Releases·Tool Release
baoer_signal_grep v1.7.0-1 makes LLM code-search judge fail in bounded batches instead of all at once
The maintainer of baoer_signal_grep, an AGPL-3.0-only open-source code-search tool, released v1.7.0-1, which runs the optional LLM judging layer ("Jev") in bounded, concurrently-started batches rather than as one all-or-nothing request. The tool now exposes explicit partial-failure state (unjudged candidates, per-batch outcomes) so agents can reason about incomplete retrieval instead of seeing a generic success or failure. It works via MCP with Claude Code, Codex, and Kimi, plus Pi and OMP integrations.
0 engagement1 sourceTue, Sep 22, 2026, 08:39 AM
AI Overview
- baoer_signal_grep (AGPL-3.0-only) caps semantic judging batches at 8 candidates and 64 KiB per request; HTTP 413 splits a batch in half and retries sequentially
- The 1.7.0-1 response exposes partial-state fields including candidatesUnjudged, batchesAttempted, batchesCompleted, batchesFailed, and batchesSplit
- A tested 20-candidate restart run produced 3 batches (8+8+4), judged all 20, and left 0 unjudged; successful batches survive later failures with deterministic local ordering
- Exact search and source inspection run locally without an LLM provider; Jev is optional for concept and hybrid retrieval, so provider failure never reads as 'no matches'
- The MCP-compatible package integrates with Claude Code, Codex, Kimi, plus Pi and OMP
Related
Releases·Tool Release
What I learned making an evidence-first local search layer for coding agents
A developer released sift-light, a local search layer for coding agents designed to prevent agents from answering from plausible-but-unverified excerpts. It prioritizes exact search, keeps literal evidence from being displaced by semantic candidates, and supports a validation step that inspects retained source against the file on disk.
A developer released sift-light, a local search layer for coding agents designed to prevent agents from answering from plausible-but-unverified excerpts. It prioritizes exact search, keeps literal evidence from being displaced by semantic candidates, and supports a validation step that inspects retained source against the file on disk.
15 engagement1 sourceredditMon, Oct 12, 2026, 11:19 AM
Releases·Tool Release
Designing a protocol for coding agents: responses that report what they left out, and refusals with explicit retry rules (Sley 2.0, open source)
Greyforge Labs released Sley 2.0, an Apache-2.0 open-source programming system where agents talk to a deterministic Rust kernel over the SMP1 framed protocol instead of editing source files. SMP1 responses carry a `bounds` record stating applied limits and what was omitted, and failures use stable error codes grouped into explicit retry classes. The design targets anyone building tool interfaces for models.
Greyforge Labs released Sley 2.0, an Apache-2.0 open-source programming system where agents talk to a deterministic Rust kernel over the SMP1 framed protocol instead of editing source files. SMP1 responses carry a `bounds` record stating applied limits and what was omitted, and failures use stable error codes grouped into explicit retry classes. The design targets anyone building tool interfaces for models.
1 engagement1 sourceredditFri, Sep 25, 2026, 11:18 PM
Products·building
Local search tool that lets AI agents find exact sentences in large repos
A small local 'librarian' tool for AI agents that indexes repositories, logs, and docs so agents can retrieve one specific sentence instead of dumping vague matches. It targets developers whose coding agents search large repos poorly. LLMs are the consuming agents; the tool provides better local retrieval for them.
A small local 'librarian' tool for AI agents that indexes repositories, logs, and docs so agents can retrieve one specific sentence instead of dumping vague matches. It targets developers whose coding agents search large repos poorly. LLMs are the consuming agents; the tool provides better local retrieval for them.
17 engagement1 sourceredditThu, Sep 10, 2026, 02:19 AM
Developments·Paper
Study finds deep agentic search beats semantic search for repo-level code QA
An empirical study on arXiv compares semantic search against deep agentic search (grep-search by subagent) for repository-level code question answering. The deep agentic approach, which delegates exploration to a subagent in an isolated context, outperforms semantic search, likely due to reduced context pollution.
An empirical study on arXiv compares semantic search against deep agentic search (grep-search by subagent) for repository-level code question answering. The deep agentic approach, which delegates exploration to a subagent in an isolated context, outperforms semantic search, likely due to reduced context pollution.
0 engagement1 sourcearxivSun, Aug 2, 2026, 09:36 PM
Releases·Tool Release
BossConsole: open-source JVM-based harness for AI agents released on GitHub
BossConsole, an open-source, multi-platform operator's console for AI agents, has been released on GitHub. Built on the JVM (not Electron), it supports running Claude Code, Codex, Gemini, or OpenCode with a real browser, terminal, editor, secrets, and over 100 MCP tools. The project targets enterprises, science, and research.
BossConsole, an open-source, multi-platform operator's console for AI agents, has been released on GitHub. Built on the JVM (not Electron), it supports running Claude Code, Codex, Gemini, or OpenCode with a real browser, terminal, editor, secrets, and over 100 MCP tools. The project targets enterprises, science, and research.
57 engagement2 sourcesgithub, redditTue, Jul 21, 2026, 01:01 AM