| name | implx |
| description | Use when the user says "implx", "use implx", "用 implx", or asks for the one-line SpecRail queue shortcut. Plain implx means drain the full actionable issue/PR queue in auto mode: run SpecRail preflight, create missing spec/task PR work before implementation, use threads for reviewer/merge-reviewer lanes when available, create per-issue implementation PRs, require CI/reviewThreads/pr_gate evidence, auto-merge on complete evidence, and perform closure audit. Say "implx review" / "implx 人工" for the review mode that keeps per-PR human merge authorization. |
Implx
Use this skill as a short operational entrypoint. It only recognizes the
implx shorthand, records the queue mode, performs the minimum startup checks,
and delegates execution policy to the focused SpecRail skills.
Do not duplicate the implementation-queue contract here. The authoritative
queue planning, spec coverage gate, context budget, runtime checkpoint, threads
orchestration, PR gate, and closure-audit rules live in
skills/specrail-implement-queue/SKILL.md.
Startup
- Run the normal SpecRail startup before remote writes:
- read the repository
AGENTS.md
- read
AGENT_USAGE.md, workflow.yaml, states.yaml, labels.yaml, and
skills/specrail-workflow/SKILL.md when present
- select the human-facing locale
- identify human gates and route-gate requirements
- Fetch current remote state before mapping a GitHub queue.
- List open issues, open PRs, current branch, dirty files, and worktrees.
- Map existing PRs before creating replacement PRs.
- Record queue scope. Plain
implx, use implx, or 用 implx means full
actionable queue drain unless the prompt explicitly narrows the scope.
Queue Mode
Use queue_mode: full_queue_drain for the plain shorthand:
These explicit forms are equivalent:
implx drain full queue
implx resume full queue
用 implx 完成所有 actionable issues 和 PRs
用 implx 做完整队列
Use queue_mode: bounded_tranche only when the prompt explicitly limits scope,
for example one issue, one PR, the current tranche, plan-only, status-only, or
review-only work.
Authorization Mode
Two modes. Record the selected mode at startup and pass it downstream.
auth_mode: auto — the DEFAULT for plain implx, use implx, 用 implx,
and any drain/resume form that does not ask for review:
- The invocation itself IS the standing merge authorization for this run.
Do not ask per-PR "can I merge" questions.
- Merge a PR when ALL evidence is current and green per
skills/specrail-implement-queue/SKILL.md (CI rollup, PR gate, resolved
review threads, clean merge state, reviewer-lane evidence).
- Use closing keywords for final slices so merged PRs close their issues;
close merged-but-open issues during closure audit.
needs_spec / needs_tasks issues are actionable: auto-draft the spec or
task packet via the focused SpecRail skills, then implement. Do not park
them waiting for human spec approval.
- Items that genuinely need a human decision (duplicate-ownership conflicts,
maintainer waivers, probe or time-window gates, destructive or irreversible
actions, conflicting review feedback, architecture-level rewrites) never
block the queue: skip them, keep draining, and report them once in a final
human_decisions list with a recommended action each.
auth_mode: review — selected by implx review, implx 审核, implx 人工,
or when the repository workflow.yaml pins auth_mode: review:
- Preserve per-PR explicit human merge authorization in the current
conversation before any merge.
- Route
needs_spec / needs_tasks to spec-writing skills but wait for
human confirmation before implementing from a freshly drafted spec.
In both modes, never force-push, delete unmerged branches, replace a
maintainer-writable PR without cause, publish releases, or act outside the
repository without explicit instruction. Auto mode does not weaken the
Bounded Tranche Hard Stop, reviewer-lane, or self-review authorization rules.
full_queue_drain means the objective spans the whole actionable queue, not
that one session runs unbounded. Execution is a sequence of bounded tranches:
each session declares a hard budget (compaction count and/or item cap) in the
runtime checkpoint at tranche start, stops when the budget is exhausted, and
hands off to a fresh session via the checkpoint. See the Bounded Tranche Hard
Stop rules in skills/specrail-implement-queue/SKILL.md.
Pass the selected modes to skills/specrail-implement-queue/SKILL.md:
implx_context:
overall_objective:
queue_mode: bounded_tranche | full_queue_drain
auth_mode: auto | review
user_authorization:
current_branch:
dirty_files:
open_issues:
open_prs:
Delegate
After startup, load skills/specrail-implement-queue/SKILL.md for any issue or
PR queue. That skill owns:
- spec coverage classification
- implementation candidate selection
- one-issue-per-PR planning
- partial versus final closing semantics
- context-budget and runtime-checkpoint behavior
- optional threads orchestration
- verification, PR-gate evidence, and closure audit
For one small scoped issue, follow that skill's instruction to route to
skills/specrail-implement/SKILL.md.
Threads
If the queue needs native parallel lanes, reviewer lanes, CI waits,
review-thread checks, merge gates, or closure audit, load
integrations/threads.md and then follow the orchestration rules in
skills/specrail-implement-queue/SKILL.md.
For GitHub issue or PR queues, reviewer lanes, merge gates, and closure audit
make native thread dispatch required whenever native subagent capability is
available. Record thread_dispatch_gate before implementation, review, or
merge work. A coordinator self-review is not a native thread and does not
satisfy merge review.
If no native threads capability is available, continue with the single-agent
SpecRail flow only after recording the fallback and reporting that no native
threads were launched.
Boundaries
- In
auto mode, merge only on complete current evidence (CI, review threads,
merge state, PR gate, reviewer lane); evidence gaps mean skip and report,
not ask.
- In
review mode, do not merge without explicit human authorization in the
current conversation.
- Do not treat green CI as merge readiness without review-thread and merge-state
truth.
- Do not close an issue from a partial implementation.
- Do not replace an existing maintainer-writable PR unless it is stale, unsafe,
unwritable, or a human approves replacement.
- Do not use old Codex session logs as queue state.
Handoff
Report the compact handoff produced by the focused queue skill. Include this
implx wrapper context when useful:
implx_handoff:
route: implement_queue
overall_objective:
queue_mode:
auth_mode:
delegated_skill: skills/specrail-implement-queue/SKILL.md
queue_truth:
open_issues:
open_prs:
current_branch:
dirty_files:
human_decisions:
focused_handoff:
thread_dispatch_gate:
native_subagents:
spawn_requirement:
native_thread_evidence: