l-gevity/l-gevity-skills v4.0.0

on GitHub

ALCHEMY — your self-orchestrating software engineering team

Give /alchemy (Claude Code) or $alchemy (Codex) an objective. It still decides which engineering questions the change must answer. It now also decides who answers them: one agent, or a team it composes and launches itself. You never assemble a team, and you keep the merge.

Routing

• A second decision after dispatch: Execution: single-agent | multi-agent.
  • SKIP, DIRECT, ADAPTIVE and FULL mean what they meant.
  • SKIP and DIRECT always stay single-agent and never load orchestration.
• When it goes multi-agent: only when the work has independent parts with disjoint files, or a high-risk change needs a reviewer who did not write it. The number of gates or skills on the route is never a reason.
• Opting out: execution.multi_agent: off in alchemy-workspace.json keeps every run single-agent. A runtime that cannot launch agents falls back to single-agent and says so.

New skill: alchemy-orchestrate

• Agents are composed from capabilities, not a fixed roster. Capabilities (architecture, engineering, verification, security, optimization, scientific validation) come from the route's skills via skill-contracts.json. Each agent loads only the skills its package needs.
• Plans are checked before anything runs. Owned work packages are scheduled in dependency waves, and check_orchestration.py plan rejects:
  • write conflicts;
  • consumers that don't wait for their contract;
  • self-review;
  • nested delegation;
  • caps above the profile;
  • merge or deploy permissions;
  • auto-merge.
• Hand-offs are claims, not facts. check_orchestration.py handoff rejects:
  • summaries contradicted by the hand-off's own evidence or by workspace digests;
  • changes outside the write set;
  • stale context;
  • credentials;
  • embedded instructions such as "skip the review".
• The gate needs evidence. check_orchestration.py gate requires every criterion verified by its assigned check at the clean committed head, or deferred by the human. It also requires:
  • no open blocking finding;
  • disputes settled on evidence, never by vote;
  • evidence on both sides goes to the human;
  • bounded rework.
• The checker is tested: 49 check codes, each exercised by one of 99 adversarial cases.

Runtime adapters

init_workspace.py --runtime claude-code|codex|unstoppable|all renders every runtime from one profile and never overwrites your files.

• Claude Code: alchemy-implementer (own git worktree) and alchemy-analyst (read-only) agents; delegation depth 1; concurrency cap; bounded background wait; ask and deny rules for push, PR, merge and force-push.
• Codex: the same two custom agents and the concurrency cap. Read-only and depth limits are enforced by instruction plus the hand-off check.
• Unstoppable:
  • an alchemy-multi-agent pipeline: plan, plan check, human approval, native sub-agents, cross-provider review, gate, human approval, PR;
  • two project actions;
  • an independent review track for the bundled Code Review pipeline.

Documentation

• ALCHEMY-MULTI-AGENT.md: the normative specification. Each rule names the check that enforces it, or is marked judgment.
• README retitled, with a new section on how the team is formed, setup, what you approve, how to inspect a run, runtime support and measured runs.

Verified by real runs (.pilot/results/)

Audit pilot, same objective, one run per arm:

That is one fixture and one run per arm. Treat it as a behavior check, not a benchmark, and keep multi-agent where your own baseline shows it pays.

Not yet verified

• Unstoppable: no pipeline run has completed review, gate and PR. Runs planned, passed the plan check and launched agents. When they could not finish, they stopped with BLOCKED or ESCALATE rather than claiming success.
• Codex: parallel implementers and agent-failure recovery have not been run.
• Not exercised: cross-repository work.

Issues #18, #19 and #20 stay open for these.

Upgrade

Re-run your installer, then:

python .agents/skills/alchemy-orchestrate/scripts/init_workspace.py --runtime claude-code   # or codex, unstoppable, all

Existing single-agent use needs no change.