Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add instructions/rocketvish/bearing/agents-mdgit clone --depth 1 https://github.com/rocketvish/bearingWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.01438 | $0.01438 |
| Opus 5 | $0.00719 | $0.00719 |
| Sonnet 5 | $0.00288 | $0.00288 |
| Haiku 4.5 | $0.00144 | $0.00144 |
Grade A, and why
bearing AGENTS.md scanned grade A with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Makes network callslowCapability
Not a fault in itself. Listed so you know the mod talks to something, and to what.
- `agent.py` -- Tool-use agent via provider adapters (OpenAI Responses or Anthropic Messages, urllib/no SDK), with prompt caching + reasoning effort Copies of this mod
1 near-identical copy found in the catalogue:
- bearing CLAUDE.md — 92% identical, 10 lines differ
How it starts
The opening of the file, as written. The whole thing — 117 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Bearing - Task Orchestrator for Codex
Architecture
bearing.py-- CLI entry point + orchestration loop (run_orchestrator,propagate_context)executor.py-- Prompt assembly (assemble_prompt) + CLI execution (run_task)relevance.py-- Embedding-based relevance scoring + compression via Ollamaeval_runner.py-- Evaluation framework: runs benchmark under 4 context format conditionseval_compare.py-- Eval compare: Bearing (8 sessions) vs single session (1 mega-prompt)agent.py-- Tool-use agent via provider adapters (OpenAI Responses or Anthropic Messages, urllib/no SDK), with prompt caching + reasoning effortcompressor.py-- Mid-conversation history compression (API or Ollama backends)retriever.py-- Embedding-based selective history retrieval via Ollama (nomic-embed-text)eval_agent.py-- Eval: agent with/without compression/caching/retrieval vs Codex -p (7 conditions)test_sanity.py-- Quick sanity test for caching + reasoning featurestasks_schema.py-- Dataclasses for Task, TaskQueue, TaskResultstatus_writer.py-- Generates status.md from task queue state
Eval Framework
Four context format conditions: prose, structured, embedding, embedding+llm.
Eval Compare (bearing eval-compare)
Compares Bearing's 8-session approach against a single mega-prompt session. This tests the core value proposition: task isolation prevents context accumulation and compaction. The single session gets the same task prompts and FOCUS directives but must build everything in one accumulated context. Key metrics: total tokens, compaction count, per-task quality.
Key Learnings
Embedding similarity baseline is high for same-domain text. nomic-embed-text cosine similarity for chunks within the same project bottoms out around 0.5-0.6, not 0.2-0.3. Default thresholds are set at keep=0.75, drop=0.55 to account for this. Benchmarks that need chunks to actually drop require genuinely unrelated domains (e.g., different languages or problem spaces), not just "parallel feature tracks" in the same codebase.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 117 lines · 1,438 tokens per session scan A 60827cc69883
bearing AGENTS.md is an instructions file published in the GitHub repository rocketvish/bearing (2 stars, last pushed 2mo ago), licensed MIT. It adds 1,438 tokens to every session, about $0.0072 per session on Opus 5. A static security scan graded it A with 1 finding (makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other instructions, from other repositories
agents
Always-loaded project anchor. Read this first. Contains project identity, non-negotiables, commands, and pointer to ROUTER.md for full context.
Tianshu-harness CLAUDE.md
Instructions for huiliyi37/Tianshu-harness, covering 天枢 (tianshu) / rivet, build & test, architecture, conventions and known constraints.
open-skill-sunset AGENTS.md
Instructions for ooocooc/open-skill-sunset, a project described as: Local, read-only audit for stale AGENTS.md, CLAUDE.md, and generic SKILL.md instructions.
codex-image-context-runtime AGENTS.md
Instructions for shixinnt/codex-image-context-runtime, covering agents.md, public boundary, runtime contract and changes.
navegador CLAUDE.md
Instructions for ConflictHQ/navegador, covering navegador — claude context, what it is, stack, package layout and falkordb connection.
contextweaver routing.instructions.md
Instructions for dgenio/contextweaver, covering routing engine — agent instructions, choicegraph validation invariants (graph.py), treebuilder grouping strategies (tree.py), router beam-search constraints (router.py) and catalog invariants (catalog.py).