Env Hub

scripts/capture.py (in this skill's folder; Python 3.9+, standard library only) does every step. Run it with the absolute path to this skill's folder.

Save

First read the user's messages in the session, numbered:

python3 <skill>/scripts/capture.py messages --from claude --json   # or: --from codex

It reads the newest session that ran in the current folder (--session ID or --transcript FILE for another; sessions --from claude lists them).

Choose the messages that state the task: the request, and later messages that correct it or add to it. Leave out what is not the task: asking to save or submit, thanks, questions about this skill, and a different task in the same session (save that one on its own). taken_by_default marks what save takes without a choice: every message before the request to save. Then:

python3 <skill>/scripts/capture.py save --from claude --session ID --messages 1,3 --json

--messages takes numbers and ranges (1,3, 2-4). The instruction is the chosen messages word for word: choose, never rewrite. The conversation kept runs from the first chosen message through your replies to the last. A task that starts at a later message starts from the last commit before it, so earlier work in the session is there only if it was committed; say so when it matters. The folder must be a git repository. Nothing leaves the machine and the repository is not changed.

Beside the task folder it keeps the conversation (what the user and the agent said, without the tools' output, anything that looks like a secret redacted) and a bundle of both, the zip Env Hub takes.

Show the user the summary: the instruction, which of their messages were taken and which left out and why (messages, messages_left_out), how many files the workspace keeps, what was left out and why (files_omitted), every file left out because it looks like a secret (secrets_left_out), and that the conversation goes with the task (conversation_turns). To leave out files, run save again with --exclude PATH (repeatable); to change what the task says, again with other --messages.

• Never edit the saved task's files.
• If save says the chosen messages hold what looks like a secret, stop and tell the user; do not work around it.

Submit (only after the user says yes)

Submitting uploads the bundle (the task and the conversation) to Env Hub under the user's Eigent account, as the contribution agreement they accept says, and starts converting it into an environment. Ask first, every time; say what will be uploaded (the bundle and its size). Then:

python3 <skill>/scripts/capture.py submit <task folder> --json

The first time, the user signs in: run capture.py login, which opens their browser (Eigent account, Google or GitHub) and asks them to accept the contribution agreement and allow the skill. Never handle their password or token; the script keeps the token in ~/.config/env-hub/.

Follow it

python3 <skill>/scripts/capture.py status <conversion id> --json
python3 <skill>/scripts/capture.py list --json     # the user's environments

status: queued, running, succeeded (environment_url is where the user sees the environment in full, verifier and solution included) or failed (error.message says why). Until then the user follows it on the web at account_url, which submit and status return; give that address, never one of your own (the hub may not be the default one). stage says more: checking, publishing, drafting (Env Hub writes the verifier and reference solution), gating (oracle and nop runs), repairing, passed, rejected, needs_attention. Drafting and gating take minutes; check back rather than waiting in a loop.

Errors you can act on: "not signed in", an expired connection or an agreement to accept → login; the day's limit → try later; the task too large → save again with --exclude for large folders. The user removes Env Hub's access from their account page on the hub.