zob-compute-profile

name	zob-compute-profile
description	Use when selecting, previewing, validating, or implementing ZOB compute/effort profiles: auto, low, medium, high, xhigh, max.

ZOB Compute Profile Skill

Purpose

Compute profiles control how much workflow effort ZOB spends on a task. They are hard policy/cap inputs, not vague encouragement to “try harder”.

Canonical plan:

• docs/ZOB_COMPUTE_PROFILE_ROUTING_PLAN.md

Profiles

auto   → preview then resolve to low/medium/high/xhigh/max
low    → single-agent or deterministic fast path
medium → small supervised workflow
high   → multi-lane workflow with validation/oracle
xhigh  → extra-high expert workflow with richer benchmarks and adversarial review
max    → maximum bounded supervised workflow; requires strict budget/human gates

TODO/delegation behavior by profile

• low: single-agent/best-effort. Do not spend extra turns on recursive split or background child orchestration; return incomplete/blocked if the box is too small.
• medium: small supervised workflow. A child may suggest a split in final output, but live/background escalation is not the default.
• high: multi-lane parent-owned workflow. Use active-session background delegation when useful, and let TODO-linked children return TODO_SPLIT_REQUEST.v1 early when the TODO is too broad for their context.
• xhigh: same as high with deeper bounded decomposition and stronger validation/oracle.
• max: same as xhigh, but strict budget/human/oracle gates are mandatory before scale-up.

Even at high/xhigh/max, child agents must not directly spawn children or mutate parent TODO state; they request split/replan and the parent applies it.

Required behavior

Before using compute to justify more work:

1. Run or inspect a compute preview when requested_profile=auto.
2. Resolve to an effective profile and hard caps.
3. Keep child dispatch parent-owned; child direct dispatch remains forbidden.
4. Keep safety, budget, path, and sandbox gates stronger than profile preference.
5. Persist only metadata/hash-safe artifacts; do not persist raw task/prompt/output bodies.
6. Treat max as approval-gated; do not auto-select it silently.

Native surfaces

• Slash commands:
  • /compute
  • /effort
• Runtime tools:
  • zob_compute_preview
  • zob_compute_resolve_profile
  • zob_compute_plan_workflow
  • zob_compute_validate_profile
  • zob_compute_write_profile_reports
• Deterministic scripts:
  • npm run preview:compute-profile:project-dna-smoke
  • npm run validate:compute-profile:project-dna-smoke
  • npm run plan:compute-workflow:project-dna-smoke
  • npm run validate:compute-workflow:project-dna-smoke
  • npm run snapshot:compute-profile:project-dna-smoke
  • npm run smoke:compute-profile-regression
• Policy files:
  • .pi/compute-profiles/defaults.json
  • .pi/compute-profiles/overrides.json
  • .pi/compute-profiles/risk-rules.json
  • npm run validate:compute-profile-policy
• Plan/doc:
  • docs/ZOB_COMPUTE_PROFILE_ROUTING_PLAN.md

No-ship rules

Block profile readiness if any occur:

• requested/effective profile mismatch is unexplained;
• caps are missing for a live or agentic workflow;
• child agents exceed maxAgents, maxParallel, or depth caps;
• strict budget is required but absent;
• oracle is required but missing;
• max is used without human/scale approval;
• profile is used to bypass secrets/path/sandbox/write policy;
• compute artifacts contain raw body/prompt/output/content/diff/patch text;
• telemetry or validation evidence is missing.

ProjectDNA mapping

ProjectDNA should use compute profile as stage selection metadata:

low    → scan + scan validation
medium → scan + validation + capsules + sample spec + one query
high   → medium + quarantine sample + sample validation + benchmark + oracle
xhigh  → high + specialist lanes + richer query suite + adversarial review
max    → xhigh + multi-reference/symbol/callgraph/promotion packet gates

Reference projects remain read-only, generated sample output remains quarantine/proposal-only, and external knowledge-backend import/sync/embed/write remains disabled unless explicitly approved.

Final response shape

<result>effective profile/caps or implemented surface</result>
<evidence>artifact paths and validation commands</evidence>
<risks_blockers>remaining gaps/no-ship risks</risks_blockers>
<compliance>profile did not bypass safety/budget/oracle gates</compliance>