Releases: RahimjonovBoburjon/multi-agent-coordination
v1.0.4 — Top-level SKILL.md for skills.sh website
What's new
Adds `SKILL.md` at the repo root so the skills.sh website can render the skill page correctly. The skills.sh CLI (`npx skills add ...`) already worked because it walks the repo with full-depth discovery, but the public website scraper only looks at the repo root or `skills//SKILL.md`. Without a root file, the public skill page showed "No SKILL.md available."
The new root SKILL.md is a short discoverability surface — YAML frontmatter (name + description), install commands for both skills.sh CLI and Claude Code plugin marketplace, quickstart, five-core-mechanics summary, slash command table, and a pointer to the full content at `plugins/multi-agent-coordination/skills/multi-agent-coordination/`. The plugin folder remains canonical.
Upgrade
```
/plugin marketplace update multi-agent-coordination
/plugin update multi-agent-coordination
```
Or via skills.sh:
```
npx skills update multi-agent-coordination
```
v1.0.3 — Marker collision fix, intro double-parens fix, redundant test-command fix
Three cosmetic bugs surfaced when testing v1.0.2 on a real project.
CLAUDE.md heading was garbled
The block heading was rendered as `## 🚨 Multi-Agent Coordinationnation -->` instead of `## 🚨 Multi-Agent Coordination`. Root cause: the legacy BEGIN marker (``) shared the word "coordination" with the next line's heading, and Claude's write path was merging the tails together.
· Fix: switched markers to `` and `` — distinct, short, no shared vocabulary with the heading.
· A blank line was added between the BEGIN marker and the heading to prevent any further visual merging during write.
· The wizard still recognizes the legacy marker format so existing v1.0.0-v1.0.2 installs upgrade cleanly on the next `/multi-agent-init`.
Developer intro had double-parens
Lines read: "You implement tasks dispatched by the Planner (the Planner (P))." Root cause: the v1.0.1 `T4 -> "the Planner (P)"` replace_all was too greedy and inserted the long form into contexts that already said "Planner".
· Fix: rewrote the developer intro phrasing to use "Planner (\`P\`)" once, no nesting.
Redundant test command in intros
When the user reuses the build command for tests (common when there's no test script in package.json), the intro printed "run \`npm run build\` and \`npm run build\` until both pass" — nonsense.
· Fix: wizard now detects `build_command == test_command` and collapses to single-command form: "run \`npm run build\` until it passes."
Self-check list expanded
Step 9 (intro output) now has a longer pre-flight check that catches all three bugs before declaring "setup complete". The wizard re-prints any block that fails a check.
Upgrade
```
/plugin marketplace update multi-agent-coordination
/plugin update multi-agent-coordination
```
Existing projects with the legacy CLAUDE.md markers will pick up the new format on the next `/multi-agent-init` (pick "Overwrite from scratch" — the wizard recognizes the old markers and replaces cleanly).
v1.0.2 — Force regenerate on overwrite + verbatim intros
Bug fixes
Overwrite from scratch now actually overwrites
In v1.0.1, when a user re-ran the wizard with "Overwrite from scratch" and the resolved settings happened to match the existing config, the wizard skipped regenerating CLAUDE.md. That meant version upgrades (like the v1.0.0 → v1.0.1 T4 → P rename) never took effect — the project stayed stuck on old conventions.
· Fix: overwrite is always full regeneration, regardless of whether settings changed. The templates themselves carry version-specific content; skipping bypasses every upgrade silently.
· If a user wants surgical changes, the "Update individual settings" path is the right tool — overwrite is for full resets.
Intros are now printed verbatim from templates
The wizard was paraphrasing the intro templates into 5-line bullet summaries instead of printing the ~50-line structured guidance. Receiving terminals missed the lock protocol details, "You MAY / You MUST NOT" sections, and the first-action prompts. Coordination drifted because the intro was incomplete.
· Fix: Step 9 of the wizard now mandates verbatim template output. Substitute placeholders, then print the entire file from line 1 to the last line. Self-check list added at the end.
Intro header format locked
Headers must match exactly:
· Developer: `━━━ T (Developer) ━━━`
· Planner: `━━━ P (Planner) ━━━`
· Solo: `━━━ SOLO ━━━`
No more `TERMINAL 1`, `TERMINAL 4`, `Terminal A` variants. The wizard self-checks each header before declaring setup complete.
Upgrade
```
/plugin marketplace update multi-agent-coordination
/plugin update multi-agent-coordination
```
Then in any project that ran v1.0.1, do a clean wipe + re-init to pick up the new conventions:
```bash
rm -rf .multi-agent active_tasks.md active_files.md
delete the multi-agent-coordination block from CLAUDE.md
```
Then `/multi-agent-coordination:multi-agent-init` — pick "Overwrite from scratch" and the new templates will land cleanly.
v1.0.1 — Planner label P, smart CLAUDE.md detection, variant explanations
What's new
Planner label changed: T4 → P
· Squad 3-terminal mode used to be confusing: T1, T2, T4 (where's T3?). Now it's T1, T2, P — natural and clear regardless of terminal count.
· Sub-agents now lock under `P-` instead of `terminal 4-`.
· Affects all intros, lock format examples, the CLAUDE.md template, and reference docs.
Lock format simplified
· Old format: `- frontend/src/views/admin/Dashboard.vue → terminal 1 @ `
· New format: `- frontend/src/views/admin/Dashboard.vue → T1 @ `
· Cleaner reads, less typing per acquire.
Wizard now explains Variant A vs B
When the user picks Custom mode, the variant question presents trade-offs explicitly:
· Variant A — feature branches + PR per task. Per-task git history; works with GitHub PR review tooling. Cross-task dependencies become awkward (S2-B can't import from S2-A until S2-A is merged); branch overhead.
· Variant B — direct commits to integration branch. Simple, fast, AI-friendly; the approval gate plays the code-review role. Less per-task rollback granularity.
No more blind picking.
New explicit "branch model" question
After Variant A/B is picked, the wizard asks how many long-lived branches:
· One branch (`main` only) — daily commits and production both on `main`. Releases tagged.
· Two branches (`dev` + `main`) — daily commits on `dev`. Promotion to `main` via release PR + tag.
Auto-detected from the repo state, but always surfaced as a deliberate choice — users may want to set up a `dev` branch even when only `main` exists.
CLAUDE.md smart-detect
Significant setup-safety improvement. When the existing CLAUDE.md has section headings that overlap with the multi-agent block (Git Workflow, Commit Format, Approval, Locking, Kanban, Terminals), the wizard now:
· Lists the overlapping headings to the user.
· Asks: append anyway / show block and let me merge manually / skip CLAUDE.md update entirely.
· If existing `` markers are found, replaces cleanly between them (the re-run case).
Prevents silent appending that creates duplicate sections.
Fixes
· Removed duplicate Step 3 / Step 4 in the wizard spec.
· CLAUDE.md template clarified: `.multi-agent/config.json` is committed, not gitignored (corrected a contradiction in the previous template text).
Install / upgrade
```
/plugin marketplace update multi-agent-coordination
/plugin update multi-agent-coordination
```