budget-term-sheet

name: budget-term-sheet description: Build a pre-deploy Rockie GPU budget term sheet before any Rocky-originated experiment submit. Trigger words "/budget-term-sheet", "quote the GPU budget", "show me the term sheet", or any workflow that is about to call /experiment for GPU / torch / triton / training / weight-download work.

/budget-term-sheet

Render the user-facing cost term sheet that Rocky must show before a Rocky-originated POST $ROCKIELAB_API_URL/api/jobs/submit call. This skill is pre-submit UX and approval gating only. It does not replace platform-context's credit reservation, day caps, or hard budget enforcement.

When to invoke

• Before /experiment submits a GPU / torch / triton / training / weight-download job.
• When the user asks for a dry-run quote without launching compute.
• When a user changes the approved budget and Rocky needs to re-render the quote.

Hard rules

• Rocky-originated GPU submits go through this skill before runtime/submit.py.
• Only explicit approval tokens count: approve, approved, or I approve.
• A user-modified budget below estimate_cents is non-submittable in this repo.
• Non-available market states never submit.
• Keep the user-facing provider deidentified: Rockie GPU, plus SKU/count/region/tier.
• Do not dispatch real compute here. Quote and approval only.

Procedure

1. Assemble a job spec with job_shape, gpu_type, gpu_count, region, tier, and either timeout_seconds or wallclock_minutes. Use region: "us" and tier: "spot" unless the user or workflow requires a different visible capacity slice.
2. Run scripts/quote_term_sheet.py to build the canonical JSON term-sheet object. Prefer live quote endpoints when credentials exist; use --market-json for tests and dry runs; use --allow-heuristic-dry-run only when the user explicitly wants a heuristic quote.
3. Render the markdown block with scripts/render_term_sheet.py. The block must show: total estimate, stage breakdown, recommended budget, wall-clock, Rockie GPU compute summary, availability/confidence, and the explicit Approve / Modify / Cancel prompt.
4. Wait for a user reply. Parse it with scripts/parse_approval.py.
5. Decision handling:
   • approve: keep the current budget and mark the term sheet approved.
   • modify: re-render with the new user_budget_cents and ask again.
   • modify_then_approve: re-render with the new budget; only submit if the modified budget is still >= estimate_cents.
   • cancel: abandon the submit.
   • clarify: ask again; do not submit.
6. Hand the final approved JSON artifact to /experiment/runtime/submit.py through --term-sheet-json, passing matching --region and --tier. The submit helper refuses artifacts or CLI requests that omit or change the approved region/tier. Do not shell raw curl POST /api/jobs/submit.

Commands

Quote JSON:

python3 ${SKILL_DIR}/scripts/quote_term_sheet.py \
  --job-spec-json /tmp/job-spec.json \
  > /tmp/term-sheet.json

Render markdown:

python3 ${SKILL_DIR}/scripts/render_term_sheet.py \
  --term-sheet-json /tmp/term-sheet.json

Parse the user's reply:

python3 ${SKILL_DIR}/scripts/parse_approval.py \
  --estimate-cents 20000 \
  --reply "cap it at $250 and I approve"

Canonical object

The term-sheet JSON must carry these fields:

• quote_id
• job_shape
• availability
• status_code
• status_reason
• quote_source
• generated_at
• compute
• stages
• estimate_cents
• recommended_budget_cents
• user_budget_cents
• wallclock_minutes
• confidence_bucket
• confidence_statement
• decision

Use references/template.md as the user-facing block shape.