feature-delivery
GitHub协调Breeze功能交付流程,包括预检、分发修复任务到后台会话、监控进度及合并PR。支持按API/Web/Mobile拆分并行子任务以优化吞吐与成本,管理模型路由与账户配额。
触发场景
安装
npx skills add LanternOps/breeze --skill feature-delivery -g -y
SKILL.md
Frontmatter
{
"name": "feature-delivery",
"description": "Use when orchestrating Breeze implementation work from this seat — dispatching waves or issue fixes to background sessions, deciding whether an open PR gets merged, handling a fixer that died or stalled, choosing the next wave, or ending a run. Triggers on \"continue the run\", \"merge on green\", \"dispatch the next wave\", \"process the board\", \"what's next\", a task-notification that a fixer finished or hit a rate limit, or any moment you are about to ask Todd \"merge?\"."
}
Feature Delivery
Overview
One loop, two entry points. An issue enters as a fixer brief; a feature enters
through feature-pipeline (spec → Gate A → plan → Gate B → register_feature) and
comes out as wave sub-issues. From dispatch onward the loop is identical:
preflight → dispatch → watch → land → record → next wave
The orchestrator (this session) owns preflight, landing, and recording. Workers own the code. The merge is the orchestrator's call, not a question for Todd (Todd, 2026-09-04): once the land gate below passes, merge.
Related skills, not repeated here: feature-pipeline (intake and gates),
feature-lifecycle (GitHub wave state), issue-to-pr (what a worker does),
delegating-to-codex (quorum), gh-queue (backlog).
Preflight (before every dispatch batch)
mcp__feature-lifecycle__get_feature_statusfor each feature: which waves are open, which are Todd-gated (prod preflight SQL, real hardware, ops steps). Never improvise around a Todd gate. Then read each candidate wave's section of the plan doc: its dependencies (only file-disjoint, dependency-free waves run in parallel), whether it adds a migration, and what it touches (sets the model tier).claude agentsandclaude-dispatch ls— other orchestrators' sessions count against the host. Cap ~8 heavy fixers machine-wide (13 sonnet fixers → load 66; 36 parallel tsc → load 48 + 6 GB swap). Load 2.5 with 3 live sessions = room for 5.claude-dispatch usage— pick the account yourself and always pass--account <letter>: b lanternops, c olivetech, d pressless (Max subscriptions; burn the soonest-reset first, skip ≥80% 5h). Never a (breezermm — retired 2026-10-23, login expired) and never e (lanternops2 — API pay-per-token). ~60 s after launch grep the session log for "Not logged in": a logged-out fixer looks launched and writes nothing. Do not fall back to in-session Agent for waves.- Re-verify the item is still open, unclaimed, and has no PR (
gh pr list --search "<N> in:title"). Pools decay ~25% per 60 merges. - Parallel fixers that add migrations each get a distinct migration slot in the
brief (
2026-10-0X-100300,100500,100700…); same prefix ties sort by slug. PRs thatCREATE OR REPLACEa shared trigger function merge one at a time. - Public roadmap label. The feature parent (and the issue, for a standalone
customer-facing fix) carries
pbefore its first wave is dispatched. See "Public roadmap (p)" below.
Dispatch
claude-dispatch run --slug wave-<sub#> --prompt-file <brief.md> --agent issue-fixer --model <tier> --account auto
# issue: same, slug fix-<issue#>
Default mode is claude --bg -w <slug>: the worktree is claude-managed and
--branch is ignored there, so the brief must name the branch the fixer creates —
feature/<parent#>-<slug>/wave-<sub#> for a wave, fix/<issue#>-<slug> for an issue.
Launch from a git repo cwd. claude-dispatch stop <slug> removes the worktree when done.
mcp__feature-lifecycle__start_wave before launching a wave. Model tier by blast
radius: opus for tenancy/RLS, auth, billing, agent-shipped Go, permission gating;
sonnet for everything else; never Fable for workers. Under an Opus outage, run
sonnet with pre-decided design points and an abort clause (round 4, 09-03: 11/11).
Parallelism and cost (Todd, 2026-10-23: "do all that, exactly")
Wave-level parallelism alone leaves the host mostly idle: a dependency chain runs one or two sessions at a time. Three levers, applied to every multi-wave feature from the plan stage onward:
- Split long waves on the API / web / mobile seam. A wave with more than ~8
tasks or any mix of
apps/apiandapps/web/apps/mobilework becomes two or three sub-issues (add_wave, titlesW05a — …,W05b — …) that run in parallel, every part based onmainand targetingmain(never stacked). The w00/orchestrator contracts are the oracle, so web/mobile parts build against the API's interfaces with mockedfetchWithAuthbefore the API merges. The API part ownspackages/shared; a web/mobile part that needs an unmerged type copies the file verbatim from the API branch (git show origin/<api-branch>:<path>) so the later rebase resolves byte-identical. An engine-vs-services seam inside the API (W05) works the same way: the second part cherry-picks the listed schema/service SHAs first. The part landing last owns whole-wave verification. - In-wave fan-out. The plan names PARALLEL GROUPS of file-disjoint tasks; the
fixer runs each group as concurrent subagents in one worktree. Concurrent subagents
never
git commit; the fixer commits once per group. Groups run in order. - Sonnet for mechanical tasks only. The session and every tenancy / auth / permission / migration task stay Opus. The plan marks Sonnet tasks (locale fan-out, page scaffolds, pure helpers, docs, DTO plumbing); they run on Sonnet subagents with an Opus reviewer. Accounts b/c/d are Max subscriptions, so cost is quota, not dollars; account e bills per token and account a is retired.
Write the three levers into the plan doc (an "Execution split and parallelism" section per wave, the task allocation and groups spelled out) and into each brief. Example: Wisp (#9906) W04–W08.
A brief is self-contained: absolute paths, the plan doc path, the wave's file
ownership (parallel waves must be file-disjoint), the PARALLEL GROUPS and Sonnet task
list from the plan, the migration slot, the exact test commands, Closes #<sub-issue>
in the PR body, and these three lines verbatim:
- "Open the PR, run
/pr-review-toolkit:review-pr, post the review summary, STOP. Do not merge, do not close the issue." - "If the premise does not hold, ABORT and report — do not improvise."
- "Do not start background CI pollers and never
pkill -f; watching CI is the orchestrator's job."
Watch
claude-dispatch ls flips a slug to done when dispatch/<slug>.result.md lands.
claude logs <id> for a live screen, claude attach <id> to steer.
| Event | Action |
|---|---|
Fixer died on session limit 429 |
Its branch and transcript persist. Resume: claude-dispatch run --resume <sid> … or SendMessage to the same agent with "resume where you stopped + remaining steps". Re-dispatch fresh only if the transcript is gone. If only the review step is missing and the code is pushed, run /pr-review-toolkit:review-pr on the PR yourself and post the summary. |
| Fixer idle with nothing committed | It backgrounded codex or a poller and "waited". Resume with "run it in the foreground". |
PR CONFLICTING with a huge conflict count |
Stacked on a squash-merged parent. git rebase --onto origin/main <last-parent-commit>, never hand-resolve. |
| Fixer reports "done" | Verify: branch pushed, PR exists, commits on the right branch, review comment posted, CI actually ran. A PR based on a sibling branch gets NO CI (ci.yml triggers on main only); dispatch it: gh workflow run CI --ref <branch>. |
Land — the merge gate
Merge yourself when all of these hold. No board ask, no second review round.
/pr-review-toolkit:review-prran on the final head: read the comment and spot-check each claimed fix against the diff (gh pr diff <N>). Findings fixed or explicitly dismissed with a reason.- Head-SHA CI is green by the aggregate, not by per-job colour:
Cancelled ≠ green.gh run list --repo LanternOps/breeze --commit $(gh pr view <N> --json headRefOid -q .headRefOid) \ --json workflowName,conclusion -q '.[]|select(.workflowName=="CI")|.conclusion' # must be "success"gh pr checksandmergeStateStatusare not the signal (admin bypass makes them meaningless). Trivy exemptions apply to Security Scanning only, never to CI. - Base is
main(stacked PR → retarget or merge the parent first; its green was never CI). - Not on the hold list.
Hold and put on the board instead when: an unresolved review finding touches tenancy/RLS, auth, billing, migrations, or agent-shipped code; the wave is marked as a Todd gate in the plan; the PR belongs to another session; or the change needs a prod preflight before it is safe on main.
Then, in order:
gh pr merge <N> --repo LanternOps/breeze # enqueues; NEVER --admin, no strategy flag
main is behind GitHub's merge queue (since 2026-09-07). A bare gh pr merge <N> enqueues the PR (the queue owns the squash strategy;
--squash only prints a warning); the queue rebuilds it on top of whatever is ahead, runs the
full CI Success gate under merge_group (no path filters, smoke jobs blocking), and
lands it serially. --admin bypasses the queue and is what caused the 09-06/09-07
pile-ups (59 of 80 main runs cancelled, siblings CONFLICTING mid-sweep) — emergency
only, and say so on the PR. If the queue run fails the PR is dequeued with a comment:
fix and re-enqueue. Watch for the merge (gh pr view <N> --json state), then
mcp__feature-lifecycle__complete_wave → dispatch the next unblocked wave (no need to
wait for the post-merge main run yourself any more; the queue already evaluated
that ref) → when every wave is merged, mcp__feature-lifecycle__close_feature with
the not-verified list.
Closing a p-labelled feature is what moves it to "Recently shipped" — confirm the
label is still on the parent (not only on the roadmap item it came from) before
closing.
Public roadmap (p)
breezermm.com/roadmap/ renders exactly the LanternOps/breeze issues carrying the
label p (marketing repo src/lib/roadmap.ts). The label is deliberately one
character so it reads as noise on a public issue. Column is derived, not chosen:
| Issue shape | Column |
|---|---|
roadmap + p, open, status:idea / status:considering |
Exploring |
roadmap + p, open, status:ready |
Planned |
feature + p, open |
In progress |
feature + p, closed |
Recently shipped (latest 8 by close date) |
enhancement only, even with p |
not rendered — needs roadmap + status:* too |
Grouping uses the first category:* label (ranked Security → Remote Access →
Patching → … → Platform); an item with none lands in "Other areas", so every p
item gets at least one category:.
Rules while delivering:
- Tag on register. When
register_featurecreates the parent, addpplus acategory:if the feature is something a buyer would recognise (a capability, integration, or workflow). Skip it for plumbing: RLS retrofits, redaction contracts, relay alerting, attestation, abuse detectors, CI/infra. - Move, don't duplicate.
promote_to_featureleaves theroadmapparent open with its ownp, so the same item shows in Planned and In progress. Removepfrom the roadmap item when the feature parent takes it. Same when a standaloneenhancementissue is superseded by a feature. - Wave sub-issues never get
p. Only the parent renders. - Standalone issue fixes that are customer-facing and feature-sized get
p,roadmap,status:considering,category:<area>when dispatched, and are closed by their PR as usual. Bug fixes and one-line enhancements stay unlabelled. - Nothing is live until the marketing site redeploys; no action needed here.
gh issue edit <parent#> -R LanternOps/breeze --add-label "p,category:<area>"
gh issue edit <roadmap#> -R LanternOps/breeze --remove-label p
Record
At the end of a run, or whenever you would end a turn with open items, append ONE
entry to ~/.claude/breeze-handoff/BOARD.md using the template in its README.md:
what merged (PR → SHA), waves completed, what is dispatched and where, Todd gates
still open, exact resume commands. Memory gets state only when it is reusable across
sessions (account health, a new trap), never the procedure.
Rationalizations
| Thought | Reality |
|---|---|
| "Tenancy PR, I'll add one more review to be safe" | One review round per change. review-pr ran and CI is green → merge. Hold only on an unresolved finding. |
| "I'll ask Todd whether to merge" | Todd delegated the merge. Asking costs six hours. Merge or hold per the gate. |
"gh pr checks is all green" |
Per-job colour merged two red PRs on 08-28. Use the head-SHA CI workflow conclusion. |
| "Two at a time to be careful" | Host cap is ~8 heavy fixers; under-dispatching wastes the day. Read the load. |
| "I'll use in-session Agent for the waves" | Waves run as claude-dispatch --bg issue-fixers in their own worktrees and account profiles; Agent inherits one account and one context. |
| "The fixer died, dispatch a fresh one" | Its branch and transcript persist. Resume first. |
"I'll tag p in a batch later" |
Batch passes miss shipped items (7 closed features were unlabelled on 09-05 and never appeared in Shipped). Tag at register, move at promote. |
| "Fixer said done" | Verify branch, PR, review comment, CI on a main-targeted PR. Subagents have reported success with uncommitted or off-branch work. |
版本历史
-
0b71144
当前 2026-11-03 11:01
新增Wisp功能的执行拆分策略,支持API/Web/Mobile并行子任务;增加并发性与成本控制的三个杠杆说明;明确Sonnet模型路由规则及账户使用规范。
- 1c3a08e 2026-09-22 22:17


