ops: verify-features drive loop on Grok — plan, directive, launcher variant

Closed
KH
2mo ago
+621-11

What

Stands up an unattended drive loop aimed at every feature in flows verifiable, then autonomous merge, running on the xAI/Grok credit pool.

  • ops/VERIFY-FEATURES-PLAN.md — measured starting state (no branch protection, path-filtered CI, zero flows in CI, no manifest, no coverage), the four verify tiers, and seven ordered work packages WP-V1 … WP-V7 with runnable definitions of done and blocked-by. Mirrors relay's verify-features shape (manifest + audit + checks.jsonl / verdict.json).
  • ops/DIRECTIVES.md — the standing directive the assess step honors before the backlog.
  • ops/gen-drive-cloud.py --cli <cli> --out <file> — emits same-steps variants of the cloud loop. workflows/drive-cloud-grok.yaml is generated; asserted equal to drive-cloud.yaml apart from cli / name / channel.
  • ops/launch-gate.sh DRIVE_WORKFLOW and ops/autodrive.sh AUTODRIVE_BRIEF_FILE — env overrides so this loop coexists with the default one. ops/AUTODRIVE_BRIEF-VERIFY.md is its brief.

What it does not do

Nothing here merges anything or touches a gate. The loop delivers PRs via ops/deliver-run.sh; humans merge until WP-V7 is enabled by an RFC-0001 #16 amendment (a human step, called out in the plan).

Evidence

$ node --test ops/*.test.mjs
# tests 64
# pass 64
# fail 0

$ python3 -c '... yaml.safe_load both, swap cli/name/channel ...'
identical apart from cli/name/channel: True

🤖 Generated with Claude Code

https://claude.ai/code/session_0166HKhhZVRRp81PTPeLw43E

[!NOTE]
Low Risk
Documentation and ops launcher/generator changes only; no runtime, CI, or merge-behavior changes in this diff.

Overview
Adds an ops-only program to drive the repo toward full feature verification and eventual autonomous merge, without implementing product features or changing merge gates in this PR.

ops/VERIFY-FEATURES-PLAN.md documents the measured CI/testing gap, four verify tiers (unit → cloud flow), what counts as a tier-2 flow test, and seven ordered work packages WP-V1 … WP-V7 with runnable definitions of done, blocked-by links, and explicit human steps (branch protection, RFC amendment for auto-merge). ops/DIRECTIVES.md elevates that plan as the assess step’s top priority until WP-V7 lands.

Autodrive / cloud launch can run a second loop in parallel: ops/AUTODRIVE_BRIEF-VERIFY.md scopes ticks to the verify plan; AUTODRIVE_BRIEF_FILE in ops/autodrive.sh selects which brief file to read; DRIVE_WORKFLOW in ops/launch-gate.sh chooses which workflow YAML to agent-relay cloud run. ops/gen-drive-cloud.py now accepts --cli, --model, --suffix, and --out to emit same-step workflow variants (agents’ CLI/model and swarm name/channel only); workflows/drive-cloud-grok.yaml is the generated opencode + xai/grok-4.6 variant for the Grok credit pool.

Explicitly out of scope here: no changes to review-swarm workflows, no CI/test harness implementation (those are later WPs), and no autonomous merging.

Reviewed by Cursor Bugbot for commit 824fc56e012d8dca8f66136dd8e51a18ede15920. Bugbot is set up for automated code reviews on this repo. Configure here.

Summary by cubic

Sets up an unattended drive loop that works toward making every feature verifiable and eventually enabling autonomous merge, running on the Grok credit pool without changing any merge behavior—humans still merge.

Details

  • ops/VERIFY-FEATURES-PLAN.md adds a measured starting state, four verify tiers, and seven ordered work packages (WP-V1 … WP-V7) with runnable definitions of done and blocked-by links; it defines a flow test as deterministic assertions over a run containing real agent steps and narrows tier 2 to five agent-path cases.
  • ops/DIRECTIVES.md gains a standing directive that makes this plan the top priority for the assess step.
  • ops/gen-drive-cloud.py now accepts --cli and --out to generate same-step variants of the cloud loop; workflows/drive-cloud-grok.yaml is generated from it and uses opencode on xai/grok-4.6 because the v1 engine's CLI registry has no grok entry.
  • ops/launch-gate.sh and ops/autodrive.sh accept DRIVE_WORKFLOW and AUTODRIVE_BRIEF_FILE env overrides so this loop can coexist with the default one; ops/AUTODRIVE_BRIEF-VERIFY.md is its brief.
  • Nothing in this PR merges anything or touches existing gates—autonomous merge is scoped to a later work package requiring an RFC amendment and a human step.