Remove (or neutralise) the Step-2c 100 GPU-h autonomous plan-approval gate

Overview

User directive (chat 2026-07-28, verbatim): "remove the 100 GPU-h gate from the workflow entirely", following "tell it to bypass the 100 GPU-h parking."

The Step-2c plan-approval GPU-hour cap is the autonomous gate that parks a plan at plan_pending when gpu_hours_total exceeds EPM_PLAN_AUTOAPPROVE_GPU_HOURS (default 100, code default 24 in task.py). It is enumerated in .claude/workflow.yaml § gates and in CLAUDE.md § Auto-continuation policy as inline gate #4.

This is an ARCHITECTURAL / public-contract change (removes a subsystem + a documented gate from the canonical gate enumeration), so the planner MUST set architectural: true and this task MUST park at plan_pending for the user's explicit greenlight on the final plan — even though the user has already directed the change in principle. The user reviews the concrete diff, not just the intent.

Goal

Remove — or, if the planner's analysis favours it, neutralise by raising the default — the Step-2c GPU-hour auto-approve cap, so an autonomous /issue session never parks a plan on GPU-hour grounds. Deliver whichever option the critic panel judges correct, with the reasoning recorded.

Two consequences the user has NOT yet weighed — surface these explicitly at plan time

1. The cap is the ONLY cost gate in the pipeline. CLAUDE.md: "cost is gated ONLY at the Step 2c plan-approval GPU-hour cap, never mid-run", and dollar caps are forbidden outright (tests/test_no_dollar_budget_caps.py). Removing it leaves NO automated cost control on any autonomous task — a planner mis-estimate of 10,000 GPU-h would auto-approve and launch.
2. The same gate is the ARCHITECTURAL-GREENLIGHT surface. .claude/rules/workflow-fix-on-bug.md § Architectural greenlight routes architectural / public-contract changes through "the /issue plan-approval gate (park at plan_pending)". If the gate is removed wholesale rather than made GPU-hour-blind, architectural workflow changes would auto-approve too — including future workflow-fix sessions editing the workflow surface itself. The architectural: true branch must survive any change here.
3. The blank-estimate fail-safe is part of the same predicate. _plan_gate_decision currently parks when gpu_hours is None — "FAIL SAFE: a missing/None gpu_hours parks (never auto-approves on a blank estimate)". That is a correctness guard against a planner that failed to estimate at all, not a cost control. Dropping it silently would let an unestimated plan run.

Options for the planner to weigh (do not pre-commit)

• A — raise the default (e.g. 100 → 1000/∞ sentinel) leaving the mechanism, the architectural: true branch, and the blank-estimate fail-safe intact. Smallest blast radius; satisfies the user's operational intent (never parks in practice).
• B — make the gate GPU-hour-blind but keep it firing on architectural: true and on a missing estimate. Removes the cost dimension only.
• C — remove the gate entirely as literally directed, accepting 1-3 above.

The planner should recommend one and state why; the critic panel gates it; the user approves the concrete diff at plan_pending.

Surfaces (grep-enumerated at filing time, 2026-07-28)

Implementation: scripts/task.py (_plan_gate_decision, --auto-approve-if-autonomous), scripts/spawn_session.py (--auto-approve-gpu-hours flag → EPM_PLAN_AUTOAPPROVE_GPU_HOURS env threading, incl. the per-child campaign cap at :2342/:2406 and the respawn re-pass at :3292), scripts/autonomous_session_watch.py, scripts/file_infra_task.py, scripts/workflow_lint.py, .claude/settings.json, .claude/workflow.yaml § gates.

Docs: CLAUDE.md § Auto-continuation policy (inline gate #4) and § PM/spawn kickoff rule; .claude/skills/issue/SKILL.md Step 2c; .claude/skills/campaign/SKILL.md (per-child cap); .claude/rules/workflow-fix-on-bug.md § Architectural greenlight.

Tests that pin current behaviour and must be updated coherently (not deleted): tests/test_autonomous_plan_gate.py (primary), tests/test_workflow_yaml.py, tests/test_workflow_lint.py, tests/test_autonomous_session_watch.py, tests/test_spawn_session_*.py, tests/test_issue_tick_skill.py, tests/test_stalled_detector_and_gc.py.

Note .claude/worktrees/issue-779/CLAUDE.md also matched the grep — that is a stale worktree copy, NOT a surface to edit.

Constraints

• Workflow-surface only. scripts/workflow_lint.py (no-flags) passes; ruff clean on touched files; uv run pytest green with tests UPDATED to the new contract rather than removed.
• .claude/workflow.yaml § gates and CLAUDE.md § Auto-continuation policy must stay mutually consistent — workflow_lint --check-asks resolves every AskUserQuestion to a documented gate, so removing a gate key without updating its citers will FAIL the lint.
• Keep the interactive path intact: a non-autonomous session returns interactive_pending today and that behaviour is out of scope unless the planner argues otherwise.

Rules to read

.claude/rules/workflow-fix-on-bug.md (§ Architectural greenlight), .claude/workflow.yaml § gates, CLAUDE.md § Auto-continuation policy + § Halt-criterion contract, .claude/rules/agents-vs-skills.md.