fix: add write-hygiene guardrails to skill review prompts (prevent SKILL.md bloat)

[gejifeng]

Problem

The background skill-review agent continuously inflates SKILL.md files across sessions. Root causes:

1. Completion bias — _SKILL_REVIEW_PROMPT said "most sessions produce at least one skill update", so the model treats no-op as a failure.
2. No read-before-write requirement — the agent never calls skill_view first, so it cannot detect duplicate content.
3. Date stamps — the model self-injects temporal markers (Updated 2026-05-10, NEW — Session ...) that accumulate across sessions.

Related upstream discussions: #12877, #20595, #22620.

---

Fix — two-layer defence

Layer 1 — Prompt-level (run_agent.py)

In both _SKILL_REVIEW_PROMPT and _COMBINED_REVIEW_PROMPT:

• Removed "Be ACTIVE — most sessions produce at least one skill update" (the completion-bias trigger).
• Added WRITE HYGIENE rules: no date stamps, no duplicate sections, read before write, prefer patch over edit, treat nothing to save as the correct default.

Layer 2 — Code-level hard rejection (tools/skill_manager_tool.py)

Added _run_write_hygiene() called by _edit_skill() and _patch_skill() before any bytes hit disk. It contains three independent hard-rejection checks that return tool-call errors, forcing the model to fix the content rather than ignoring prompt instructions:

Check	What it catches	Where
_check_date_stamps()	Updated 2026-05-10, Added 2024-11, <!-- session … --> etc.	Both edit and patch
_check_duplicate_headings()	Same ## heading appears twice in new content	Both edit and patch
_check_content_growth()	Single write more than doubles file size (>100% + >500 chars)	edit only

The growth guard is disabled for patch (targeted by design); date-stamp and duplicate-heading checks apply to both.

---

Why prompt-only is insufficient

As noted in #20595: "Completion bias is a cognitive failure mode, not a prompt-attention failure." The model will comply with soft wording when convenient and ignore it when under pressure to produce output. Code-level checks are the only way to guarantee enforcement.

---

Testing

# All pass:
assert _check_date_stamps("Updated 2026-05-10 rule")   # blocked
assert _check_date_stamps("## Style Rules") is None      # allowed
assert _check_date_stamps("<!-- session 2025-03-01 -->") # blocked
assert _check_date_stamps("Version 3.2.1 info") is None  # allowed
assert _check_duplicate_headings("## Foo\n\n## Foo\n")   # blocked
assert _check_duplicate_headings("## Foo\n\n## Bar\n") is None  # allowed
assert _check_content_growth("x" * 2001, "x" * 1000)    # blocked (200%)
assert _check_content_growth("x" * 1800, "x" * 1000) is None  # allowed (180%)

[teknium1]

Thank you for identifying the completion-bias issue; it remains present on current main in agent/background_review.py:171-175 and :282-284.

Problems

• The prompt source moved: run_agent.py:1576-1582 imports these constants from agent.background_review, so the run_agent.py portion needs to be ported rather than applied directly.
• The proposed hard checks run for all create/edit/patch callers. Current main deliberately distinguishes autonomous review from foreground user-directed edits in tools/skill_manager_tool.py:297-312; please avoid applying heuristic date/heading/growth rejections to foreground writes.
• The PR replaces tests/tools/test_skill_manager_tool.py with +62/-907. Please retain the existing unrelated regression coverage and add focused tests instead.

Suggested changes

• Port the prompt change to agent/background_review.py.
• Build on the existing background-review read-before-write guard at tools/skill_manager_tool.py:382-409 (commit 20871c1d941a697723aae24cd8e88afe701793ae) and scope any new autonomous-only enforcement accordingly.

Automated hermes-sweeper review.

[teknium1]

Please retain the existing test module and add focused hygiene cases to it. This rewrite removes 907 lines of unrelated CRUD, validation, path-safety, external-directory, dispatcher, and pinned-skill coverage while adding only 62 lines.

[gejifeng]

Addressed the latest review and rebased the PR onto current main.

Changes:

• moved the prompt updates to agent/background_review.py;
• scoped date-stamp, duplicate-heading, and suspicious-growth checks to autonomous background-review writes only;
• layered the checks after the existing exact-file read-before-write guard for edit/patch;
• left foreground user-directed create/edit/patch behavior unchanged;
• retained the full existing skill-manager test module and added focused background/foreground coverage;
• updated the existing prompt behavior tests for signal-gated reviews and write hygiene.

Local verification:

• scripts/run_tests.sh tests/tools/test_skill_manager_tool.py tests/run_agent/test_review_prompt_class_first.py -q (133 passed)
• background-review regression suite (51 passed)
• ruff check and git diff --check passed.