feat(p2c): persist CodeMemory experiences to SQLite. Related to #162 by wen-placeholder · Pull Request #225 · jerry609/PaperBot

wen-placeholder · 2026-03-03T22:08:38Z

Closes #162

vercel · 2026-03-03T22:08:44Z

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
paper-bot	Ready	Preview, Comment	Mar 3, 2026 10:08pm

coderabbitai · 2026-03-03T22:08:46Z

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests
Post copyable unit tests in a comment
Commit unit tests in branch wenjing0224

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

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

gemini-code-assist · 2026-03-03T22:09:02Z

Summary of Changes

Hello, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces a crucial persistence layer for the CodeMemory component of the Paper2Code pipeline. By storing code generation experiences such as successful patterns, verified structures, and failure reasons in a SQLite database, the system can now leverage past knowledge across different runs. This enhancement allows the LLM to pre-load relevant historical data, improving the efficiency and accuracy of future code generation by learning from previous successes and failures.

Highlights

Database Persistence for CodeMemory: A new SQLite table repro_code_experience and an associated SQLAlchemy model have been introduced to store code generation experiences, ensuring data survives process restarts.
Experience Store Implementation: A ReproExperienceStore class was added to manage CRUD operations for these persisted experiences, allowing for adding new records and retrieving them by paper or pack ID.
CodeMemory Integration: CodeMemory now accepts an experience_store to load prior experiences (success patterns, verified structures, failure reasons) and inject them into the Large Language Model's context, improving future generation.
Automated Experience Recording: The DebuggingAgent, GenerationNode, and VerificationNode have been updated to automatically record relevant experiences (failure reasons, success patterns, verified structures) to the new persistence layer.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Changelog

alembic/versions/0021_repro_code_experience.py
- Added a new Alembic migration script to create the repro_code_experience table with columns for pack_id, paper_id, pattern_type, content, code_snippet, and created_at, along with appropriate indexes.
src/paperbot/infrastructure/stores/models.py
- Added the ReproCodeExperienceModel SQLAlchemy ORM model, mapping it to the new repro_code_experience table.
src/paperbot/infrastructure/stores/repro_experience_store.py
- Added a new file defining the ReproExperienceStore class, providing methods to add new experience records and retrieve them by paper_id or pack_id.
src/paperbot/repro/agents/debugging_agent.py
- Modified the DebuggingAgent constructor to accept an optional experience_store.
- Implemented logic to persist failure_reason experiences to the experience_store when a repair is successful.
src/paperbot/repro/memory/code_memory.py
- Updated CodeMemory to accept an optional experience_store in its constructor.
- Added _prior_experiences attribute to store loaded experiences.
- Modified get_relevant_context to include prior experiences from the database in the LLM's context.
- Added methods load_experiences_from_db, record_success_pattern, record_verified_structure, and record_failure_reason to interact with the experience_store.
- Updated the clear method to also clear _prior_experiences.
src/paperbot/repro/nodes/generation_node.py
- Modified the GenerationNode constructor to accept and pass an experience_store to CodeMemory.
- Added logic to pre-load prior experiences from the database using self.memory.load_experiences_from_db at the start of execution.
- Implemented recording of success_pattern experiences after each successful file generation.
src/paperbot/repro/nodes/verification_node.py
- Modified the VerificationNode constructor to accept an optional experience_store.
- Implemented logic to record verified_structure experiences to the experience_store when all essential verification checks pass.
tests/unit/test_repro_code_experience.py
- Added a new unit test file to cover the functionality of ReproExperienceStore and the persistence methods within CodeMemory.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	`/gemini review`	Performs a code review for the current pull request in its current state.
Pull Request Summary	`/gemini summary`	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	`/gemini help`	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

gemini-code-assist

Code Review

This pull request introduces a persistence layer for CodeMemory experiences using a SQLite database, enabling the system to learn from past code generation and debugging sessions. However, a critical prompt injection vulnerability has been identified: maliciously crafted filenames or directory names, when re-injected into the LLM prompt context from these persisted experiences, could lead to arbitrary code execution. It is crucial to sanitize or clearly delimit this context data before including it in prompts to mitigate this risk. Additionally, I recommend improving the robustness of the error handling in the new persistence logic by logging exceptions that are currently being swallowed silently, which will aid in debugging database connection or data saving issues.

gemini-code-assist · 2026-03-03T22:11:44Z

+        # 4. Prior experiences from DB (success patterns / verified structures)
+        if self._prior_experiences and remaining_chars > 200:
+            exp_lines = []
+            for exp in self._prior_experiences[:5]:
+                ptype = exp.get("pattern_type", "")
+                content = exp.get("content", "")
+                if ptype in ("success_pattern", "verified_structure") and content:
+                    exp_lines.append(f"  [{ptype}] {content}")
+            if exp_lines:
+                prior_ctx = "# === Prior Experience (same paper) ===\n" + "\n".join(exp_lines)
+                if len(prior_ctx) < remaining_chars:
+                    context_parts.append(prior_ctx)


The get_relevant_context method retrieves 'prior experiences' from the database and directly formats their content into the LLM prompt context. These experiences are populated in GenerationNode and VerificationNode using filenames and directory names. If an attacker can influence these names (e.g., by providing a malicious paper that tricks the planning agent into using specific filenames), they can inject malicious instructions into the prompt. Since the generated code is subsequently executed by the VerificationNode during import checks, this could lead to arbitrary code execution on the system.

gemini-code-assist · 2026-03-03T22:11:44Z

+                            code_snippet=repair_result.original_error[:1000],
+                        )
+                    except Exception:  # noqa: BLE001
+                        pass


Silently swallowing exceptions with except Exception: pass can hide important issues with the database or persistence logic. It's better to log the exception, even if you don't want it to interrupt the agent's execution. This will make debugging much easier if experiences are not being saved as expected. Consider using self.logger.warning with exc_info=True.

Suggested change

pass

self.logger.warning("Failed to persist code experience for failure reason.", exc_info=True)

gemini-code-assist · 2026-03-03T22:11:44Z

+                    paper_id=paper_id,
+                )
+            except Exception:  # noqa: BLE001
+                pass


Similar to my other comment, it's better to log exceptions here rather than silently swallowing them. This will help diagnose any problems with persisting the verified_structure experience.

Suggested change

pass

self.logger.warning("Failed to persist verified structure experience.", exc_info=True)

address Gemini code review issues from PR #224 and #225

* fix: validate studio output dirs and add p2c module design docs :wq * feat(PaperToContext): M3 #140 Closes #140 * feat(p2c): implement ContextEngineBridge to inject user context into extraction Related to #157 * activate paper-scope memory read/write path. Raletd to #158 * fix(p2c): address Gemini code review issues from PR. Related to #157 * fix(p2c): address Gemini code review issues. Related to #158 * fix(p2c): sanitize XML tag content to prevent tag-escape prompt injection * feat(memory): add FTS5 full-text search and sqlite-vec hybrid search. Related to #161 * feat(p2c): persist CodeMemory experiences to SQLite. Related to #162 * address Gemini code review issues from PR #224 and #225 * fix(memory): address Gemini code review issues from #153 epic audit P0 fixes: - Replace unbounded daemon threads with ThreadPoolExecutor(max_workers=2) for embedding writes; atexit + close() ensure graceful shutdown so in-flight embeddings are not lost on process exit - Restore apprise>=1.9.0 and feedgen>=1.0.0 removed in epic branch, which would have broken Epic #179 push/RSS features on merge P1 fixes: - _escape_fts: replace double-quote-only escaping with a whitelist regex [A-Za-z0-9_+-] so FTS5 operators (*, NEAR, NOT, ^) cannot alter query semantics; empty queries now short-circuit to [] - _hybrid_merge: skip items with None/invalid id instead of defaulting to id=0, preventing silent score collisions across unrelated records - ReproExperienceStore: add application-level dedup check + UNIQUE constraint (paper_id, pattern_type, content) + IntegrityError fallback to prevent duplicate experiences from accumulating across retries Migration: - 0022_repro_experience_dedup: adds uq_repro_exp_paper_type_content unique constraint to existing repro_code_experience table Tests: 47 unit tests pass (+3 new tests covering the fixes above) * fix: harden repro experience isolation and wire persistence into repro pipeline Add user-scoped isolation for repro_code_experience and enforce dedup semantics with migration 0022_repro_experience_dedup. Inject ReproExperienceStore through ReproAgent/Orchestrator/CodingAgent/GenerationNode and propagate user_id/pack_id in generation, verification, and debugging persistence paths. Update /api/gen-code to accept user_id and extend unit tests for user isolation and persistence behavior. * feat(memory): introduce memory decay mechanism (#163) Add decay-aware scoring that combines relevance (confidence), recency (exponential decay with 90-day half-life), and usage frequency to re-rank search results. New memories default to expires_at = created_at + 365 days. search_memories() now auto-touches usage on hits. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat(memory): add cross-track batch search (#164) Add search_memories_batch() that queries multiple scope_ids in a single SQL call, eliminating the N+1 loop in build_context_pack(). Engine now uses this batch method for cross-track memory retrieval. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat(context): implement layered context loading (#165) Refactor build_context_pack() into 4 layer methods: - Layer 0: user profile (cached with 5-min TTL, ~200 tokens) - Layer 1: track context — tasks/milestones (~500 tokens) - Layer 2: query-relevant + cross-track memories (~1000 tokens) - Layer 3: paper-scoped memories (on-demand) Return value adds context_layers metadata while remaining fully backward-compatible. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * refactor(memory): align decay with OpenClaw patterns - Use standard ln(2)/halfLifeDays lambda formula (matches OpenClaw temporal-decay.ts:toDecayLambda) - Default half-life lowered to 30 days (was 90, now matches OpenClaw) - Evergreen memories (global scope, preference kind) are immune to recency decay (inspired by OpenClaw isEvergreenMemoryPath) - Add _to_decay_lambda() and _is_evergreen_memory() helpers - Expand tests for lambda math, half-life precision, and evergreen logic Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(context): isolate layer0 cache and touch batch hits (#236) * feat(memory): upgrade batch retrieval with hybrid and MMR options (#237) * feat(context): add embedding fallback chain and token guard config (#238) * fix(memory): bound batch hybrid candidates by scope_ids * docs: update README with macOS python3 tips * docs: update README with more badges * docs: update README with deleting extra badges * feat(search): add offline retrieval benchmark harness Add a deterministic retrieval benchmark fixture, scorer, CLI, docs, and CI smoke gate for PaperSearchService.\n\nCloses #284\nRefs #283 * feat(context): add offline context-engine benchmark Add deterministic fixtures, scoring, CLI, and smoke coverage for layered assembly, token guard, and advisory routing in ContextEngine.\n\nCloses #286\nRefs #283 * feat(memory): add scope isolation acceptance bench Add an offline scope-isolation benchmark for memory retrieval paths, extend the metric collector with cross-user and cross-scope leak rates, and wire the new check into CI.\n\nCloses #285\nRefs #283 * feat(memory): add offline injection robustness detector Add a deterministic prompt-injection pattern detector, labeled offline fixtures, an acceptance benchmark, and CI coverage for Injection Robustness L1.\n\nCloses #287\nRefs #283 * feat(memory): add offline performance benchmark harness Add a deterministic synthetic benchmark for memory search latency baselines across 10k/100k/1M scales, plus docs and smokeable unit coverage.\n\nCloses #288\nRefs #283 * feat: add ROI benchmark for repro memory Closes #289 Refs #283 * docs: add MemoryBench epic completion report Refs #283 * docs: add runtime memory benchmark report Refs #283 * Docs: update README with our new name (#290) * docs: update README with macOS python3 tips * docs: update README with more badges * docs: update README with deleting extra badges * docs: update README with new god name --------- Co-authored-by: 林杰 <linjie@linjiedeMacBook-Air.local> Co-authored-by: 林杰 <linjie@v8d1ef64c.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef43e.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef6c5.dip.tu-dresden.de> * Fix: show friendly error when backend is unreachable (#291) * docs: update README with macOS python3 tips * docs: update README with more badges * docs: update README with deleting extra badges * docs: update README with new god name --------- Co-authored-by: 林杰 <linjie@linjiedeMacBook-Air.local> Co-authored-by: 林杰 <linjie@v8d1ef64c.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef43e.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef6c5.dip.tu-dresden.de> * feat: improve memory ROI and effectiveness benchmarks * fix: avoid importing missing data template in main * fix: stabilize live memory roi benchmark * feat: expand multi-session memory effectiveness benchmark * feat: implement MemoryBench evaluation suite with 4 bench suites Add comprehensive memory module evaluation aligned with LongMemEval (ICLR 2025), LoCoMo (ACL 2024), Mem0, and Letta benchmarks. - Retrieval Bench v2: IR metrics (Recall@5=0.873, MRR@10=0.731, nDCG@10=0.747) with 40 annotated queries across 5 question types and 5 memory dimensions - Scope Isolation + CRUD: zero-leak verification across user x scope matrix, Mem0-aligned CRUD lifecycle (add/update/delete/dedup) - Context Extraction: L0-L3 layer completeness, precision, token budget guard, TrackRouter accuracy (100%), graceful degradation - Injection Robustness L1: offline pattern detection (0% pollution, 0% FP) - Fixture dataset: 45 memories (2 users), 12 injection patterns - Testing documentation with methodology, validity analysis, and results Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: boyu <oor2020@163.com> Co-authored-by: WenjingWang <jingnvx@outlook.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: 林杰 <linjie@linjiedeMacBook-Air.local> Co-authored-by: 林杰 <linjie@v8d1ef64c.dip.tu-dresden.de> Co-authored-by: Linjie-top <linjie666z@gmail.com> Co-authored-by: 林杰 <linjie@v8d1ef43e.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef6c5.dip.tu-dresden.de>

* fix: validate studio output dirs and add p2c module design docs :wq * feat(PaperToContext): M3 #140 Closes #140 * feat(p2c): implement ContextEngineBridge to inject user context into extraction Related to #157 * activate paper-scope memory read/write path. Raletd to #158 * fix(p2c): address Gemini code review issues from PR. Related to #157 * fix(p2c): address Gemini code review issues. Related to #158 * fix(p2c): sanitize XML tag content to prevent tag-escape prompt injection * feat(memory): add FTS5 full-text search and sqlite-vec hybrid search. Related to #161 * feat(p2c): persist CodeMemory experiences to SQLite. Related to #162 * address Gemini code review issues from PR #224 and #225 * fix(memory): address Gemini code review issues from #153 epic audit P0 fixes: - Replace unbounded daemon threads with ThreadPoolExecutor(max_workers=2) for embedding writes; atexit + close() ensure graceful shutdown so in-flight embeddings are not lost on process exit - Restore apprise>=1.9.0 and feedgen>=1.0.0 removed in epic branch, which would have broken Epic #179 push/RSS features on merge P1 fixes: - _escape_fts: replace double-quote-only escaping with a whitelist regex [A-Za-z0-9_+-] so FTS5 operators (*, NEAR, NOT, ^) cannot alter query semantics; empty queries now short-circuit to [] - _hybrid_merge: skip items with None/invalid id instead of defaulting to id=0, preventing silent score collisions across unrelated records - ReproExperienceStore: add application-level dedup check + UNIQUE constraint (paper_id, pattern_type, content) + IntegrityError fallback to prevent duplicate experiences from accumulating across retries Migration: - 0022_repro_experience_dedup: adds uq_repro_exp_paper_type_content unique constraint to existing repro_code_experience table Tests: 47 unit tests pass (+3 new tests covering the fixes above) * fix: harden repro experience isolation and wire persistence into repro pipeline Add user-scoped isolation for repro_code_experience and enforce dedup semantics with migration 0022_repro_experience_dedup. Inject ReproExperienceStore through ReproAgent/Orchestrator/CodingAgent/GenerationNode and propagate user_id/pack_id in generation, verification, and debugging persistence paths. Update /api/gen-code to accept user_id and extend unit tests for user isolation and persistence behavior. * feat(memory): introduce memory decay mechanism (#163) Add decay-aware scoring that combines relevance (confidence), recency (exponential decay with 90-day half-life), and usage frequency to re-rank search results. New memories default to expires_at = created_at + 365 days. search_memories() now auto-touches usage on hits. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat(memory): add cross-track batch search (#164) Add search_memories_batch() that queries multiple scope_ids in a single SQL call, eliminating the N+1 loop in build_context_pack(). Engine now uses this batch method for cross-track memory retrieval. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat(context): implement layered context loading (#165) Refactor build_context_pack() into 4 layer methods: - Layer 0: user profile (cached with 5-min TTL, ~200 tokens) - Layer 1: track context — tasks/milestones (~500 tokens) - Layer 2: query-relevant + cross-track memories (~1000 tokens) - Layer 3: paper-scoped memories (on-demand) Return value adds context_layers metadata while remaining fully backward-compatible. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * refactor(memory): align decay with OpenClaw patterns - Use standard ln(2)/halfLifeDays lambda formula (matches OpenClaw temporal-decay.ts:toDecayLambda) - Default half-life lowered to 30 days (was 90, now matches OpenClaw) - Evergreen memories (global scope, preference kind) are immune to recency decay (inspired by OpenClaw isEvergreenMemoryPath) - Add _to_decay_lambda() and _is_evergreen_memory() helpers - Expand tests for lambda math, half-life precision, and evergreen logic Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(context): isolate layer0 cache and touch batch hits (#236) * feat(memory): upgrade batch retrieval with hybrid and MMR options (#237) * feat(context): add embedding fallback chain and token guard config (#238) * fix(memory): bound batch hybrid candidates by scope_ids * docs: update README with macOS python3 tips * docs: update README with more badges * docs: update README with deleting extra badges * feat(search): add offline retrieval benchmark harness Add a deterministic retrieval benchmark fixture, scorer, CLI, docs, and CI smoke gate for PaperSearchService.\n\nCloses #284\nRefs #283 * feat(context): add offline context-engine benchmark Add deterministic fixtures, scoring, CLI, and smoke coverage for layered assembly, token guard, and advisory routing in ContextEngine.\n\nCloses #286\nRefs #283 * feat(memory): add scope isolation acceptance bench Add an offline scope-isolation benchmark for memory retrieval paths, extend the metric collector with cross-user and cross-scope leak rates, and wire the new check into CI.\n\nCloses #285\nRefs #283 * feat(memory): add offline injection robustness detector Add a deterministic prompt-injection pattern detector, labeled offline fixtures, an acceptance benchmark, and CI coverage for Injection Robustness L1.\n\nCloses #287\nRefs #283 * feat(memory): add offline performance benchmark harness Add a deterministic synthetic benchmark for memory search latency baselines across 10k/100k/1M scales, plus docs and smokeable unit coverage.\n\nCloses #288\nRefs #283 * feat: add ROI benchmark for repro memory Closes #289 Refs #283 * docs: add MemoryBench epic completion report Refs #283 * docs: add runtime memory benchmark report Refs #283 * Docs: update README with our new name (#290) * docs: update README with macOS python3 tips * docs: update README with more badges * docs: update README with deleting extra badges * docs: update README with new god name --------- Co-authored-by: 林杰 <linjie@linjiedeMacBook-Air.local> Co-authored-by: 林杰 <linjie@v8d1ef64c.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef43e.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef6c5.dip.tu-dresden.de> * Fix: show friendly error when backend is unreachable (#291) * docs: update README with macOS python3 tips * docs: update README with more badges * docs: update README with deleting extra badges * docs: update README with new god name --------- Co-authored-by: 林杰 <linjie@linjiedeMacBook-Air.local> Co-authored-by: 林杰 <linjie@v8d1ef64c.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef43e.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef6c5.dip.tu-dresden.de> * feat: improve memory ROI and effectiveness benchmarks * fix: avoid importing missing data template in main * fix: stabilize live memory roi benchmark * feat: expand multi-session memory effectiveness benchmark * feat: implement MemoryBench evaluation suite with 4 bench suites Add comprehensive memory module evaluation aligned with LongMemEval (ICLR 2025), LoCoMo (ACL 2024), Mem0, and Letta benchmarks. - Retrieval Bench v2: IR metrics (Recall@5=0.873, MRR@10=0.731, nDCG@10=0.747) with 40 annotated queries across 5 question types and 5 memory dimensions - Scope Isolation + CRUD: zero-leak verification across user x scope matrix, Mem0-aligned CRUD lifecycle (add/update/delete/dedup) - Context Extraction: L0-L3 layer completeness, precision, token budget guard, TrackRouter accuracy (100%), graceful degradation - Injection Robustness L1: offline pattern detection (0% pollution, 0% FP) - Fixture dataset: 45 memories (2 users), 12 injection patterns - Testing documentation with methodology, validity analysis, and results Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * docs: add MemoryBench quantitative results to README Add evaluation results section with metrics from 4 bench suites: - Retrieval quality (Recall@5=0.873, MRR@10=0.731, nDCG@10=0.747) - Scope isolation + CRUD lifecycle (zero leaks, all CRUD pass) - Context extraction (100% precision, 100% router accuracy) - Injection robustness (0% pollution, 0% false positive) Includes LoCoMo question-type breakdown and run instructions. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: boyu <oor2020@163.com> Co-authored-by: WenjingWang <jingnvx@outlook.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: 林杰 <linjie@linjiedeMacBook-Air.local> Co-authored-by: 林杰 <linjie@v8d1ef64c.dip.tu-dresden.de> Co-authored-by: Linjie-top <linjie666z@gmail.com> Co-authored-by: 林杰 <linjie@v8d1ef43e.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef6c5.dip.tu-dresden.de>

* fix: validate studio output dirs and add p2c module design docs :wq * feat(PaperToContext): M3 #140 Closes #140 * feat(p2c): implement ContextEngineBridge to inject user context into extraction Related to #157 * activate paper-scope memory read/write path. Raletd to #158 * fix(p2c): address Gemini code review issues from PR. Related to #157 * fix(p2c): address Gemini code review issues. Related to #158 * fix(p2c): sanitize XML tag content to prevent tag-escape prompt injection * feat(memory): add FTS5 full-text search and sqlite-vec hybrid search. Related to #161 * feat(p2c): persist CodeMemory experiences to SQLite. Related to #162 * address Gemini code review issues from PR #224 and #225 * fix(memory): address Gemini code review issues from #153 epic audit P0 fixes: - Replace unbounded daemon threads with ThreadPoolExecutor(max_workers=2) for embedding writes; atexit + close() ensure graceful shutdown so in-flight embeddings are not lost on process exit - Restore apprise>=1.9.0 and feedgen>=1.0.0 removed in epic branch, which would have broken Epic #179 push/RSS features on merge P1 fixes: - _escape_fts: replace double-quote-only escaping with a whitelist regex [A-Za-z0-9_+-] so FTS5 operators (*, NEAR, NOT, ^) cannot alter query semantics; empty queries now short-circuit to [] - _hybrid_merge: skip items with None/invalid id instead of defaulting to id=0, preventing silent score collisions across unrelated records - ReproExperienceStore: add application-level dedup check + UNIQUE constraint (paper_id, pattern_type, content) + IntegrityError fallback to prevent duplicate experiences from accumulating across retries Migration: - 0022_repro_experience_dedup: adds uq_repro_exp_paper_type_content unique constraint to existing repro_code_experience table Tests: 47 unit tests pass (+3 new tests covering the fixes above) * fix: harden repro experience isolation and wire persistence into repro pipeline Add user-scoped isolation for repro_code_experience and enforce dedup semantics with migration 0022_repro_experience_dedup. Inject ReproExperienceStore through ReproAgent/Orchestrator/CodingAgent/GenerationNode and propagate user_id/pack_id in generation, verification, and debugging persistence paths. Update /api/gen-code to accept user_id and extend unit tests for user isolation and persistence behavior. * feat(memory): introduce memory decay mechanism (#163) Add decay-aware scoring that combines relevance (confidence), recency (exponential decay with 90-day half-life), and usage frequency to re-rank search results. New memories default to expires_at = created_at + 365 days. search_memories() now auto-touches usage on hits. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat(memory): add cross-track batch search (#164) Add search_memories_batch() that queries multiple scope_ids in a single SQL call, eliminating the N+1 loop in build_context_pack(). Engine now uses this batch method for cross-track memory retrieval. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat(context): implement layered context loading (#165) Refactor build_context_pack() into 4 layer methods: - Layer 0: user profile (cached with 5-min TTL, ~200 tokens) - Layer 1: track context — tasks/milestones (~500 tokens) - Layer 2: query-relevant + cross-track memories (~1000 tokens) - Layer 3: paper-scoped memories (on-demand) Return value adds context_layers metadata while remaining fully backward-compatible. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * refactor(memory): align decay with OpenClaw patterns - Use standard ln(2)/halfLifeDays lambda formula (matches OpenClaw temporal-decay.ts:toDecayLambda) - Default half-life lowered to 30 days (was 90, now matches OpenClaw) - Evergreen memories (global scope, preference kind) are immune to recency decay (inspired by OpenClaw isEvergreenMemoryPath) - Add _to_decay_lambda() and _is_evergreen_memory() helpers - Expand tests for lambda math, half-life precision, and evergreen logic Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(context): isolate layer0 cache and touch batch hits (#236) * feat(memory): upgrade batch retrieval with hybrid and MMR options (#237) * feat(context): add embedding fallback chain and token guard config (#238) * fix(memory): bound batch hybrid candidates by scope_ids * docs: update README with macOS python3 tips * docs: update README with more badges * docs: update README with deleting extra badges * feat(search): add offline retrieval benchmark harness Add a deterministic retrieval benchmark fixture, scorer, CLI, docs, and CI smoke gate for PaperSearchService.\n\nCloses #284\nRefs #283 * feat(context): add offline context-engine benchmark Add deterministic fixtures, scoring, CLI, and smoke coverage for layered assembly, token guard, and advisory routing in ContextEngine.\n\nCloses #286\nRefs #283 * feat(memory): add scope isolation acceptance bench Add an offline scope-isolation benchmark for memory retrieval paths, extend the metric collector with cross-user and cross-scope leak rates, and wire the new check into CI.\n\nCloses #285\nRefs #283 * feat(memory): add offline injection robustness detector Add a deterministic prompt-injection pattern detector, labeled offline fixtures, an acceptance benchmark, and CI coverage for Injection Robustness L1.\n\nCloses #287\nRefs #283 * feat(memory): add offline performance benchmark harness Add a deterministic synthetic benchmark for memory search latency baselines across 10k/100k/1M scales, plus docs and smokeable unit coverage.\n\nCloses #288\nRefs #283 * feat: add ROI benchmark for repro memory Closes #289 Refs #283 * docs: add MemoryBench epic completion report Refs #283 * docs: add runtime memory benchmark report Refs #283 * Docs: update README with our new name (#290) * docs: update README with macOS python3 tips * docs: update README with more badges * docs: update README with deleting extra badges * docs: update README with new god name --------- Co-authored-by: 林杰 <linjie@linjiedeMacBook-Air.local> Co-authored-by: 林杰 <linjie@v8d1ef64c.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef43e.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef6c5.dip.tu-dresden.de> * Fix: show friendly error when backend is unreachable (#291) * docs: update README with macOS python3 tips * docs: update README with more badges * docs: update README with deleting extra badges * docs: update README with new god name --------- Co-authored-by: 林杰 <linjie@linjiedeMacBook-Air.local> Co-authored-by: 林杰 <linjie@v8d1ef64c.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef43e.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef6c5.dip.tu-dresden.de> * feat: improve memory ROI and effectiveness benchmarks * fix: avoid importing missing data template in main * fix: stabilize live memory roi benchmark * feat: expand multi-session memory effectiveness benchmark * feat: implement MemoryBench evaluation suite with 4 bench suites Add comprehensive memory module evaluation aligned with LongMemEval (ICLR 2025), LoCoMo (ACL 2024), Mem0, and Letta benchmarks. - Retrieval Bench v2: IR metrics (Recall@5=0.873, MRR@10=0.731, nDCG@10=0.747) with 40 annotated queries across 5 question types and 5 memory dimensions - Scope Isolation + CRUD: zero-leak verification across user x scope matrix, Mem0-aligned CRUD lifecycle (add/update/delete/dedup) - Context Extraction: L0-L3 layer completeness, precision, token budget guard, TrackRouter accuracy (100%), graceful degradation - Injection Robustness L1: offline pattern detection (0% pollution, 0% FP) - Fixture dataset: 45 memories (2 users), 12 injection patterns - Testing documentation with methodology, validity analysis, and results Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * docs: add MemoryBench quantitative results to README Add evaluation results section with metrics from 4 bench suites: - Retrieval quality (Recall@5=0.873, MRR@10=0.731, nDCG@10=0.747) - Scope isolation + CRUD lifecycle (zero leaks, all CRUD pass) - Context extraction (100% precision, 100% router accuracy) - Injection robustness (0% pollution, 0% false positive) Includes LoCoMo question-type breakdown and run instructions. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * feat(AgentSwarm): add Claude/Codex workflow, workspace persistence, and human review UX (#296) * feat(AgentSwarm): add Claude/Codex workflow, workspace persistence, and human review UX - 添加 /api/agent-board 路由和 Swarm Commander/Dispatcher 基础架构 - 将生成的任务文件持久化到已配置的工作区 - 在工作区中为每个任务生成用户审核文档（reviews/<任务>-user-review.md） - 添加代理看板 UI，包含看板、任务详细日志和人工审核操作 - 要求“全部运行”任务设置工作区，并添加 VS Code 打开操作 - 实现 Studio 存储和 SSE 任务的更新/插入/同步行为 - 为路由流程、持久化和超时处理添加单元/功能测试 - 更新 Studio 布局、PostCSS 配置、后端 URL 助手和锁定文件 Closes #197 * (fix)修改AI review问题 * refactor: share research fetch helpers (#295) Co-authored-by: 林杰 <linjie@v8d1ef6c5.dip.tu-dresden.de> * feat: add research/yearCombobox. Realted to #297 * refactor(research): dedupe fetch helpers across Research pages; fix Radix popover import; add @radix-ui/react-popover dep * fix(web): move @radix-ui/react-popover to dependencies * fix(research): keep track list order stable across activation * refactor(research): share stable track merge helper * fix(research): make paper feedback togglable * feat(research): hide 'Open Discovery Workspace' behind feature flag * feat(research): gate track memory button behind feature flag (#333) * feat(research): gate track memory button behind feature flag * fix(research): avoid reactivating already active track * ci(vercel): add auto deploy workflow for dev branch * feat: calm dashboard workspace (#334) * feat: harden search and connector infrastructure (#339) * refactor(infra): unify async request layer for connectors Implements issue #262 by moving Arxiv/OpenAlex/PapersCool connectors onto the shared async transport with retry support, updating async call sites, and adding focused tests. * perf(infra): batch OpenAlex ID lookups Implements issue #267 by replacing per-ID OpenAlex work fetches with batched filter queries and adds focused coverage for the batched path. * refactor(infra): unify Semantic Scholar client stack Implements issue #266 by routing both the shared mixin and the scholar-tracking agent through the same SemanticScholarClient surface, removing the duplicated legacy API client path, and adding focused unification tests. * feat(infra): harden SSE transport handling Implements issue #263 by moving heartbeat, timeout, cancellation cleanup, and X-Accel-Buffering handling into the shared SSE wrapper and switching the streaming routes onto the common response helper with regression coverage. * fix(infra): bound ARQ jobs and reuse event log Implements issue #268 by adding timeout/max_tries metadata to worker functions, reusing the SqlAlchemyEventLog singleton inside the worker module, and covering both behaviors with focused tests. * feat(search): add three-tier paper deduplicator (DOI/arxiv_id/rapidfuzz) Replaces the identity-key-only dedup in PaperSearchService._fuse_with_rrf() with a dedicated PaperDeduplicator that matches across DOI, arxiv_id (version- stripped), and fuzzy title similarity via rapidfuzz. Merged papers accumulate the best metadata (highest citations, longest abstract, union of identities). Closes #317 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(search): prevent conflicting identity dedup merges Refs #317 * fix(embeddings): add CJK character support to hash embedding tokenizer Extends HashEmbeddingProvider regex to include CJK Unified Ideographs (U+4E00–U+9FFF) and Extension A (U+3400–U+4DBF), enabling proper embedding of Chinese text. Closes #276 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(embeddings): extend hash tokenizer beyond Han-only CJK Refs #276 --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> * fix(research): restore feedback track ux on current dev (#341) * refactor(research): dedupe fetch helpers across Research pages; fix Radix popover import; add @radix-ui/react-popover dep * fix(web): preserve item order in research tracks selection Closes #304 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * refactor(web): remove unused components Closes #305 Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(web): keep active research track visible Refs #304 * fix(research): align paper feedback toggles with persisted state Refs #324 * fix(web): add missing @radix-ui/react-popover dependency Research page crashed with "Module not found: Can't resolve '@radix-ui/react-popover'" because SearchBox.tsx imports popover.tsx which depends on this package. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * fix(research): restore feedback track ux on current dev --------- Co-authored-by: 林杰 <linjie@v8d1ef6a5.dip.tu-dresden.de> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> * fix(web): resolve DeepCode Studio page issues * fix(repro): improve memory and workflow correctness * feat: add intelligence radar backend * fix(api): block cross-user intelligence feed access * fix(intelligence): avoid thread-unsafe refresh races * refactor: enforce alembic-only store initialization * fix(db): bootstrap sqlite schemas for lazy stores * fix(stores): align ports with store interfaces * fix(stores): align remaining port contracts * fix(api): harden request boundaries and settings handling * fix(api): stop trusting spoofable forwarded hosts * fix(api): gate studio code execution by default * fix(api): stream request-size checks safely * feat(export): add Obsidian filesystem exporter * feat(cli): add Obsidian export command * feat(obsidian): finish configurable vault export workflow (#348) * refactor(web): move workflow workbench out of dashboard * ci(vercel): use --preview flag for dev deploy * fix(ci): remove invalid Vercel preview flag * fix(ci): keep Vercel secrets free of inline comments * refactor: share candidate search boundary and simplify dashboard * refactor(api): split paperscool search curate and ingest * refactor(web): flatten dashboard brief snapshot * refactor(web): reduce dashboard brief visual noise * refactor(web): feature top dashboard signals * fix(obsidian): harden export follow-ups * fix(obsidian): tighten review follow-ups * feat(web): add Obsidian handoff to Research workspace (#352) * feat(web): add Obsidian handoff to Research workspace * fix(web): address obsidian workspace review feedback * feat: ship track-centric research context stack (#356) * feat(api): add track context read model service * feat(api): expose consolidated track context endpoint * refactor(web): migrate research page to track context endpoint * refactor(memory): wrap track-scoped memory access * chore: document track-centric research model and stabilize tests * docs: refresh readme screenshots * fix(deps): align package manifests with runtime imports * docs: refresh email push screenshot * feat(research): close citation graph and obsidian export gaps (#360) * feat(obsidian): add bidirectional vault sync * fix(paper): improve saved papers table layout * feat(openclaw): add paperbot plugin bridge * fix(obsidian): address sync review feedback * feat(web): flatten dashboard action bands * refactor(web): condense dashboard next-up panel * refactor(web): simplify dashboard around recommendations * refactor(web): simplify dashboard surface and workflow copy * feat(web): add dashboard queue actions * fix(web): harden dashboard queue links and brief parsing * fix(paper): unify unsave with feedback API * fix(paper): scope saved papers per track * fix: unblock dashboard build and e2e * ci: add vercel pr preview automation * fix: repair vercel preview workflow setup * fix: run vercel build from repo root * fix: skip preview smoke without bypass secret * feat(auth): add multi-user authentication foundation. Related to #151 - Add User domain model and SQLAlchemy UserModel with soft-delete support - Add SqlAlchemyUserStore with email/GitHub user CRUD, password auth, reset tokens - Add JWT signing/verification (python-jose), bcrypt password hashing - Add FastAPI auth dependencies: get_user_id (optional fallback) and get_current_user (strict) - Add /api/auth routes: register, login, github/exchange, me, forgot/reset-password - Add Alembic migrations for users and password_reset_tokens tables - Add AUTH_OPTIONAL env var for gradual migration from legacy 'default' user - Fix account lifecycle bugs: reactivate on OAuth re-login, reject inactive on password login - Add auth API tests * fix(auth): address code review feedback on PR #365 * feat(document): add explicit evidence indexing pipeline * docs(benchmark): define document evidence eval contract * feat(benchmark): add document evidence eval scaffold * fix(api): restore py39 auth compatibility after dev rebase * chore(logging): surface cleanup and FTS5 failures * fix(paper): refine saved filters and cleanup code * fix(paper): refine paper context cleanup and year filter * feat(eval): support dedicated embedding benchmark providers * feat(settings): add embedding endpoint configuration * refactor(settings): align embedding endpoint ux with cc-switch * refactor(settings): simplify embedding endpoint panel * refactor(settings): tighten embedding layout on wide screens * feat(studio): refine paper gallery icon animation and context workspace layout * ci: disable native vercel git deploys * docs: restore master demo gallery assets * fix: unblock ci for merge-dev-into-master * fix: address codeql alerts * fix: harden agent board workspace path validation --------- Co-authored-by: boyu <oor2020@163.com> Co-authored-by: WenjingWang <jingnvx@outlook.com> Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com> Co-authored-by: 林杰 <linjie@linjiedeMacBook-Air.local> Co-authored-by: 林杰 <linjie@v8d1ef64c.dip.tu-dresden.de> Co-authored-by: Linjie-top <linjie666z@gmail.com> Co-authored-by: 林杰 <linjie@v8d1ef43e.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef6c5.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef6a5.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef4ea.dip.tu-dresden.de> Co-authored-by: 林杰 <linjie@v8d1ef41f.dip.tu-dresden.de>

feat(p2c): persist CodeMemory experiences to SQLite. Related to #162

4384ad4

wen-placeholder requested a review from jerry609 as a code owner March 3, 2026 22:08

wen-placeholder merged commit fefe5b6 into epic/p2c-memory-and-context-optimization Mar 3, 2026
4 checks passed

gemini-code-assist Bot reviewed Mar 3, 2026

View reviewed changes

wen-placeholder added a commit that referenced this pull request Mar 3, 2026

address Gemini code review issues from PR #224 and #225

ab32340

wen-placeholder added a commit that referenced this pull request Mar 3, 2026

Merge pull request #226 from jerry609/wenjing0224

4e1b7c6

address Gemini code review issues from PR #224 and #225

jerry609 pushed a commit that referenced this pull request Mar 5, 2026

address Gemini code review issues from PR #224 and #225

fb191f8

jerry609 pushed a commit that referenced this pull request Mar 5, 2026

address Gemini code review issues from PR #224 and #225

20cb103

jerry609 pushed a commit that referenced this pull request Mar 5, 2026

address Gemini code review issues from PR #224 and #225

85cfba0

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(p2c): persist CodeMemory experiences to SQLite. Related to #162#225

feat(p2c): persist CodeMemory experiences to SQLite. Related to #162#225
wen-placeholder merged 1 commit into
epic/p2c-memory-and-context-optimizationfrom
wenjing0224

wen-placeholder commented Mar 3, 2026

Uh oh!

vercel Bot commented Mar 3, 2026

Uh oh!

coderabbitai Bot commented Mar 3, 2026

Review skipped

Uh oh!

gemini-code-assist Bot commented Mar 3, 2026

Uh oh!

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot Mar 3, 2026

Uh oh!

gemini-code-assist Bot Mar 3, 2026

Uh oh!

gemini-code-assist Bot Mar 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

	pass
	self.logger.warning("Failed to persist code experience for failure reason.", exc_info=True)

	pass
	self.logger.warning("Failed to persist verified structure experience.", exc_info=True)

Conversation

wen-placeholder commented Mar 3, 2026

Uh oh!

vercel Bot commented Mar 3, 2026

Uh oh!

coderabbitai Bot commented Mar 3, 2026

Review skipped

Uh oh!

gemini-code-assist Bot commented Mar 3, 2026

Summary of Changes

Highlights

Footnotes

Uh oh!

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Mar 3, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Mar 3, 2026

Choose a reason for hiding this comment

Uh oh!

gemini-code-assist Bot Mar 3, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant