aragora: Skill for Claude Code

Elves (aragora-native)

This is an aragora-bound fork of the Elves autonomous-development skill. The unattended Ralph Loop (try → check → feed back → repeat across sprint-sized batches, surviving context compaction) is unchanged. What changes is the gate. Upstream Elves closes a batch on npm test plus a GitHub PR comment. In aragora, "the tests are the watch" is necessary but not sufficient: a batch is only complete when it has passed aragora's own evidence-first governance.

Positioning. aragora's README states it "does not sell lights-out autonomy as the default story" and "does not advance work without evidence, review, and clear terminal states." This fork honors that. It keeps Elves' overnight leverage but subordinates it to receipts, model quorum, and human settlement. The elves work the night shift; aragora decides what is allowed to land.

When to use

Trigger inside ~/development/aragora (or any aragora checkout) when the user wants a long, unattended run: "run overnight", "I'm going offline until 8", "implement this plan end-to-end", "keep going without me", "do not stop". For one-off single edits, do not use this skill.

The two-call handoff (unchanged from Elves)

1. Stage the run — clean the plan, generate the survival guide / learnings / execution log from the templates in references/, claim a dedicated branch + worktree, run preflight, and stop. See references/kickoff-prompt-template.md.
2. Launch in a fresh call with a short, behavior-heavy prompt.

The aragora validation gate (the core difference)

Every batch runs through references/validation-gate-aragora.md. In short, a batch is not complete until all of the following hold, in order:

1. Local truth. Project validation passes against current HEAD: pre-commit run --all-files, mypy (no new errors above .mypy-baseline), and the relevant pytest slice. Total test count must never decrease.
2. Adversarial review = a debate, not a comment. Run a heterogeneous-model adversarial review of the batch diff via aragora itself (e.g. aragora ask framed as a review of git diff <default-branch>...HEAD, or the repo's aragora-review-gate path). Capture the exact head SHA reviewed, the model/provider families, independence from the authoring lane, the recommendation, and any dissent. Unresolved dissent blocks the batch.
3. Receipt. Produce a DecisionReceipt for the batch decision and verify it: aragora verify <receipt.json> (and/or aragora receipt verify). The receipt binds the decision to the reviewed SHA. No receipt → batch not closed.
4. Tier classification + settlement. Classify the batch against docs/REVIEW_AUTHORITY_PRINCIPLES.md (Tier 0-4):
   • Tier 0-2 (docs/tests, additive internal, live automation/CLI/observability): a green, receipt-backed model quorum with no unresolved dissent is sufficient to record settlement autonomously and continue. Still never merge by default — settlement records the authorization packet; the human or a recorded merge-on-green preference plus helper authorization does the merge. §Conductor never grants merge authority by itself.
   • Tier 3-4 (semantic correctness, persistence, security/RBAC/auth, public API/SDK, migrations; or secrets/deployment/workflow policy/destructive ops/merge-authority self-modification): HARD STOP. The model quorum prepares the packet, but the run must pause and require explicit human risk acceptance (the aragora/human-settlement signal) before that batch is considered landed. Continue with other non-blocked batches if any exist; otherwise checkpoint and wait.