Repository navigation
Replies: 2 comments
|
On 1 and 2, I'd split by what each layer costs to carry. Anything in a file that loads every session is sent again on every step, so Layer 1 earns its place only as a short trigger rule and Layer 2 doesn't earn it at all. I keep my own rule as a skill whose description line is always visible and whose body loads when it fires. The worktree recipe then sits in the project, because the slot derivation depends on the stack. On 3, I run the threshold at 2. The test that matters is shared state rather than count. Two items that touch one file stay serial, and six that share nothing run in parallel. Two things I'd add from running it. Cap what each agent returns (a summary plus a path to a file), since a long report pasted back is read again for the rest of the session. And build the coordinator's checklist from the original issue list, never from the agents' reports, or it inherits whatever they skipped. On 4, I'd leave simulator limits to the project recipe. I maintain a reference architecture that documents the cost side of this at https://github.com/jimy-r/agent-workspace-architecture/blob/main/PATTERNS.md#18-position-is-price--a-token-costs-more-the-earlier-you-add-it Drafted with Claude Code and reviewed by me before posting. |
|
On question 1: whatever ends up opt-in, I’d keep “the coordinator merges branches and runs the full suite itself” in the default rule. It’s the line that catches what per-worktree tests can’t. I set that case up on purpose on lukeed/mri. Claude Code renamed the internal helper toVal to assign, and Codex routed --no- flags through toVal. Both edited src/index.js, git merged them with no conflict, and each branch passed npm test on its own. Merged, the suite failed with ReferenceError: toVal is not defined. Running each branch alone as well as the merge is what showed it was the combination and not one bad branch. One small repo, and I picked the two tasks so they’d touch the same helper, so it shows this can happen, not how often. I build Pytxo, and |
Uh oh!
There was an error while loading. Please reload this page.
The problem
A batch of 5 or 6 independent issues (stakeholder feedback, a bug list, a round of copy edits) almost always gets worked serially in one long session. The pieces are independent and could run as parallel subagents, each owning one issue with its own failing test, but nothing in the template or the plugin nudges toward that.
Two layers are involved, and they seem to want different homes.
Layer 1: the delegation rule
A behavioral rule that fits the template's existing voice:
This is short enough for
MODELCLAUDE.md.Layer 2: worktree isolation
A worktree isolates files only. Ports, databases, Docker Compose projects, gitignored
.envfiles, dependency installs and caches are all shared until separated. A workable recipe assigns each agent a slot number N and derives everything from it (port = base + N,<db>_wt<N>,COMPOSE_PROJECT_NAME=<project>-wt<N>), copies.env*into each worktree, and keeps shared cloud state (BigQuery, HubSpot, Sentry, deploy config, schema migrations) with the coordinator.That is 40 or more lines of stack-specific setup. Putting it in a global CLAUDE.md taxes every session, including ones editing markdown.
Questions
MODELCLAUDE.md, or is it opinionated enough to be a separate opt-in snippet?fullstack-dev-skills(something likeparallel-worktrees), a per-project CLAUDE.md recipe the rule defers to, or both?Drafts of both layers exist and can be posted here if there is interest.
🤖 Generated with Claude Code
All reactions