Files

26 KiB

CI — PR 合入门禁

The pull-request admission layer has exactly nine required external contexts:

Required context Contract
Lint Stable PR revision/risk classification plus applicable formatting, go vet, and Actionlint
Test Tier-selected race/unit/release-script tests plus representative cross-platform compilation
Coverage Scope-matched overall non-regression and 100% changed-code coverage
Policy Repository policy and the fail-closed CHANGELOG contract
Edition Edition contract tests
Interface Integrity CLI, Schema, Skill, and stable-release compatibility
AI Behavior Base-owned policy for PRs labeled ai-generated
CLI Smoke Offline help for every public top-level command
Mock MCP HTTP and stdio MCP lifecycle smoke tests

The workflow display name is CI. Parallel helper jobs may implement Test and Coverage, but they are not ruleset contexts. Do not require an aggregate alias or a downstream integration check in place of the nine contracts above.

AI Behavior is evaluated by a pull_request_target workflow that never checks out or executes PR code. It writes the exact AI Behavior status to the current PR head. Its Files API read is bracketed by base/head revision checks, so a synchronize race fails closed. The same workflow supplies a successful AI Behavior check run on protected main pushes for release governance.

Exact CHANGELOG-only fast path

A pull request qualifies only when GitHub reports exactly one changed file, that file is an in-place modification of CHANGELOG.md, and the base and head both retain it as a regular non-executable 100644 blob. Add, delete, rename, symlink, executable-mode, and second-file changes do not qualify.

Lint classifies the Files API result only after verifying that the API's base and head equal the event revision both before and after pagination. Policy checks out GitHub's PR merge ref and verifies its parents:

HEAD^1 = pull_request.base.sha
HEAD^2 = pull_request.head.sha

It then runs:

./scripts/policy/check-changelog-pr.sh \
  --fast-path "$PR_BASE_SHA" HEAD

The exact fast path remains limited to historic one-file maintenance. A release-seal PR uses --content-only, which permits the generated CHANGELOG.md change together with archival moves from .changes/ to .changes/released/; it receives the normal scoped admission instead of this fast path. Ordinary PRs must not modify CHANGELOG.md; they add a standalone release fragment instead. The validator and its policy dependencies in that merge tree are byte-for-byte the current base versions. Validation targets the synthetic merge tree, not the feature-branch tree, so a stale branch cannot supply an older validator or combine with newer base notes into an invalid final CHANGELOG.

All nine admission contexts are still emitted and must succeed. Expensive implementation helpers are skipped; the named contexts record that their code surface is unaffected.

The protected main push keeps that fast path only when all of these fail-closed conditions hold:

  • the event is a non-forced update of the existing refs/heads/main;
  • the event after SHA is the exact workflow SHA, and both event SHAs are complete, non-zero commit IDs;
  • GitHub's comparison reports the previous main tip as the unique linear merge base, with no commits behind it;
  • the complete resulting tree diff is exactly one in-place modification of CHANGELOG.md;
  • the previous main tip already has successful GitHub Actions checks for all nine Code Admission contexts.

Policy then independently checks out the pushed revision and runs the same check-changelog-pr.sh --fast-path contract from the event's before SHA to its after SHA. If identity, ancestry, file scope, tree mode, CHANGELOG content, or predecessor admission cannot be proved, classification falls back to the complete main admission suite. A source change can therefore never inherit the CHANGELOG-only result.

Any PR that touches CHANGELOG.md but also changes another file runs the same content contract in Policy with --content-only. That mode accepts only fragment archival moves (.changes/<name>.md to .changes/released/<version>/<name>.md) alongside the changelog; source and documentation changes are rejected. It still rejects invalid dates or versions, missing bullets, placeholder TODO/TBD, unmanaged-section changes, and unsafe tree modes.

Risk tiers and downstream boundaries

Lint resolves the complete base/head diff before any helper is skipped. Unknown or truncated input fails closed into the high-risk tier.

Tier Selection Admission work
Documentation-only Only prose/documentation assets; no executable, generated, workflow, packaging, or interface surface Documentation and repository-asset validation; expensive code helpers skip while every required context still succeeds
Standard Ordinary code change with a stable package graph Race tests for changed Go packages and their reverse dependencies; candidate and merge-base coverage over the same impacted scope and coverpkg; representative Darwin/Windows compilation
High-risk / protected main Workflow/policy, package add/remove/rename, generated Schema/registry, platform, auth/keychain, installer, packaging, release, transport, recovery, or an unprovable infrastructure classification Complete race suite and full native macOS/Windows tests, plus every affected domain gate

Domain helpers (Edition, Interface Integrity, CLI Smoke, and Mock MCP, for example) execute their substantive suites when the diff can affect that contract or when the high-risk tier is selected. Otherwise their stable named contexts still report a successful, explicit unaffected result. Release-script tests follow the same impact rule. This preserves the ruleset contract without charging every developer for unrelated work.

Platform-sensitive changes additionally run native changed-code coverage. Protected main always runs native tests; generic portable changes are held to the Linux changed-code gate rather than being forced to manufacture platform-only coverage.

Complete Multi-profile E2E is not a PR admission context. It belongs to the Main Integration — 主干集成 workflow and runs only after a push to main (or an explicit manual dispatch). A failing downstream run remains a real regression and must be repaired, but it must not be represented by a synthetic successful PR check.

flowchart TB
  PR["Pull request"] --> ADMISSION["CI"]
  ADMISSION --> L["Lint"]
  ADMISSION --> T["Test"]
  ADMISSION --> C["Coverage"]
  ADMISSION --> P["Policy"]
  ADMISSION --> E["Edition"]
  ADMISSION --> I["Interface Integrity"]
  ADMISSION --> A["AI Behavior"]
  ADMISSION --> S["CLI Smoke"]
  ADMISSION --> M["Mock MCP"]
  ADMISSION --> MAIN["Protected main"]
  MAIN --> NATIVE["Full native platform matrix"]
  MAIN --> E2E["Multi-profile E2E"]
  MAIN --> RELEASE["Release delivery"]

Review ownership and auto-merge

A base-owned pull_request_target workflow routes newly opened, updated, reopened, or newly ready PRs targeting main to one eligible peer reviewer. It does not check out or execute PR code, excludes both the author and the known latest pusher, and balances the open requested-review load across the reviewed maintainer pool. A current-head approval or change request is preserved; after a new push, stale activity does not suppress a fresh request, and an outstanding change requester is preferred for continuity.

The branch rulesets keep one human approval and all nine strict required contexts, require someone other than the latest pusher to approve after the most recent head update, and restrict main updates to the dedicated Reviewer Router App in pull-request mode plus the designated Formula publishers and break-glass identities. Repository auto-merge is enabled for ready PRs, so the App-owned request records the automation intent while the App's synchronous merge path waits for that approval and the current revision's nine green checks. If main advances, strict checks rerun before merge. The reviewer routing job uses the built-in GITHUB_TOKEN to request reviewers with Contents: read and Pull requests: write. A separate base-owned cleanup job isolates the merge-authority permissions (Contents: write and Pull requests: write), revalidates the exact event base/head, and uses the built-in token only to disable an existing request owned by github-actions[bot] or one whose title or merge metadata requests that GitHub skip workflows; it never enables auto-merge. The job then mints a current-repository installation token for the dedicated Reviewer Router GitHub App, proves its emitted slug matches the reviewed REVIEWER_ROUTER_APP_SLUG, replaces every non-App request, and enables native auto-merge with fixed metadata. This identity boundary is required because GitHub suppresses most workflow events created by the built-in token; using it for auto-merge would silently skip the merge commit's protected-main CI and baseline-cache producer. Token minting or takeover fails closed without falling back to GITHUB_TOKEN: the unsafe request is cleared before credentials are read, and the required Test context live-verifies the exact main-merge-writers update rule. GitHub's read APIs may omit parameters for the strict update_allows_fetch_and_merge: false value, so the gate accepts only that exact omission or a one-field parameters object containing explicit boolean false; every other present shape or value fails closed. The gate then binds the same ruleset node through GraphQL and requires its non-null updateAllowsFetchAndMerge value to be exactly false. It also requires its own built-in token to report current_user_can_bypass: never. The minted App separately requires pull_requests_only on that writer rule and never on every other active main ruleset before it can enable, reconcile, or synchronously merge. The read-only Test token may receive a repository projection with both merge-default properties omitted; it accepts only that complete omission or exact MERGE_MESSAGE plus PR_TITLE/BLANK, while partial or malformed projections fail closed. The same unprivileged pull_request job may receive an empty repository-variable projection for an external fork. Only when the event head repository differs from the base repository does it substitute the exact reviewed public slug dingtalk-dws-reviewer-router for identity comparison. An empty variable on a same-repository PR and every malformed non-empty value still fail closed. This fallback neither mints a token nor grants merge authority; the base-owned Router continues to require its minted App slug to equal the repository variable before any mutation. The minted App's Contents: write token must observe the exact reviewed defaults before either mutation path proceeds. GitHub hides the complete bypass_actors list from low-privilege callers, so the rollout audit must still keep the writer list at exactly the Reviewer App, haofeng0705 (ID 30925823), and PeterGuy326 (ID 47820304). The required check finally accepts a null or exact App-owned request after a short takeover grace period. Null is safe from the suppressed event path because the built-in Actions identity cannot update main; other permitted identities produce either a main push or the trusted closed-PR repair. Drafts skip the identity step, while ready_for_review, edited, auto_merge_enabled, and auto_merge_disabled explicitly start fresh admission for readiness, title, and merge-request changes. Router does not react to auto_merge_disabled, so a human can deliberately leave the PR manual-only for break-glass handling. Reviewer routing remains available. The protected-main push that deploys the workflow automatically migrates every open, ready non-App request and repairs unsafe App metadata; it disables workflow-skipping requests for correction. Because GitHub's deferred native auto-merge path does not reliably apply an App's pull-request-only ruleset bypass, a zero-permission approval-signal workflow and completed CI / Code Admission — AI Behavior workflows wake the same trusted default-branch reconciliation through workflow_run. That event is only a wake-up signal: the privileged job does not consume its pull-request payload or artifacts and does not check out the triggering run's code. It re-enumerates open main PRs through the API and attempts only an exact App-owned request through the synchronous PR merge endpoint. Immediately before each attempt it revalidates the App's ruleset boundary and PR intent, supplies the current head SHA, and treats server-declared not-ready or concurrent-revision responses as retriable. The live preflight requires the exact repository-owned approval ruleset and exact nine-check strict quality ruleset, with every context bound to the GitHub Actions App (integration_id=15368) and the Reviewer Router App unable to bypass either; a missing, disabled, incorrectly sourced, or weakened gate fails closed before merge. GitHub—not the workflow—decides whether the merge is admissible. A staggered twice-hourly schedule provides eventual recovery, and a manual workflow_dispatch from main is the immediate idempotent retry path. Reconciliation never enables an originally null request. The reviewer router is orchestration, not a quality context, and must not be added to the ruleset.

Disabling the App-owned request before the reconcile job's final PR read keeps the PR manual-only. GitHub can atomically bind the subsequent merge to the head SHA, but it cannot bind that call to the auto-merge intent; a disable racing after the final read may therefore lose to the in-flight merge. Closing the PR or changing its head blocks the attempt only if GitHub observes that state before accepting the merge endpoint call; no client-side action can revoke a merge that the server has already accepted.

The merge endpoint has no expected-base precondition. Reconciliation checks that the base is this repository's main immediately before and after the call, but a retarget racing after the final read is not atomically preventable in the workflow. Operators must disable the App-owned intent and wait for all running Reviewer routing reconciliation jobs to finish before retargeting a PR; a stronger adversarial guarantee requires a GitHub-side branch/ruleset control.

GitHub may omit pull_request_target for security-sensitive head branch names, including names that look like commit SHAs. Those PRs cannot use Router App takeover or the closed-event repair: rename the branch for the supported path, or use the designated break-glass identity with a safe final message so main push CI remains the exact-SHA producer.

Running focused gates locally

Run the contracts relevant to the change. Ordinary contributors are not expected to repeat every CI job locally:

make build
make policy
make interface-integrity BASE_REF=<merge-base> STABLE_REF=<stable-tag> CANDIDATE_REF=<candidate-sha>
make schema-compatibility BASE_REF=<merge-base> STABLE_REF=<stable-tag> CANDIDATE_REF=<candidate-sha>
make skill-command-integrity
make cli-smoke
make mock-mcp-smoke
go test -v -count=1 ./pkg/editiontest/...

CI 先解析并核对精确的 merge-base、最近可达且未撤回的 stable GA tag 和已提交的 candidate SHA,再调用 make authoritative-interface-integrity。本地 make interface-integrity 与该 CI target 都只委托给同一个 modern authoritative wrapper,不存在第二个比较 入口。省略 BASE_REF 时本地 target 默认比较 origin/main,省略 STABLE_REF 时自动 选择该 base 可达且未撤回的最近 stable GA tag,省略 CANDIDATE_REF 时比较已提交的 HEAD。 需要逐字复现某次 CI 时,应显式传入该次运行记录的 merge-base、stable tag 和 candidate SHA。

make update-interface-baseline / make reset-interface-baseline 只维护 test/fixtures/cli-interface-baseline.txt 这一份非权威 CLI Smoke fixture。底层旧 check-interface-baseline.sh 不再作为本地或 CI 的兼容性审批入口,也不能用于批准 flag 迁移。

Schema compatibility 使用同一组 base、stable、candidate refs,以及 base-owned flag 与 command migration ledgers。merge-base-owned checker 分别规范化 merge-base 与 stable 的完整 Schema,并让 candidate 对两份历史 contract 独立执行检查;它只把已通过 Interface lifecycle 的 exact rename、command move 或 flag extraction 规范化到当前历史 副本,不会维护第二份 allowlist,也不会放宽其他 Schema 历史字段。

For a release-seal branch that archives rendered fragments:

base_ref=$(git merge-base HEAD origin/main)
./scripts/policy/check-changelog-pr.sh --content-only "$base_ref" HEAD

make coverage-gate is an enforcement step, not a profile generator. For a standard PR, CI derives changed packages and their reverse-dependency test closure, then generates candidate and merge-base profiles with the same test scope and coverpkg. High-risk and protected-main runs use the complete profiles. The complete candidate profile is produced by disjoint per-shard helper jobs (scripts/ci/test-packages.sh list-coverage, kept serial with -p 1 inside each shard; verify proves the shard union equals the full-suite scope exactly once) and concatenated in the aggregate job before enforcement. The complete merge-base profile is restored from an exact-key cache written by the last green main push of that same commit (key: merge-base SHA plus resolved Go version); any miss falls back to recomputing it in a merge-base worktree. The trusted main producer and PR consumer use the same dedicated cache profile path because GitHub includes that path in the cache version; the runtime-facing candidate and baseline filenames remain separate. Near-miss reuse is forbidden — the caches carry no prefix restore keys, because a neighbouring commit's profile would compare the candidate against the wrong baseline. PR concurrency is keyed by PR number, so a later revision cancels the stale run instead of letting obsolete test matrices compete with the replacement for hosted runners. If cancellation interrupts a cold-cache fallback, the latest run recomputes the same exact merge-base profile authoritatively. Main concurrency remains keyed by pushed SHA, so a newer main push cannot cancel a predecessor's producer.

Every supported main advancement path has an exact-SHA producer. The required Test context rejects GitHub workflow-skip directives in PR and auto-merge metadata, reruns when that metadata is enabled, disabled, or edited, and verifies the live App/writer-ruleset identity contract. Reviewer Router additionally binds auto-merge to the exact head OID and writes a fixed safe merge headline/body. The sole break-glass publisher must retain a safe final message; the release-controlled Formula-only path is the sole supported use of [skip ci]. A full source push saves the assembled profile after the aggregate gate passes. A trusted documentation or release-seal push independently verifies that the complete before...after diff contains only the reviewed metadata allowlist, restores only the exact before cache, recomputes the full profile if the chain is cold, and makes that helper a dependency of the required Coverage context. Release-generated Formula commits intentionally retain [skip ci]; after their nine synthetic contexts are sealed, an independent release-governance job creates an acknowledgement and emits a coverage-baseline-promote repository dispatch. The default-branch promotion workflow revalidates the exact single-parent Formula identity, successful parent and target contexts, and main containment before it promotes the exact parent cache or performs the same full fallback. Every target-main producer follows its save with a lookup-only restore and requires cache-hit=true for the exact key; this turns the cache action's otherwise warning-only upload failure or prefix match into a hard failure. Formula promotion additionally updates one release-created Coverage Baseline Cache check. A separate confirmation job waits for that exact check-run ID while npm and mirrors remain dependent only on the immutable publication job; cache failure therefore makes the final delivery gate red without creating a partially published release. Once Formula sealing exposes its SHA, a later publication verification failure cannot suppress that confirmation job.

A separate base-owned pull_request_target: closed safety net covers the final merged SHA even if a human or integration changes the merge message after PR checks finish. Skip directives alone do not suppress pull_request_target, subject to GitHub's separate security-sensitive branch-name restriction above. That job executes no PR code and only dispatches after binding the exact closed-event PR number and stable head SHA to merged-PR facts (merged_at, base.ref, and merge_commit_sha) and proving main containment. It does not compare the later REST base.sha, which follows the live base branch after merge. Because GitHub makes default-branch caches read-only to pull_request_target, the dispatcher first waits up to one minute for a run from the exact protected .github/workflows/ci.yml workflow and exits when that normal producer exists. A successful main CI hard-verifies the exact key itself. A completed non-success run starts a separate base-owned workflow_run dispatcher, which binds the exact CI workflow ID/path, run ID/attempt, conclusion, upstream repository, main branch, and head SHA. That trigger is also cache-read-only, so either trusted dispatcher uses repository_dispatch; its producer revalidates the merged-PR or failed-CI identity, checks out the contained SHA, and produces/verifies the exact full cache. An hourly schedule and a main-only manual dispatch repair the event-time main SHA after a direct break-glass push or cache eviction. The dispatch exception is intentional: unlike an ordinary event created by GITHUB_TOKEN, GitHub allows repository_dispatch to start another workflow. A legacy built-in-token merge can suppress the closed event too, which is why the required Test identity gate and dedicated Reviewer Router App are still mandatory.

A cold miss can still occur during a producer race or after cache eviction, but it remains fail-safe: the PR recomputes the authoritative baseline with a 30-minute job budget and saves a PR-scoped copy for same-PR reruns. It is no longer possible for a supported main-advance path to omit its producer silently. That PR-scoped fallback save remains a best-effort acceleration and does not replace the normal push, metadata, Formula, and merged-PR repair producers. Supporting and (when platform-selected) native profiles are generated before the aggregate Coverage context evaluates them. The aggregate and native gates require 100% coverage for changed executable Go statements. Overall coverage remains an unrounded, zero-tolerance, scope-matched merge-base non-regression check. Candidate and baseline profiles are evaluated by the same block-deduplicating checker; supporting policy and shortcut profiles contribute to changed-code coverage only. The checked-in badge is presentation only and is never read as a gate input.

CLI 兼容检查只使用 modern Interface Snapshot 这一处权威比较 seam,并从 PR merge-base 和最近的可达 stable release 生成权威快照。本治理机制合入后, merge-base 拥有生成器、比较器和已审批迁移清单,因此 candidate 不能通过修改 helper、fixture 或在同一 PR 新增 self-approval 记录来放行 breaking change。首次 bootstrap 仍由 merge-base 已有的 modern helper 做无豁免比较,并只接受 candidate 提交中的规范空清单;完整边界见下方治理文档。

精确的两阶段 flag 迁移生命周期见 CLI flag 兼容迁移治理。治理 PR 只能在 surface 未变化时新增 pending;后续产品 PR 达到审批的精确 surface 后,才能 消费 base-owned 记录并改为 consumed。在 main 与 stable 都达到 after 状态前 必须保留该回执,之后再由单独 PR 清理。机制只放行记录中的 legacy visible-to-hidden,以及 canonical required 新增或提升;删除、type、scope、 shorthand、no-opt 和任何无关漂移仍然阻塞。Schema 可以新增;历史 product、 tool、parameter、mapping、positional execution、constraint 与 safety 语义继续 受保护。alias_of 只是一项由 FlagSpec.Aliases 产生的框架关系证据,不是 payload 等价证明;产品 PR 仍须证明 canonical 与 legacy 的最终运行 payload 等价并在 transport 前拒绝冲突输入。当前迁移清单为空,不授权 PR #904。

Required GitHub repository settings

The main quality ruleset must enable strict required-status-check policy (strict_required_status_checks_policy=true) so a PR is revalidated whenever main advances. Every entry must select the GitHub Actions App (integration_id=15368), not “any source”. It must require these exact context/source pairs and no legacy aliases:

  • Lint
  • Test
  • Coverage
  • Policy
  • Edition
  • Interface Integrity
  • AI Behavior
  • CLI Smoke
  • Mock MCP

Do not require helper jobs, Multi-profile E2E, or an aggregate admission alias. Update ruleset contexts only after the new names have appeared on the protected branch, so a rename cannot silently remove enforcement or leave an unproducible required context.

The branch ruleset also requires one approval after the latest push. Enable repository auto-merge and automatic head-branch deletion; keep the base-owned reviewer router outside the required-context list. Install its dedicated GitHub App only on this repository with Contents: read and write and Pull requests: read and write; do not grant Actions, Workflows, or Administration. Give it pull-request-only bypass on main-merge-writers and no bypass on any other ruleset. Store the App client ID and lowercase slug in repository variables REVIEWER_ROUTER_APP_CLIENT_ID and REVIEWER_ROUTER_APP_SLUG, and its private key in repository secret REVIEWER_ROUTER_APP_PRIVATE_KEY. Do not reuse release, Homebrew, or personal tokens for this boundary.