开发日志#

开发日志

v0.2.0 — 2026-04-20#

1. 论文Q&A — 多轮问答与工具调用#

写作阶段完成后，Paper 标签页新增了交互式Q&A对话面板。用户可以对生成的论文提问，理解论文内容，并辅助撰写审稿回复（rebuttal）。

工作原理：

  • PaperQAAgent 将完整LaTeX源码放入系统提示词（利用Prompt缓存实现高效多轮对话）

  • 工具调用支持：arxiv_search、semantic_scholar、web_search、latex_section_read——Agent可检索最新文献并定位论文具体章节，使回答有据可查

  • 每个问题最多进行5轮工具调用；回答实时流式输出至CLI和浏览器UI

  • Q&A历史持久化至会话目录下的 paper_qa_history.jsonl，浏览器刷新和会话恢复后均不丢失

新工具 latex_section_read：

Agent可通过名称读取论文任意章节，无需在每条消息中嵌入完整文档：

latex_section_read(section="定理2")
→ 返回从 \section{} 到下一个标题之间的内容

CLI交互示例：

Review the paper? [y/N]: y

Question (Enter to accept): 定理2的bound看起来比较松——O(n²) 是紧的吗？

  ✓ arxiv_search: "spectral gap tight bound"
  ✓ latex_section_read: section "定理2"

O(n²) 的bound来自Weyl不等式。[Chen et al. 2024] 指出，对于正则图，
利用Cauchy交错定理可将bound收紧至 O(n log n)。当前证明未利用正则性，
仍有改进空间。

What next?  [a]ccept  [q]uestion  [r]ewrite: q

新增文件： eurekaclaw/agents/paper_qa/agent.py（多轮 ask() 方法）、eurekaclaw/tools/latex_section.py、eurekaclaw/orchestrator/paper_qa_handler.py

2. 论文重写 — 基于Q&A反馈的定向修改#

用户现在可以直接在论文Q&A门控中触发完整的论文重写——浏览器UI和CLI均支持。重写操作将把修改指令注入为反馈，重新运行理论证明→实验验证→写作的完整流水线。

工作原理：

  • 用户在Q&A对话框中输入修改指令（例如："利用引理2.1中的Cauchy交错定理，将定理2的bound从O(n²)收紧至O(n log n)"）

  • POST /api/runs/{id}/rewrite 立即返回 202 Accepted——流水线在后台运行

  • Live 标签页实时显示理论→实验→写作的进度

  • 成功后：保存新版本论文（paper_v2.tex、paper_v3.tex……），Paper标签页自动刷新

  • 失败后：前一版本论文保持不变，对话框中出现错误标记

统一重写接口 —— 无论会话是否仍在运行（门控模式）还是已完成（历史会话），接口行为完全一致。彻底解决了历史会话重写中因代理超时（30秒）导致的 Failed to fetch 报错。

重写范围： 理论证明 → 实验验证（遵循 EXPERIMENT_MODE 配置）→ 写作。实验Agent自身的跳过逻辑（EXPERIMENT_MODE=false）依然有效，无需额外配置。

论文版本管理：

~/.eurekaclaw/runs/<session_id>/
├── paper_v1.tex              # 原始版本
├── paper_v2.tex              # 第一次重写后
├── paper_qa_history.jsonl    # 完整Q&A及重写记录日志
└── rewrite_context_v2.json   # 本次重写使用的Q&A摘要与修改指令

错误恢复：

失败场景

处理方式

理论Agent失败

回退至 paper_v{n-1}.tex；对话框出现错误标记

写作Agent失败

回退至 paper_v{n-1}.tex；对话框出现错误标记

重写已在进行中

返回 409 Conflict——UI显示提示；已有重写继续执行

会话状态缺失/损坏

返回 400 Bad Request——对话框出现错误标记；论文不变

修改文件： eurekaclaw/ui/server.py（新增 /rewrite 接口、后台任务、_ensure_bus_activated 辅助函数）、eurekaclaw/orchestrator/paper_qa_handler.py（_do_rewrite 扩展至实验阶段）、frontend/src/components/workspace/paper-review/usePaperSession.ts（简化状态机，统一 onRewrite）

2026-03-20 (continued — PRs #23, #28, #29, #30, #31, #33)#

8. Principled Prover Confidence Scoring#

The Prover now replaces purely format-based heuristics with a structured LLM self-assessment block appended to every proof response.

Prompt change: the system prompt instructs the LLM to append a \``json` block after the proof body:

{
  "confidence": 0.0-1.0,
  "completeness": "complete|partial|sketch",
  "gaps": ["...uncertain step descriptions..."],
  "weakest_step": "...",
  "techniques_used": ["..."]
}

Calibration guide embedded in the prompt:

Range

Meaning

≥ 0.95

Every step elementary or from a named theorem; no hand-waving

0.80–0.94

All key steps present; at most one routine calculation left implicit

0.60–0.79

Sketch correct but one non-trivial step not fully worked out

0.40–0.59

Main idea right but genuine gap flagged with [GAP: ...]

< 0.40

Incomplete proof or possibly wrong approach

_parse_proof_attempt() pipeline (6 steps):

  1. Extract inline [GAP: ...] tags

  2. Parse the structured JSON self-assessment block

  3. Heuristic fallback if no valid block (presence of "QED", "□", "this completes")

  4. Citation-integrity check: penalty −0.15 per uncited lemma ID (cap −0.30)

  5. Weasel-word penalty: −0.05 per "clearly"/"obviously"/etc. (cap −0.15)

  6. Strip JSON block from stored proof_text; extract optional Lean4 sketch

9. Verifier Bug Fix: Format Code Error#

Verifier._peer_review() was raising Unknown format code 'f' for object of type 'str' because settings.verifier_pass_confidence returned a str via Pydantic's env-var parsing. Fixed by explicit float() cast before :.2f formatting.

10. pdfTeX Font Expansion Fix#

! pdfTeX error (font expansion): auto expansion is only possible with scalable fonts caused by microtype's automatic font expansion being incompatible with lmss (Latin Modern Sans) on TeX Live 2022. Fixed by passing expansion=false to \RequirePackage[expansion=false]{microtype} in eureka.cls.

11. Token Limit UI Sliders (Extended)#

Five additional token-limit controls added to the Settings tab:

Setting

UI label

Default

MAX_TOKENS_ASSEMBLER

Assembler

4096

MAX_TOKENS_CRYSTALLIZER

Crystallizer

2500

MAX_TOKENS_ARCHITECT

Proof architect

3000

MAX_TOKENS_ANALYST

Analysis stages

1600

MAX_TOKENS_SKETCH

Proof sketch

600

server.py's _CONFIG_FIELDS mapping updated to expose all 12 token-limit knobs to the UI.

All hardcoded max_tokens=... literals across theory and learning agent files replaced with settings.max_tokens_* references.

12. Human Intervention for Empty Survey#

When the survey stage completes with zero papers found, the pipeline pauses and prompts the user:

⚠ Survey stage completed but found 0 papers.
Please provide a comma-separated list of paper IDs/titles to retry,
or press Enter to proceed without papers:

If the user provides titles, the survey task is re-injected with the paper list and re-executed once. Implemented in MetaOrchestrator._handle_empty_survey_fallback().

13. ProofCheckpoint Bug Fixes (Issue #27)#

Bug

Fix

cp.checkpoint_path AttributeError — public property was missing

Added @property checkpoint_path → self._checkpoint

cli.py resume used wrong meta key "research_brief" (dict)

Fixed to json.loads(meta.get("research_brief_json", "{}") or "{}")

cli.py resume used wrong attribute exc.paused_before_stage

Fixed to exc.stage_name

14. Minimax LLM Backend (PR #23)#

New backend shortcut LLM_BACKEND=minimax:

Variable

Description

MINIMAX_API_KEY

Minimax API key

MINIMAX_MODEL

Minimax model name (e.g. abab7-chat)

active_model / active_fast_model properties added to Config — resolve the correct model string for any backend (Anthropic, Minimax, OpenAI-compat). All agents and theory stages updated to use settings.active_model / settings.active_fast_model instead of hardcoded model names.

15. PDF Extraction via Docling (PR #23)#

PaperReader._extract_from_paper_pdf() fetches the full arXiv PDF via Docling, filters to theorem/lemma sections with a regex pass, then runs the LLM extractor over the rich excerpt instead of the abstract only.

Enable with the [pdf] optional extra:

pip install eurekaclaw[pdf]

16. Memory Relevance-Based Retrieval (PR #30)#

Tier 4 domain memories are now ranked by cosine similarity to the current task query instead of most-recently-written ordering.

  • _index.json now stores a "embedding" field per memory file (computed at write time using embedding_utils.get_embedding())

  • MemoryManager.load_for_injection(domain, k, query) accepts a query argument; if provided, ranks candidates by cosine similarity and returns the top-k most relevant files

  • New helper module: eurekaclaw/memory/embedding_utils.py (get_embedding(), cosine_similarity())

17. smile.sty Math Macro Package (PR #28)#

A comprehensive SMiLe-group math macro package (smile.sty) is now bundled with the template:

  • Blackboard bold shortcuts: \RR, \EE, \PP, \NN, \ZZ, \QQ, \CC, …

  • Calligraphic shortcuts: \cA … \cZ

  • Bold vectors (mathbf family): \xb, \Wb, … and (bm family): \bx, \bW, …

  • Bold Greek: \balpha, \bbeta, \bGamma, …

  • Norm/bracket macros: \norm{x}, \nbr{x}, \bignorm{x}, \opnorm{x}{p}

The writer agent system prompt is updated to list all available macros and instruct the LLM to use them instead of redefining equivalents.

18. OpenClaw Hub Integration (PR #31)#

Skills can now be installed directly from the ClawHub registry:

eurekaclaw install-skills <skillname>
# e.g. eurekaclaw install-skills steipete/github

install-skills with no argument continues to install all bundled seed skills. The clawhub CLI must be installed separately. The implementation lives in eurekaclaw/skills/from_hub.py.

19. Bug Fix: pause AttributeError (PR #33)#

cli.py referenced exc.paused_before_stage but ProofPausedException stores the field as stage_name. Fixed both call sites so pausing a run no longer crashes with AttributeError.

20. Direction Planning Fallback (PR #33)#

When DivergentConvergentPlanner.diverge() returns an empty list or throws an exception, the orchestrator now halts and prompts the user to enter a research direction manually instead of silently continuing with no direction. The survey's open problems are shown as context. Empty input or Ctrl+C raises RuntimeError and exits cleanly.

2026-03-20#

1. Theory Review Gate#

A new theory_review_gate orchestrator task is inserted between the TheoryAgent and WriterAgent in default_pipeline.yaml. It is always shown, regardless of --gate mode.

  • Displays a numbered lemma chain (L1, L2, …) with ✓ verified / ~ low confidence tags.

  • y / Enter → proceed to writer.

  • n → user specifies the most problematic step (by number or ID) and describes the issue. The TheoryAgent re-runs once with the feedback injected into its task description, then shows the updated sketch once more.

2. Pause / Resume#

Two new CLI commands and a graceful SIGINT handler:

Command

Description

eurekaclaw pause <session_id>

Write pause.flag; theory agent stops at next stage boundary and saves checkpoint.json

eurekaclaw resume <session_id>

Load checkpoint and continue from the saved stage

Ctrl+C during prove

Same as pause — writes flag instead of raising KeyboardInterrupt

Checkpoint file: ~/.eurekaclaw/sessions/<session_id>/checkpoint.json

3. ProofArchitect Improvements#

  • source=None crash fixed: item.get("source") or "" instead of item.get("source", "") so a JSON null value never passes through as None.

  • 3-layer fallback: full plan (4–10 lemmas) → 3-lemma simplified plan → single main_result. Single-goal is now truly a last resort.

4. Citation Quality Improvements#

  • Assembler prompt requires explicit [lemma_id] citations in the assembled proof (e.g., By [arm_pull_count_bound], …).

  • ConsistencyChecker receives the list of proved lemma IDs and verifies all appear in the proof; returns uncited_lemmas.

  • Retry routing: when ConsistencyChecker failure is citation-related ("uncited" / "missing citation"), the retry loop re-runs the Assembler as well as the Crystallizer — not just the Crystallizer.

5. TheoremCrystallizer Fixes#

  • max_tokens raised from 1500 → 2500 to prevent mid-expression truncation of LaTeX formulas.

  • Added explicit no-truncation constraint for math environments.

6. Knowledge Graph Write Timing#

Tier 3 KG writes now trigger whenever proven_count > 0 (i.e., after any lemma is proved), not only when the full consistency check passes. Lemma nodes and dependency edges are preserved even when crystallization fails.

7. Bug Fixes#

File

Bug

Fix

agents/theory/inner_loop_yaml.py

cp.delete() AttributeError at checkpoint clear

cp.delete() → cp.clear()

agents/theory/inner_loop_yaml.py

ProofPausedException swallowed by bare except Exception

Added isinstance(e, ProofPausedException): raise before handler

agents/theory/agent.py

ProofPausedException swallowed in agent execute

Added explicit re-raise

orchestrator/meta_orchestrator.py

_init_brief created a new session UUID, mismatching pause.flag path

session_id reuses bus.session_id

2026-03-19#

1. Robust Lemma Decomposer Parsing#

_parse_lemmas in agents/theory/decomposer.py now uses a 4-pass extraction strategy instead of 2, preventing the "Empty lemma list from decomposer" fallback in most cases:

Pass

Strategy

1

JSON inside ```json ``` or plain ``` ``` code fence

2

First JSON object {...} — checks 7 key names: lemmas, steps, subgoals, proof_steps, lemma_list, components, parts

3

First JSON array [...] — accepted directly as lemma list

4

Plain-text numbered/bulleted list heuristic

_normalize_list accepts flexible field names per item (id/lemma_id/name/title, statement/formal_statement/hypothesis/content, etc.) so variant LLM output schemas are handled without falling back to single-theorem mode.

The same 4-pass strategy was also applied to ProofArchitect._parse_lemmas.

2. UI Polling Log Suppression#

GET /api/runs/<id> 200 status-poll requests are now logged at DEBUG level instead of INFO, removing repetitive log noise during long runs.

3. Bug Fixes#

File

Bug

Fix

agents/survey/agent.py

ValueError: substring not found on unclosed ```json block

Wrapped text.index in try/except

agents/base.py

run_agent_loop ignoring SURVEY_MAX_TURNS setting

Uses dynamic AsyncRetrying now

main.py

NameError: name 'Path' is not defined in save_artifacts

Added from pathlib import Path

ui/server.py

GET /api/runs/... spamming the log

Demoted to DEBUG for 200 polling responses

4. Always-On Stage Summary Cards#

orchestrator/gate.py now prints a Rich summary card after every completed pipeline stage regardless of GATE_MODE:

Stage

Card shows

survey

Papers found, open problems, key mathematical objects

theory

Proof status, per-lemma breakdown with confidence tags

experiment

Alignment score, per-lemma numerical check results

writer

Full session summary before final output

5. Human Gate Improvements#

  • Text feedback input: after approving a gate, users can type a correction or hint injected into the next agent's task via get_user_feedback()

  • Auto-escalation: if ≥1 lemma has verified=False after theory stage, gate auto-escalates from auto to human with full lemma confidence breakdown

  • Default changed: GATE_MODE default changed from none to auto

6. Proof Readability Enforcement (WriterAgent)#

ENFORCE_PROOF_STYLE=true (default) injects _PROOF_STYLE_RULES into writer prompts:

  • No skip words ("clearly", "trivially", "by standard arguments") without immediate justification

  • Every inequality must name the lemma or theorem it uses

  • Each lemma proof opens with 1–2 sentences of informal explanation

  • Low-confidence lemmas tagged \textcolor{orange}{\textbf{[Unverified step]}} in PDF

  • Limitations section explains all unverified steps

  • Added \usepackage{xcolor} to LATEX_PREAMBLE

7. Targeted Numerical Testing (ExperimentAgent)#

ExperimentAgent now separates proven lemmas into verified and low_confidence groups:

  • For each low-confidence lemma: runs dedicated numerical test, computes violation_rate

  • Lemmas with violation_rate > 1% flagged as numerically_suspect

  • Experiment stage summary card shows per-lemma check results with color coding

  • WriterAgent adds stronger warnings for numerically_suspect lemmas

8. CLI Default Output Directory#

All three CLI commands (prove, explore, from-papers) now default to ./results if --output is not specified.

9. Bibliography & Citation Fix#

Removed \cite{} → ? bug in output PDFs:

  • _fix_missing_citations() surgically removes \cite{} keys that have no matching entry in references.bib

  • LATEX_END split into LATEX_END_WITH_BIB / LATEX_END_NO_BIB — \bibliography{references} only added when .bib is non-empty

  • WriterAgent prompt instructs LLM to use exact cite keys (same as _generate_bibtex)

10. LaTeX Extraction Improvements#

Extended _extract_latex in WriterAgent with:

  • Markdown heading → LaTeX section conversion

  • tikzpicture environment removal

  • Environment name normalization additions: rem→remark, rema→remark, prop→proposition, defin→definition, corolary→corollary, Cor→corollary, Thm→theorem, Lem→lemma

  • \endproof → \end{proof} substitution

  • QED box removal (\begin{flushright}$\square$\end{flushright})

  • Two-pass _close_open_environments: (1) remove orphaned \end{X}, (2) append missing \end{X}

2026-03-18#

1. Context Compression#

Optimization summary:

Stage

Before

After

Saving

Formalizer model

opus

haiku

~90% cost per call

Formalizer re-run

Every iteration

Only on change

Saves N-1 calls

Proven lemma context

200-char proof text

Statement only (120)

~40% input tokens

Verifier (high-conf)

Always LLM call

Auto-accept ≥ 0.85

~30% fewer calls

Verifier proof text

Raw (up to 3000 chars)

Head+tail (1000 chars)

~67% input tokens

Agent loop history

Accumulates forever

Compressed every 6 turns

~60% input tokens

Stagnation

Wastes 3+ iterations

Forced refinement

Saves full iterations

Run performance: typical run now ~20 LLM calls (down from ~35); worst case ~55 (down from ~100).

Agent max_turns reductions:

Agent

Before

After

SurveyAgent

15

8

IdeationAgent

5

3

ExperimentAgent

10

5

WriterAgent

5

3

New config knobs: CONTEXT_COMPRESS_AFTER_TURNS, AUTO_VERIFY_CONFIDENCE, STAGNATION_WINDOW

2. Configurable Token Limits Per Call Type#

7 new .env variables:

Variable

Default

Applies to

MAX_TOKENS_AGENT

8192

All agent reasoning loops

MAX_TOKENS_PROVER

4096

Proof generation

MAX_TOKENS_PLANNER

4096

Research direction planning

MAX_TOKENS_DECOMPOSER

2048

Lemma decomposition

MAX_TOKENS_FORMALIZER

2048

Formalization, refiner, counterexample

MAX_TOKENS_VERIFIER

1024

Proof verification

MAX_TOKENS_COMPRESS

512

Context compression

UI sliders added in the Settings tab for all 7 values.

3. Multi-Backend LLM Support#

Three named backends in config.py and llm/factory.py:

Backend

LLM_BACKEND=

Notes

Anthropic native

anthropic

Default

OpenRouter

openrouter

Set OPENAI_COMPAT_API_KEY=sk-or-...

Local (vLLM/Ollama)

local

Defaults to http://localhost:8000/v1

ccproxy / OAuth fallback: if ANTHROPIC_API_KEY is empty, automatically reads ~/.claude/.credentials.json and routes through ccproxy (allows Claude Pro/Max users to run without a separate API key).

4. Additional Tuning Knobs#

Variable

Default

Effect

SURVEY_MAX_TURNS

8

Tool-use turns in SurveyAgent

THEORY_STAGE_MAX_TURNS

6

Turns per theory stage

WRITER_MAX_TURNS

4

Turns for paper generation

ARXIV_MAX_RESULTS

10

Hard cap on arXiv results

LLM_RETRY_ATTEMPTS

5

Retry attempts on 5xx / rate-limit errors

LLM_RETRY_WAIT_MIN / MAX

4 / 90

Exponential backoff bounds

5. Domain Plugin Architecture#

New three-tier plugin system:

EurekaClaw (general pipeline)
    └── DomainPlugin (e.g. MAB)
            └── Workflow (per-domain prompt guidance)

New files:

  • domains/base.py — DomainPlugin ABC with 4 methods

  • domains/__init__.py — @register_domain decorator + resolve_domain()

  • domains/mab/ — MAB domain plugin (GaussianBandit, BernoulliBandit, 4 tools, 4 skills, 3-level benchmark)

Core changes:

  • SkillRegistry.add_skills_dir() — load skills from domain directories

  • build_default_registry() — now domain-agnostic; domain tools registered via plugin

  • MetaOrchestrator — accepts domain_plugin parameter

  • EurekaSession.run() — auto-detects domain plugin from InputSpec.domain

6. LaTeX Compilation Robustness#

  • Added 7 \newtheorem declarations to LATEX_PREAMBLE: assumption, maintheorem, conjecture, claim, example, fact, observation

  • Environment name normalization in _extract_latex (step 6): mis-cased names corrected

  • _close_open_environments() — new static method, stack-based, prevents \begin{tabular} truncation from causing fatal errors

  • Removed _rescue_compile and paper_rescue.tex fallback entirely

  • Full bibtex-aware compile sequence: pdflatex → bibtex → pdflatex → pdflatex

7. Bibliography & Reference Resolution#

  • Write order fixed: references.bib saved before _compile_pdf() is called

  • bibtex added to compile sequence: previously only pdflatex ran twice

  • Cite key consistency: _compute_cite_keys() in WriterAgent uses identical algorithm to _generate_bibtex in main.py

  • Duplicate key handling: _generate_bibtex deduplicates with a, b, … suffixes

  • ResearchOutput.bibliography_json field added

8. Experiment Mode Control#

New EXPERIMENT_MODE env var:

  • auto — skip experiment if formal statement has no quantitative signals

  • true — always run experiment stage

  • false — always skip experiment stage

当前页面
由 GitHub 编辑