Phinneas

run-engagement

Run one draft-review cycle for an ESRS E1 disclosure and stop for consultant approval.

Phinneas 0 Updated 4h ago
GitHub

Install

npx skillscat add phinneas/esrs-e1/run-engagement

Install via the SkillsCat registry.

SKILL.md

Run engagement

You cannot list folders. Read files by name.

  1. Read client.yaml. If company.legal_name or company.reporting_period is empty, stop and say which fields are missing.

  2. Read inputs/company.yaml and inputs/emissions.yaml. If either is missing, stop.

  3. Run the drafter:
    gitagent --dir agents/drafter -p "Draft the ESRS 2 and E1 disclosure. Follow skill draft-disclosure."

  4. Read outputs/coverage/E1-11.csv. It is written last, so if it exists the draft is complete.
    If it is missing, run the drafter again with
    -p "Resume the draft. Follow skill draft-disclosure and skip DRs whose coverage file is complete."
    Try at most 2 resumes, then stop and report what is missing.

  5. Run the reviewer:
    gitagent --dir agents/reviewer -p "Review round 1. Follow skill review-disclosure."

  6. Read review/findings.csv. Count open findings by severity, separately for check = data_gap and everything else.

  7. If any open finding with check other than data_gap is blocking or major, and rounds used is below review.max_rounds:

    • run the drafter with -p "Revise against open findings in review/findings.csv. Follow skill revise-disclosure."
    • run the reviewer with -p "Review round N. Follow skill review-disclosure."
    • repeat from step 6.

    Do not run another round only for data gaps. Rewording cannot fix missing data.

  8. Stop. Report to the consultant:

    • rounds run
    • open draft problems by severity, with requirement IDs for every blocking one
    • the number of data gaps, and that the list is in review/findings.md under "Data the client must supply"
    • the exact files to read: outputs/disclosure.md, review/findings.md
    • that approval is a git commit of approvals/<period>.md from the consultant's own account

You do not set time limits. The guard sets each sub-agent run to one hour.

Prompts passed with -p must not contain quotes, $, backticks, ;, &, |, <, >, or brackets. The guard blocks them.