Internal shared skill. Prevents mode stickiness — the user being unaware that the AI is still making automatic decisions. Referencing workflows declare it in frontmatter
dependenciesand abort at startup if it is missing.
Mode recognition
- On hosts that list ai-proxy triggers, overlay phrases (「ai-proxy 模式」「AI 代理模式」 / "ai-proxy mode", or mid-run 「切换 ai-proxy」 / "switch to ai-proxy") request overlay first (section below). If both an overlay trigger and 「自动*」 appear, overlay+freeze wins over naked auto.
- Else trigger contains "自动" (auto) → auto mode; otherwise → manual mode (default).
- Mid-run switching: user says "切换自动模式" / "切换手动模式" to switch; overlay mid-run switch is 「切换 ai-proxy」 / "switch to ai-proxy".
- Manual: pause at each stage exit for user confirmation. Auto: proceed end-to-end, pausing only at workflow-defined limits (e.g. review-loop cap).
ai-proxy overlay (not a third control-flow mode)
Hosts that list ai-proxy triggers in description recognize an overlay on top of manual/auto:
- 「ai-proxy 模式」「AI 代理模式」 / "ai-proxy mode" request overlay; 「切换 ai-proxy」 / "switch to ai-proxy" switches mid-run.
- Overlay maps to auto carrier plus pending
Stage-exit policy: ai-proxy(verbal trigger is not occupancy; freeze writes the policy). Occupancy and thin freeze live inai-proxy-discipline— this skill MUST NOT restate the charter. - If both an auto trigger (「自动*」) and an overlay trigger appear, overlay + freeze wins over naked auto (named-escape pass-through).
- On full-flow completion or any interruption, overlay reverts to manual and clears
Stage-exit policy(same revert table as auto). Re-entering overlay requires an explicit trigger. Implicit continuation — "继续", "再改一下", "深入分析" — MUST NOT re-activate overlay. - Hosts whose
descriptionomits ai-proxy triggers (includingwrite-workflow) MUST ignore overlay; 「ai-proxy 模式」 does not change their two-state recognition.
Core rule: auto always reverts to manual
Auto mode automatically reverts to manual mode when:
| Scenario | Note |
|---|---|
| Full flow completes normally | Including the case where a new cycle is decided — a new cycle starts in manual by default |
| Flow is interrupted in any way | Failure abort, user-initiated stop, termination after a review-cap pause |
The workflow-specific completion point (e.g. PR merged, archive done) is defined by each workflow's own differences block, but the revert rule itself is universal.
Re-entering auto mode: explicit only
After reverting to manual, re-entering auto mode requires an explicit trigger:
- "自动 xxx" / "自动分析" / "自动解决" / "切换自动模式"
Implicit continuation — "继续", "再改一下", "深入分析" — must NOT re-activate auto mode.
Batch scenarios
In batch orchestration (e.g. goal-driven-queue; jira-fix-queue / opsx-jira-fix-queue are trigger shells that enqueue into it), the orchestrator passes Stage-exit policy (and thus overlay vs auto vs manual) explicitly per sub-invocation. The single-run revert rule above does not propagate across sub-invocations.
Integration guide (for referencing workflows)
- Keep in your own body: your trigger-word table (triggers differ per workflow), the manual/auto stop-point summary per stage, and a workflow-specific differences block.
- Delegate to this skill: the revert-to-manual core rule, explicit re-entry, implicit-continuation prohibition, the batch rule, and overlay recognition (if the host lists ai-proxy triggers). Do NOT copy their full text inline. Do NOT copy the
ai-proxy-disciplinecharter here. - Differences block examples (stay in the workflow, never move here):
--retryresets to manual;--resumekeeps the checkpoint mode; validation-failure rollback keeps mode within N attempts; archive failure counts as interruption.