fix: bound reflexion and judge prompts

[SkyeAv]

Hard-bounds the reflexion and judge prompts so a pathological article cannot outgrow the model's context window. Fleet distill telemetry: the three largest of 534 LLM calls were all reflexion prompts of 1.6M and 2.3M characters (689,241 and 524,336 input tokens) — past every model's window, failing outright after paying for the serialization. Cause: llm_propose_config_edit interpolated the full coverage report and context with no cap; _build_judge_prompt interpolated the raw build report.

Bounds

• agent.compact_coverage_report (pure, total, non-mutating): caps every reachable unresolved list (top level, sections[i], sections[i].per_column[col], back-compat top-level per_column) at the first 20 terms (UNRESOLVED_CAP) plus one visible +N more marker — the same convention as compact_audit_report — and adds unresolved_count siblings naming the ORIGINAL lengths so the scale signal survives. Beyond max_chars, per-column unresolved lists are dropped entirely (counts/fractions retained); the serialized block is bounded by the caller, never by parsing truncated JSON back.
• agent._truncate_for_prompt(text, limit, *, what) with a verbatim …[X truncated: N of M chars omitted] marker; limit <= 0 raises.
• Budgets: config 8,000 (MAX_PROMPT_CONFIG_CHARS) + coverage 8,000 (COVERAGE_PROMPT_CHARS) + context 40,000 (MAX_PROMPT_CONTEXT_CHARS) + report 8,000 (MAX_PROMPT_REPORT_CHARS) ≈ worst case 58 K chars ≈ 15 K tokens — a ≥ 27× reduction, room left for the ~16 K-token output.
• Public APIs unchanged: map_coverage, _measure_section, build_and_audit, and the supervisor's tier-1 deterministic proposer still receive full uncapped reports; compaction happens only at the tier-2 prompt-serialization boundary. Never-raises contract and the validate_table_config gate preserved.

Testing

• uv run pytest tests/test_agent_prompt_bounds.py -q --no-cov -n 0 → 13 passed in 1.16s: caps-every-list, purity/non-mutation, totality on hostile shapes, verbatim markers, unresolved_count, 50,000-term regression (< 60 K prompt, 3 sections × 2 columns), overflow drop path, 400-section survival, judge compaction (kgx_path absent, both reverse modes), determinism
• Full gate on the stack tip: uv run pytest -q → 1565 passed, 15 skipped; ruff/format/pyright clean; CI-equivalent base env passes
• Tier-2 review caught and fixed a Blocker in the first round: the overflow path parsed a mid-JSON-truncated string back (json.loads always raised, silently disabling reflexion via the caller's blanket except); the parse-back pass was deleted and pinned by overflow tests, re-reviewed PASS (31/31 adversarial checks)

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 9d791768-fd3e-4266-8de7-78853faccf92

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}