sdlcTasksExecutions

SDLC Tasks Executions

One slash → numbered tasks → one agent per task → full software lifecycle per lane.

Slash: /sdlc-tasks-executions. Aliases still recognized: coordinate-worktrees, “setup worktrees”, “one agent per task”, “fan out lanes”.

Former name: coordinate-worktrees. All sibling skills that said that name mean this skill.

Modes

Default when the user pastes a numbered list or says full SDLC / execute: execute.
Default when a sibling says “setup-lanes” / “land-lanes”: that mode only.

One task = one agent = one worktree = one branch = one issue = one PR.

Reuse first (do not reimplement)

This skill owns: task parsing, lane isolation, SDLC mandate, matrix board, merge sequencing when authorized. Details: REFERENCE.md.

Safety

• Never commit product work on the default branch. Topic branches only.
• Never delete remote branches, tags, or PRs unless the user explicitly asks.
• Own-lane isolation: each agent edits only its worktree and path globs. Foreign dirty trees are user-owned.
• Reject or sequence tasks that share the same files/schemas until land order is clear.
• Freeze or coordinate active writers before land-lanes. Snapshots go stale while another agent writes.
• No broad reset/clean/checkout-overwrite, blanket stash, forced branch switch, shared rebase, or ours/theirs conflict strategy.
• Treat untracked, ignored-but-important, symlink, mode, submodule, sparse-checkout, detached-HEAD, and sequencer state as first-class data.
• execute authorizes per-lane merge only when the user said merge / “to main” / full ship, and hard gates pass (unit + e2e green, confidence ≥ 6). Default without merge language: open PR only.
• Headless e2e default. Headed only if user said headed / visible / ui.
• Never print secrets from .env / fixtures.
• Agent host is agnostic: host subagents, cmux per lane, or briefs-only.

Workflow

0. Choose mode

Mode: execute (default — full SDLC)

1. Parse tasks

1. Repo root, default branch from origin/HEAD, git fetch (do not mutate other worktrees).

2. Read AGENTS.md, CODE-STYLE.md, ADRs, package verify/test scripts.

3. Split the request into independent numbered tasks. Accept forms: 1. …, 1) …, - [ ] …, or explicit “Task N:”.

4. Name each with a short kebab slug. Flag overlaps; sequence or batch only when paths do not collide.

5. Mint a run-scoped board path (never a fixed file that parallel runs overwrite):

   RUN_ID=$(date -u +%Y-%m-%dT%H%M%SZ)
   AGENT_DOCS="docs/agent/sdlc-tasks/$RUN_ID"
   mkdir -p "$AGENT_DOCS"
   printf '%s\n' "$RUN_ID" > docs/agent/sdlc-tasks/CURRENT
   

   Write BOARD.md (and optional STATE.md) only under $AGENT_DOCS. Put AGENT_DOCS in every LANE-BRIEF.md so all lanes share this run. Resume → use CURRENT / explicit run-id (do not mint a new one). Never put campaign MD at repo root or under product docs/agents/.

6. Host mode: A subagents (default “just run it”), B cmux watchable, C briefs only.

2. Fan out lanes (same mechanics as setup-lanes)

For each task:

1. Create GitHub issue (gh issue create) with title, acceptance, out-of-scope — or bind existing #N if the line already has an issue.

2. Worktree under repo root:

   git fetch origin
   git worktree add ".worktrees/<type>-<issue>-<slug>" -b "<type>/<issue>-<slug>" "origin/<default>"
   

   • Branch: feat|fix|refactor|chore|docs/<issue>-<short-slug>
   • Dir: .worktrees/<type>-<issue>-<slug> (slashes → dashes). Ensure .worktrees/ is gitignored when the repo agrees.

3. Write <worktree>/LANE-BRIEF.md with the full SDLC mandate (template in REFERENCE). Include path globs, acceptance, unit + e2e requirement, UI QA when product is UI, “no main commits”, merge policy.

4. Spawn one agent with cwd = worktree (or cmux workspace / brief-only).

3. Per-agent SDLC mandate (every lane)

Each agent must complete only its task:

1. Clarify acceptance against issue body; stop if blocked (record on BOARD).
2. Implement using repo patterns (reuse-first-audit only when build-vs-buy is ambiguous).
3. Unit happy path — add/extend; run repo unit command → green. Hard stop if red.
4. E2E/integration happy path — real suite or smallest durable e2e; headless default → green. Hard stop if red when stack exists; honest skip only if env missing (cap confidence).
5. Manual / UI QA — load preview-and-prove for user-visible web/app flows; record evidence or N/A.
6. Commits — organized-commits on the topic branch only.
7. Ship pre-merge — finish-and-ship: verify, push, open/update PR with Fixes #<n>, Summary, Confidence 1–10, test plan (unit/e2e/QA/act).
8. act / CI when workflows exist — red blocks merge.
9. Merge only if orchestrator authorized and confidence ≥ 6; else stop at open PR.
10. Do not delete remotes; do not touch other lanes.

4. Orchestrator after lanes

1. Refresh BOARD: task → issue → worktree → branch → PR → unit → e2e → QA → confidence → merge.
2. If merge authorized: order merges (contracts → domain → UI → e2e-heavy last); gh pr merge; rebase remaining open PRs on conflicts; never force-push default.
3. Optional reinstall/smoke on default tip when user asked full ship (see REFERENCE).
4. Done receipt (Verification section). Unfinished lanes stay open.

Flags: no-merge (default without merge language), merge / to-main, no-cmux, headed, agent=grok|claude|codex, max-lanes=N (default all, soft cap 8).

Mode: setup-lanes

Same fan-out as execute steps 2–3 without forcing full SDLC gates unless the calling skill’s brief requires them. Each lane: implement → narrow/full verify as brief says → organized-commits + finish-and-ship push/PR. Do not merge unless asked. Report matrix. Used by messy-repo / test-gap-ship / lean-prove / ux-journey.

Mode: land-lanes

1. Target = remote default (or named product tip). Fetch without mutating worktrees.
2. Inventory every worktree (include .worktrees/*): path, branch, upstream, ahead/behind, dirty, ignored-important, stash, sequencer.
3. Classify: clean committed, dirty-only, already reachable, patch-equivalent, divergent, blocked. Use patch-ID + history, not subjects alone.
4. Map path and semantic overlap (schema/consumer, API/caller, migration/model).
5. Backup refs for target and every committed source. Preserve dirty/untracked with verified methods.
6. Make each lane reviewable (organized-commits when authorized). Prefer landing via open PRs.
7. Integrate one logical lane at a time; mechanical conflicts only when intent is clear; stop on competing logic.
8. Focused checks per lane; full gate after convergence.
9. Prove original commits + preserved artifacts reachable before proposing local cleanup. Remote delete still forbidden unless user asks.

Verification

execute

mode: execute
repo: <path>
default: <branch> @ <sha>
board: docs/agent/sdlc-tasks/<run-id>/BOARD.md
run_id: <YYYY-MM-DDTHHMMSSZ>

host: A|B|C
lanes:
  - N | task | issue | worktree | branch | pr | unit | e2e | qa | confidence | merge
skills_reused: organized-commits, finish-and-ship, preview-and-prove[, …]
residual: <blocked / deferred / overlaps>

Done only when each started lane has issue + worktree + branch + PR (or honest block), unit/e2e evidence (or skip reason), no default-branch product commits, no remote deletes, and merges only when authorized + gates green. “Agents ran” is not done.

setup-lanes

Per lane: path, branch, issue URL, host/cmux, LANE-BRIEF.md, PR URL, head SHA, verify outcome; no default commits; no remote deletes; leftover unassigned tasks listed.

land-lanes

Matrix of every worktree disposition, backups, merge order, conflicts, focused + full gate results, reachability proof, cleanup authorization (local vs remote separate).

“Everything landed” = reachability + byte preservation — not empty git status alone.
“Agents fired” = issue + branch + PR per lane — not directories alone.