refactor(context_compressor): extract message marker helpers

[andrexibiza]

Summary

Extract the three message-marker helpers and _DB_PERSISTED_MARKER from agent/context_compressor.py into the leaf module agent/context_compressor_message_helpers.py.

The original module re-exports every moved name, preserving import compatibility and the established monkeypatch surface used by ContextCompressor.compress().

Exact scope

• Feature Package pin: f15a38ee73631b3cd5f7d30765c37d5f0245d403
• Original window: agent/context_compressor.py:173-237
• Moved bytes: 3,204
• SHA-256: 8c6860876e9a511927f9eecffce5bb878179d590218df284edd48c3cf3adc31d
• Moved names:
• _fresh_compaction_message_copy
• _template_visible_role
• _strip_persistence_markers
• _DB_PERSISTED_MARKER

The complete owner file remained byte-identical from the Feature Package pin through rebased base df7e17841959d08254d923869a1060005680d4a2, so the approved seam did not drift.

Compatibility and behavior

• agent.context_compressor still exposes the three helpers and marker constant.
• Re-exported helper objects are identical to the leaf-module definitions.
• Runtime tests prove a monkeypatch applied to agent.context_compressor._fresh_compaction_message_copy controls the real ContextCompressor.compress() path.
• Fresh outer-message copies, nested-data identity, role-template behavior, and persistence-marker stripping are preserved.
• AST checks reject duplicate definitions in the original module.
• The leaf module has no import edge back to agent.context_compressor.

Verification

Post-rebase at commit 95900a9fc4752232358c578869e753cd08b5d785:

• focused compressor/seam battery: 180 passed
• Ruff check: passed
• Ruff format check: passed
• git diff --check: passed
• clean-pin application checks: all five independent lane patches applied
• exact moved-byte and AST ownership receipt: passed

The five-lane implementation wave produced four exact-byte clean extractions. Two independent lanes converged on the accepted leaf-module name; one divergent implementation failed the Golden-SHA gate and was excluded. Stronger compatible assertions from the lane set were consolidated into the committed seam suite.

Feature Package status

This is one coherent shard toward the 2,000-line law. agent/context_compressor.py remains above 2,000 lines; this PR does not claim the final god-file .

Part of #78645

Part of #78647

🔒 : #78645 · indexed by the All Gods tracker #78647.

Related interlock closure

• Current summary-kernel shard: refactor(context_compressor): extract summary kernel (LB8) #81105 / refactor(context_compressor): extract summary kernel and preserve compatibility #81181
• Prior compressor shard PRs: refactor(context_compressor): extract text-util leaves into context_compressor_text_utils #80626 refactor(context_compressor): extract skill-prune/ghost-skill cluster (LB2) #80628 refactor(context_compressor): extract token-budget estimators into context_compressor_budget (LB3) #80634 refactor(context_compressor): extract content/media strip helpers (LB4) #80636 refactor(context_compressor): extract threshold coercion mixin (LB7) #80644 refactor(context_compressor): extract tool-result summarizers (LB5a) #80645 refactor(context_compressor): extract durable session guards mixin (LB6) #81074
• Related collider/base-drift PRs: fix(agent,cli): case-insensitive context length lookup, multimodal token estimation, /q alias #6648 fix(context_compressor): harden tool-call previews in summaries #12588 Fix long-context compression with hierarchical summaries #25716 feat(context_compressor): add compression efficiency ratio logging #29859 fix(compression): attribute auxiliary failures to the actual wire route #72637 fix: add usedforsecurity=False to hashlib.md5 in context_compressor (#FIPS) #74632
• Referenced issue nodes: [Bug] context compaction summaries still emit function-call-shaped pseudo tool calls with truncated arguments #12586 fix(agent): error display attributes auxiliary compression 401 to main model #72636 fix: hashlib.md5 FIPS crash in context_compressor.py #56715 fix: hashlib.sha1 FIPS crash in codex_responses_adapter.py #56716 fix: hashlib FIPS crash in gateway and tools modules #56719

Every listed node is preserved for coordination; this PR does not duplicate their extracted windows.

[Enough1122]

AI code review — automated review for reference, author can ignore or act on any point.

refactor(context_compressor): extract message marker helpers

1. The byte-hash assertions are a change-detector test — tests/agent/test_context_compressor_message_helpers_seam.py test_leaf_source_is_cycle_free_exact_move_and_only_definition_owner (~L218-251): len(approved) == EXPECTED_BYTES and sha256(...) == EXPECTED_SHA256 over the raw source bytes of the extracted functions. Any docstring edit, comment tweak, or whitespace change in context_compressor_message_helpers.py breaks CI for zero behavioral reason — exactly the source-reading / change-detector pattern the repo's AGENTS.md bans ("Never read source code in tests", "Don't write change-detector tests"). Recommend dropping the byte-length/SHA assertions and keeping the identity, behavior, and patch-authority checks, which already fully cover the seam.
2. helper_bytes.index(b"def _fresh_compaction_message_copy") (~L249) is position-based slicing of the file — if a docstring or comment above that def ever contains the same byte sequence, the slice starts early and the hash trivially fails. ast-based function-range extraction (already used elsewhere in the same test for the cycle-free check) would be robust — or simply drop the byte checks per point 1.
3. The rest of the extraction is exemplary — identity re-export checks in both directions, import-order cycle-freeness in fresh interpreters, patch-authority preservation on agent.context_compressor, a parametrized _template_visible_role contract test, and a compress() integration test with a leaking-copy mock proving the terminal sweep invariant.