yaj-a

Autonomous development ritual with role-based coordination. Loops until complete using specialized agents (hotṛ→research, adhvaryu→implement, udgātṛ→test).

SKILL.md

GitHub

Yajña (यज्ञ) - Autonomous Development Ritual

Status as of 2026-09-13: rewritten as constraints. The iteration pseudo-code, the validation snippets, and the per-role model table were removed — the model plans the loop; this file says what must hold while it runs.

An autonomous development loop over three roles: hotṛ researches, adhvaryu implements, udgātṛ validates. Use it for a task list long enough that stopping for confirmation between items costs more than it saves.

What must hold

  • Do not pause for routine decisions. The point of the ritual is that it runs unattended. Stop only for a blocker, for completion, or on interrupt.
  • Every iteration starts with a live daemon. health_check first; an unreachable daemon is a blocker, not something to work around, because the loop's state and memory both live behind it.
  • Implementation happens in a worktree, never on main. Enter it before the first mutation, leave it on completion or failure, and surface patches as artifacts rather than applying them directly.
  • Validate that work landed, don't trust the report. An agent that says it edited a file and a file that changed are different claims. Check the file.
  • Three strikes ends an iteration. Same error three times, or a file still missing after three retries, is stagnation — stop and report rather than burning the budget on a loop that has stopped converging.
  • Address agents to explicit absolute paths. Vague targets are the dominant cause of an agent editing the wrong file.

Delegation follows the repo-wide policy in CLAUDE.md: research cheap, implementation to Codex gpt-6-astra in the worktree, review in the orchestrator. hooks/pre-tool-hook.sh enforces the routing, so pass model only where the default would be wrong.

Blocker or not

A failing test, a partial result, or a slow task is not a blocker — log it and keep going. A dead daemon, a permission denial, a critically broken build, or three consecutive identical failures is — stop, state what is needed, and leave the ritual resumable.

Output

Status lines while looping, not commentary: which role ran, what changed, tasks remaining. Full detail only on a blocker or at completion.

MCP Tools

Tool Purpose
health_check Pre-flight daemon check
long_task_start Initialize ritual
long_task_active Check for existing
long_task_snapshot Load context
long_task_update Record progress
long_task_event Log blockers/checkpoints
long_task_complete Mark done