normalization correctly does not set mutated.
• repair_message_sequence (:88) only resets _db_flush_scan_prefix when repairs > 0, guarded by hasattr — no behavior change on clean histories or marker-less agents.
• New test file pins all four legs (user merge, assistant merge, no-repair untouched, missing attr). Good.
• Non-blocking: the _DB_PERSISTED_MARKER import is function-local in three places (presumably to dodge a cycle). If the cycle allows, hoisting to module level would be cleaner — cosmetic only.

Verdict

Correct persistence-invariant fix. Non-blocking.