Getting it into your agent
It runs from inside its repository, so the clone comes first — what it calls does not travel with the file alone.
git clone --depth 1 https://github.com/wan-huiyan/agent-traffic-controlnpx agentmods add skills/wan-huiyan/agent-traffic-control/multi-phase-skill-disk-reading-strategyWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/wan-huiyan/agent-traffic-control/multi-phase-skill-disk-reading-strategy)<a href="https://agentmods.dev/skills/wan-huiyan/agent-traffic-control/multi-phase-skill-disk-reading-strategy"><img src="https://agentmods.dev/badge/skills/wan-huiyan/agent-traffic-control/multi-phase-skill-disk-reading-strategy/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/wan-huiyan/agent-traffic-control/multi-phase-skill-disk-reading-strategy"><img src="https://agentmods.dev/badge/skills/wan-huiyan/agent-traffic-control/multi-phase-skill-disk-reading-strategy.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What 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 5.1 | $0.00157 | $0.01480 |
| Opus 5 | $0.00078 | $0.00740 |
| Sonnet 5 | $0.00031 | $0.00296 |
| Haiku 4.5 | $0.00016 | $0.00148 |
Grade A, and why
multi-phase-skill-disk-reading-strategy scanned grade A with 0 findings 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 9d 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.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 151 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Multi-Phase Skill: Disk-Reading Strategy for Late-Pipeline Subagents
Sister skill:
multi-agent-skill-silent-phase-compressioncovers the output-direction symmetric problem (subagent outputs flooding back into orchestrator → silent compression of mandatory phases). Same fix family (file-based state passing), opposite direction. If a multi-agent skill is failing, read both — the bug is usually one or the other (sometimes both).
Problem
In Claude Code skills with many phases (10+), late-pipeline subagents silently fail to execute when the orchestrator's context window is near capacity. The orchestrator tries to inject large amounts of structured data (earlier phase outputs) into the subagent's prompt, but context exhaustion causes the launch to either fail silently or produce a degraded prompt that the subagent can't follow properly.
Symptoms:
- Earlier phases complete successfully but the final phase's output is missing
- No error reported — the orchestrator skips to the completion message
- When the user manually asks for the missing output, the orchestrator produces a generic version from memory rather than following the spec
- The failure is intermittent: works on short inputs, fails on long ones
Context / Trigger Conditions
- A Claude Code skill runs 10+ sequential phases with subagents
- A late phase (e.g., Phase 15 of 16) needs data from earlier phases
- The orchestrator currently INJECTS that data into the subagent prompt (embedding 500+ lines of structured data or process history)
- The earlier phases already WROTE their output to disk (.md, .json, etc.)
- The failure appears as a silent skip — no error, just missing output
Solution
1. Make late phases sequential, not parallel
If Phase N-1 and Phase N currently run in parallel, make them sequential. Phase N runs AFTER N-1 so that N-1's output file exists on disk.
Latency impact is usually negligible — orchestrator-assembled phases (no subagent) complete in seconds.
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.
- 9d ago First seen · 151 lines · 157 tokens per session scan A b18285b880ae
multi-phase-skill-disk-reading-strategy is a skill published in the GitHub repository wan-huiyan/agent-traffic-control (3 stars, last pushed 4d ago), licensed MIT. It adds 157 tokens to every session and 1,480 once invoked, about $0.0008 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
cloudflare-workers-observability
Cloudflare Workers observability with logging, Analytics Engine, Tail Workers, metrics, and alerting. Use for monitoring, debugging, tracing, or encountering log parsing, metric aggregation, alert configuration errors.
cloudflare-workers-dev-experience
Cloudflare Workers local development with Wrangler, Miniflare, hot reload, debugging. Use for project setup, wrangler.jsonc configuration, or encountering local dev, HMR, binding simulation errors.
cloudflare-workers-performance
Cloudflare Workers performance optimization with CPU, memory, caching, bundle size. Use for slow workers, high latency, cold starts, or encountering CPU limits, memory issues, timeout errors.
api-error-handling
Implements standardized API error responses with proper status codes, logging, and user-friendly messages. Use when building production APIs, implementing error recovery patterns, or integrating error monitoring services.
skeptical-triage
Reusable 3-round self-challenge + arbiter pattern for filtering false positives from findings/verdicts. Use when the cost of a false-positive gate block exceeds the cost of 4 extra LLM turns.
anti-patterns
Catalogue of known SDLC anti-patterns that greatcto agents must actively reject when reviewing architecture, plans, code, or post-mortems. Used by architect (pre-impl), pm (planning), senior-dev (impl), l3-support (post-incident).