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 skills add wenhaochai/claude-plugins --skill extract-computation-graphgit clone --depth 1 https://github.com/wenhaochai/claude-pluginsWrote 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/wenhaochai/claude-plugins/extract-computation-graph)<a href="https://agentmods.dev/skills/wenhaochai/claude-plugins/extract-computation-graph"><img src="https://agentmods.dev/badge/skills/wenhaochai/claude-plugins/extract-computation-graph/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/wenhaochai/claude-plugins/extract-computation-graph"><img src="https://agentmods.dev/badge/skills/wenhaochai/claude-plugins/extract-computation-graph.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 | $0.00162 | $0.00994 |
| Opus 5.5 | $0.00065 | $0.00398 |
| Sonnet 5.5 | $0.00032 | $0.00199 |
| Haiku 4.5 | $0.00016 | $0.00099 |
Grade A, and why
extract-computation-graph 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 14d 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 — 58 lines — stays where its author put it; the contents beside it link to each section on GitHub.
extract-computation-graph
In: the job's code, at a known commit.
Out: records/computation_graph.json and records/execution_schedule.json in the workspace’s records/ directory.
This runs nothing. It records what you expect the job to do. The profile records what it actually did, and you read one against the other.
Two files, and what each is for
records/computation_graph.json |
records/execution_schedule.json |
|
|---|---|---|
| says | what is computed | when and where it is computed, and what moves between devices |
| lets you attribute | a measured kernel back to the module, layer or op that issued it, times its repeat count | a measured idle gap back to the copy, collective or sync point behind it |
| lets you check | whether the cost sits where the structure says the work is | whether what the code meant to overlap actually did |
The fields below are examples. Add what this job needs.
records/computation_graph.json
| field | why it is there |
|---|---|
job_type: training, inference serving, or batch |
each has a different skeleton. Do not assume training |
commit |
every file:line below means nothing without the sha it was read at |
entrypoint |
the step function, as a file:line |
modules, down to op level (model -> modules -> layers -> ops), each naming its enclosing parent and its execution order |
findings land on ops, not on files |
shapes on the boundaries you can read |
shape decides which kernel gets picked and how much memory the tensor takes |
code_anchor on every entry, written model/dit.py:DiTBlock.forward:214 |
this is where a root cause has to land later. Symbol before line, so it survives an edit above it |
repeat on every unit that repeats: the unit and its count |
a cost inside a unit multiplies by N. Rank by per-unit cost x N, never by one instance: forty blocks turn a millisecond into forty |
records/execution_schedule.json
| field | what to record |
|---|---|
phases |
startup, data, forward, backward, optimizer, periodic eval or checkpoint, teardown. Each with a kind of one-off, repeated or periodic, and who it runs_on |
data_path |
every hop from disk to device: worker processes, collation, pinned staging, the copy itself. Mark which hops are asynchronous and which block the step |
host_device |
the H2D and D2H copies, whether their source is pinned or pageable, and every sync point: .item(), .cpu(), .numpy(), .tolist(), torch.cuda.synchronize(), printing a tensor, any Python branch that reads one |
device_device |
the collectives, each tagged with the parallelism it serves (data, tensor, pipeline, expert, sequence) and whether the call blocks or was issued with async_op=True |
compute |
forward, backward, which blocks recompute under activation checkpointing, gradient accumulation, the optimizer step, and the torch.compile regions with the graph breaks between them |
overlap |
what is supposed to run at the same time as what, and on which stream |
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.
- 14d ago Changed · +162 tokens per session 654c9d50c52a
- 17d ago First seen · 58 lines · 0 tokens per session scan A 8ca75c9cd1e4
extract-computation-graph is a skill published in the GitHub repository wenhaochai/claude-plugins (17 stars, last pushed 2d ago), licensed MIT. It adds 162 tokens to every session and 994 once invoked, about $0.0006 per session on Opus 5.5. 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 2026-10-01.
Other skills, from other repositories
trace
Evidence-driven tracing lane that orchestrates competing tracer hypotheses in Claude built-in team mode.
smiles-validation
Strict SMILES validation, structural comparison, and modification verification. Catches invalid LLM-generated molecules.
nonlinear-solvers
Select and configure nonlinear solvers for f(x)=0 or min F(x). Use for Newton methods, quasi-Newton (BFGS, L-BFGS), Broyden, Anderson acceleration, diagnosing convergence issues, choosing line search vs trust region, and analyzing Jacobian quality.
linear-solvers
Select and configure linear solvers for systems Ax=b in dense and sparse problems. Use when choosing direct vs iterative methods, diagnosing convergence issues, estimating conditioning, selecting preconditioners, or debugging stagnation in GMRES/CG/BiCGSTAB.
numerical-stability
Analyze and enforce numerical stability for time-dependent PDE simulations. Use when selecting time steps, choosing explicit/implicit schemes, diagnosing numerical blow-up, checking CFL/Fourier criteria, von Neumann analysis, matrix conditioning, or detecting stiffness in advection/diffusion/reaction problems.
social-physics-filter
Unified boundary enforcement, interpersonal diagnostic, and relational audit engine. Absorbs 40 psychology + 2 social protocols and all relationship case studies.