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 skills/icesixgod/codex-trajectory/inspect-codex-trajectorynpx skills add icesixgod/codex-trajectory --skill inspect-codex-trajectorygit clone --depth 1 https://github.com/icesixgod/codex-trajectoryWrote 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/icesixgod/codex-trajectory/inspect-codex-trajectory)<a href="https://agentmods.dev/skills/icesixgod/codex-trajectory/inspect-codex-trajectory"><img src="https://agentmods.dev/badge/skills/icesixgod/codex-trajectory/inspect-codex-trajectory.svg" alt="Measured on agentmods" 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 | $0.00078 | $0.00758 |
| Opus 5 | $0.00039 | $0.00379 |
| Sonnet 5 | $0.00016 | $0.00152 |
| Haiku 4.5 | $0.00008 | $0.00076 |
Grade A, and why
inspect-codex-trajectory 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 5d 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.
What it actually says
Inspect Codex Trajectory
Use the plugin's read-only MCP tools instead of opening raw files under the Codex home directory.
- For the active task or the most recently updated task, call
show_codex_trajectorywithout a session ID. Keep the defaultdetailLevel: summary; this renders the interactive trajectory UI without tool inputs or outputs. - For a historical task, call
list_codex_sessions, choose the exact session ID, then callshow_codex_trajectorywith that ID. - Use
get_codex_trajectorywhen the user wants analysis without the interactive UI. Cite turn numbers and record indexes from its structured result. If the requested evidence predates the returned page, call it again with the response'spagination.nextBeforeRecordasbeforeRecord; continue only as far back as the request requires. - When the user wants a live or picture-in-picture view, call
show_codex_trajectoryin summary mode. The user can then click Live window in the rendered viewer. Inside Codex it fills the supported host side-panel presentation with frozen whole-task Token totals, a remaining-quota badge when official windows are present, one-click and opt-in threshold stop controls, and a chronological per-record Token stream; a regular Chromium page uses native video PiP without stop controls. Stop controls require the user's explicit experimental loopback CDP opt-in. The standard Apps surface uses an app-private direct-stop tool that pauses an active Goal beforeturn/interrupt, never posts a follow-up message or enters the steering queue, and fails closed while CDP is unavailable. Manual Stop rearms for later turns; automatic Stop remains latched throughout the same quota cycle, including/goalcontinuations, and rearms only after quota recovery, a window reset, or a guard configuration change. A rejected stale turn is rebound once and transient failures retry without consuming the latch. Selecting another trajectory disables stopping. On macOS and Windows, the token-bearing View trajectory Browser shortcut additionally requires per-connection operating-system authentication of the Codex desktop peer; failures and unsupported platforms keep it unavailable without disabling direct stop. The live window does not launch a standalone app or touch composer drafts and attachments. Do not call the app-only live-update, toolbar-setting, or direct-stop helpers directly. - Use
detailLevel: fullonly after the user explicitly asks to inspect full tool input or output. Treat those values as potentially sensitive and quote no more than the task requires. Never exposesession_meta.base_instructions, encrypted reasoning, credentials, or environment variables. - Explain that Codex logs do not expose DeepSeek Harness step boundaries directly. The plugin starts a new approximate step when model output resumes after one or more tool results.
- Treat read warnings as evidence of skipped malformed records. If a paginated lineage fails validation, report that the local task history is inconsistent; do not bypass the plugin by opening inherited rollout files directly.
The user-facing trajectory tools are read-only. They read local sessions/ and, when requested, archived_sessions/ beneath CODEX_HOME or ~/.codex. The viewer alone may use its private CDP setting and direct-stop helpers after the user changes the experimental switch or activates an enabled stop control.
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.
- 5d ago First seen · 19 lines · 78 tokens per session scan A 29da5b70032d
inspect-codex-trajectory is a skill published in the GitHub repository icesixgod/codex-trajectory (245 stars, last pushed today), licensed MIT. It adds 78 tokens to every session and 758 once invoked, about $0.0004 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 2026-08-30.
Other skills, from other repositories
matlab-integrate-pytorch-vision
Creates MATLAB interfaces to Python image processing and computer vision models from GitHub repositories or pip-installable packages using MPyReq. Use when asked to interface MATLAB with a Python CV/image model (segmentation, depth estimation, object detection, image generation, super-resolution, etc.), given a GitHub…
pytorch-patterns
PyTorch deep learning patterns and best practices for building robust, efficient, and reproducible training pipelines, model architectures, and data loading.
cuml-machine-learning
Use for GPU-accelerated machine learning on tabular data using NVIDIA cuML. Triggers when tasks involve classification, regression, clustering, dimensionality reduction, or model training on datasets.
optimize-for-gpu
GPU-accelerates scientific Python on NVIDIA hardware and verifies that the result is correct and faster. Use for CUDA/GPU optimization; CPU-bound NumPy, SciPy, pandas, scikit-learn, NetworkX, scikit-image, vector-search, image-processing, graph, simulation, or file-I/O workloads; CuPy, cuDF, cuML, cuGraph, cuVS…
hf-gradio
Use Gradio applications via API. Use when the user asks for to generate a prediction from a Gradio app on Hugging Face spaces or public URL. For example, "Generate an image using black-forest-labs/FLUX.2-dev".
pydantic-ai
Build production-ready AI agents with PydanticAI — type-safe tool use, structured outputs, dependency injection, and multi-model support.