This document is the single rule source + master control routing for the iteration pipeline; the responsibilities of all sub-skills are listed in Section 8, and the overall rules are subject to the following sections. Version number = current Git branch name (obtained by reading

git branch --show-current

only; if failed/empty → prompt manual branch switch, prohibit using empty string/guessed value). Global Path Convention:

Project Root/...

always refers to the Git repository root directory (where

AGENTS.md

is located, e.g.,

Project Root/.gatekeeper/

);

template/

of the

gatekeeper

skill refers to its own resource directory, and the two must not be confused.

1. Manual trigger only via the

   /gatekeeper

   command (set

   allow_implicit_invocation: false

   in

   agents/openai.yaml

   to prohibit implicit invocation by Agent);
2. Used to advance iteration, confirm current progress, or resume from an interrupted point.

Manual Requirement Confirmation

Agent High-Level Design

Manual Review Traceability + Approval

Agent Coding + Unit Testing

Agent Dev Environment Testing (traceability in Part 3 of review.md)

Manual Code Review Traceability + Approval

Manual Commit Merge + Staging Deployment

Manual Staging Testing Verification (traceability in test.md)

Manual-Assisted Production Deployment Finalization

• Requirement confirmation criteria:

  requirement.md

  has no HTML placeholders, and "Requirement Confirmer/Date" is filled.
• Modifications mentioned in chat: In requirement/high-level design phase → direct code modification is prohibited, require manual completion of requirement.md; minor changes in coding phase (not involving reviewed interfaces/data structures/design) → can modify code directly and proceed with code review.
• Minor change iteration: Mark "Skipped" in Phase 2 of progress.md, skip high-level design and design review, directly enter T03.
• T07 issue attribution: If covered by requirements → return to Phase 3/4 for repair; if not covered by requirements → manual requirement supplement before re-development, no need to re-run the full process.
• Historical archiving: Completed iterations are stored in

  Project Root/.gatekeeper/history/

  , and are completely ignored during iteration. After T08 is completed or

  /gatekeeper-abort

  is executed to terminate, migrate the entire iteration to

  Project Root/.gatekeeper/history/${version number}/

  .

1. The only basis for review is

   review.md

   ; the only basis for staging testing is

   test.md

   ; dev testing records are in Part 3 of

   review.md

   . Ignore temporary chat comments.
2. Two rounds of review (high-level design, code) + one round of testing (staging); multiple rounds of rectification and review are supported, and the round number is recorded in the "Round" column of the same table, no need to copy the header.
3. Work logic: Only read and rectify rows with [Status = Pending Rectification] in a targeted manner, no new functions; actively prompt manual review after rectification, do not judge completion privately; this round can only end after manual marking [Review Passed]; mark "No Rectification Needed" with reasons for false reports, and let manual personnel decide, do not modify or clear privately.
4. Prohibit rectification without basis, out-of-scope rectification, and private clearing of issue records.
5. Output a repair list after fixing, inform manual personnel for review.
6. Only after all issues are [Review Passed] + manual update of progress.md phase + issuance of advance instruction can enter the next phase; AI does not actively poll or advance on its own.
7. Minimum Elements for Code Review (Mandatory for Gate 2): Before approval, ① check the diff scope item by item according to the "Change List" in design.md to identify out-of-scope changes; ② interface contract changes must be synchronized to

   crmeb/apidoc/

   and stated in the rectification result; ③ test coverage description. Empty table/zero issues cannot be approved directly; for "zero issues", manual personnel must specify the check scope and sampling items in the approval remarks, and manually sample and check at least 1-2 change points.
8. Dev Testing and Acceptance Standard Mapping (Mandatory for T04): Each test item in Part 3 of review.md must fill in "Corresponding Acceptance §" (corresponding to Section 7, Acceptance Standards in requirement.md); un-mapped acceptance standards mean incomplete dev testing.
9. Staging Attribution Judgment (Mandatory for test.md): Must read and quote the original entry in requirement.md before judging "Covered by Requirements / Not Covered by Requirements"; only these two enumerated values are allowed in the attribution column, actively prompt manual correction for illegal values; if repair leads to interface contract changes, must state that apidoc has been synchronized.
10. Cross-Iteration Regression (Mandatory): The change list in design.md must state "Associated Historical Iteration Regression Items"; when touching previous functions, dev testing must retest and leave trace in review.md ③, otherwise dev testing is incomplete.

1. Only complete 1-2 tasks in one round of dialogue, prohibit completing the entire iteration at once.
2. Prohibit cross-phase work: Coding is not allowed before passing the high-level design review (or marking as skipped).
3. Progress, phases, and review results are all recorded in files, do not rely on AI dialogue memory.
4. After Agent fills in the rectification result, prompt manual personnel to submit traceability files; Git commits are executed manually.

1. Design review: Stop coding, wait for manual rectification instruction before reading the issue list in review.md.
2. If [Pending Rectification] is manually entered: Fix and fill in item by item, set status to "Rectified Pending Review"; mark "No Rectification Needed" with reasons for false reports.
3. Manual review: Mark "Review Passed" if approved, and the issue enters the next round if rejected.
4. Only after all are reviewed and passed + manual update of progress.md + issuance of advance instruction can enter the next link.
5. Code review logic is consistent with design review.
6. Staging testing: Manual personnel enter issues in test.md → Agent judges attribution → Fix and fill in → Manual re-test and review.
7. Progress Location: progress.md does not maintain a "current phase" pointer, locate the first uncompleted phase row according to the [Status] of each phase; verify that the status is consistent with the traceability conclusion before archiving.

1. Output Specification (P0): Only output change summary + increment in replies, prohibit full table display of traceability files; only show rows related to the current action.
2. Template Header Simplification (P0): 1-2 lines of header description for traceability files; delete redundant "Usage Instructions" during instantiation.
3. Incremental Reading (P1): Locate the scope according to [Status = Pending Rectification]/latest round, only read pending rectification rows and the latest round; do not re-read rows already read in the same round.
4. Codegraph Convergence (P2): Precise symbol names + maxFiles limit; prioritize trace/callers for link queries.
5. Single Rule Source (P0): Global rules are unified in this document, avoid repeated injection.

• Project Root/.gatekeeper/${version number}/

  does not exist → No process exists, initiate

  /gatekeeper-init

  , stop and wait for manual filling of requirement.md.
• Exists → Continue.

• Only read

  progress.md

  , locate the first uncompleted phase row (see 6.7); only read

  review.md

  /

  test.md

  to check [Pending Rectification] rows.
• If

  .init-incomplete

  mark exists → Return to

  gatekeeper-init

  (complete missing files).

• Only output: Version number, current phase, dispatched sub-skill and reason (one line), subsequent manual actions; prohibit full table display of traceability files (see 7.1).

1. This skill prohibits automatic usage by Agent; if no process directory is detected, only

   /gatekeeper-init

   can be initiated, and cannot advance requirements on behalf of manual personnel.
2. Only perform dispatch, do not act on behalf of sub-skills.
3. requirement.md / progress.md are maintained manually, Agent is read-only (see Section 3).
4. If the version number directory exists and there is no

   .init-incomplete

   , do not return to init (avoid overwriting manual content).
5. One round of

   /gatekeeper

   only advances the current phase (see 5.1).