ContextSniper

Definition

ContextSniper is AntTrail’s code-repair module for precision evidence selection in repository-level program repair (arXiv 2607.01916, Luk et al.). LLM agents repairing real repository issues spend large context budgets on whole-file reads, broad searches, and long terminal outputs where useful evidence is mixed with irrelevant code and logs.

Key Points

• Four-step pipeline: (1) index code and action memory as three abstract levels; (2) retrieve candidates with a hybrid ranker; (3) filter long tool output through an intention-aware context gate; (4) return compact evidence packets, with full source kept recoverable on demand.
• Results (matched 50-task-per-condition, SWE-bench Lite): OpenClaw — token use −51.5%, logged cost −36.4%; Claude Code — token use −38.9%, logged cost −27.3%; submitted-resolution rates “essentially unchanged.”
• Mechanism contrast: unlike naive compression, evidence packets are selected by an intention-aware gate and full source stays recoverable — the information is not destroyed, just deferred.
• Benchmark: beats existing memory- and RAG-style integrations on token efficiency in a five-task comparison.

Implications

ContextSniper is the strongest measured counterexample to the paper’s warning: it reduces both tokens and logged cost (−36.4% / −27.3%) because it preserves evidence recoverability (select, don’t destroy). The design difference is exactly what RTK’s issue #582 lacked: full source is recoverable on demand, so the model can re-read when needed. It validates the “success-adjusted billed cost” lens as the right evaluation metric.

Implementation (confidence: medium)

The GitHub repo Calluking/ContextSniper (capture 2026-07-31) is the same project family as the arXiv paper — identical name, 3-stage pipeline, dual Claude Code + OpenClaw hosts, matched SWE-bench Lite methodology, near-matching headline numbers. The relationship is never declared in the repo (no arXiv citation, different author handle, paper points to gitcode.com), so authorship is unverified at medium confidence.

• Ships two local plugins: claude-plugin/ (MCP tools: contextsniper_search_code, contextsniper_edit_file, contextsniper_index_codebase, contextsniper_health) and openclaw-plugin/ (native tools + prompt policy injection). Both start a local backend + AGFS service (c4pt0r/agfs via pyagfs), stopped when host exits.
• Repo numbers: Claude Code −39% total tokens (−47% with Filter); OpenClaw −41% (−52% with Filter); embedding via text-embedding-3-large. Repo omits cost figures that the paper reports.
• Config: CONTEXTSNIPER_FILTER_ENABLED (default 1), CONTEXTSNIPER_INJECT_FILTERING_PROMPT toggle for the shipped prompt-policy injection files (claude-plugin/prompts/code_policy_injection.txt, openclaw-plugin/prompts/code_policy_injection.txt). License MulanPSL-2.0. Requirements: Python 3.11+, Go 1.22+, Claude Code/OpenClaw CLI, OpenAI-compatible embedding endpoint.

Open Questions

• Whether “logged cost” is provider-billed or an estimate (methodology not fully specified in capture).
• Repo availability (gitcode.com) vs the GitHub ecosystem used by other tools here.
• Authorship unverified: repo never cites arXiv 2607.01916 or AntTrail; paper’s lukchiwang/ContextSniper lives on gitcode.com while this is Calluking on GitHub; whether Calluking = author Chiwang Luk is unverifiable from captures.
• Number discrepancy: repo −39%/−41% (plain) and −47%/−52% (Filter) tokens vs paper −38.9%/−51.5% tokens and −27.3%/−36.4% cost; repo omits cost entirely. Token %s roughly align; treat as config/version-sensitive, not a hard contradiction.

Related

token-reduction-is-not-cost-reduction — the evaluation lens this tool passes rtk-rust-token-killer — the compression failure mode ContextSniper avoids token-usage-reduction — the cost-lever space it operates in codebase-memory — analogous knowledge-graph memory approach context-mode — mechanism parallel: select-don’t-destroy intent filtering at tool-output scale

Sources

• raw/external/arxiv-org-2607-01916-96bad417.md
• raw/external/github-com-contextsniper-7749d2f0.md