01a07197 - Define immediate draft publication and CI failure ownership

AGENTS.md

@@ -1,12 +1,15 @@
 # Agent
 
-Read [CONTRIBUTING.md](CONTRIBUTING.md) and [DESIGN.md](DESIGN.md) before
+Read [CONTRIBUTING.md](CONTRIBUTING.md), [DESIGN.md](DESIGN.md), and
+[docs/pull-request-lifecycle.md](docs/pull-request-lifecycle.md) before
 changing this repository. Skill contracts live next to the client:
 `agent skills path` (spine, review-loop, pr-review, error-fix).
 
-A draft plus local tests is not done. Pull-request done is signed commits
-on a branch in this repository, grok quality and logic then Codex quality
-and logic on this head with zero findings, CI green on this head, then
-ready. The authoring session does not sit those PR reviews. A human merges.
-The local-CI comment schema for private product repositories is
-`docs/local-ci-v1.md`.
+Draft publication is immediate after the first signed task commit; see the
+lifecycle. A draft plus local tests is not done. Ready for review is signed
+commits on a branch in this repository, grok quality and logic then Codex
+quality and logic on this head with zero findings, CI green on this head, then
+leave-draft. Ready for review is still not merge and not completion. The
+authoring session does not sit those PR reviews. A human merges; claim
+completion only after that merge is verified. The local-CI comment schema for
+private product repositories is `docs/local-ci-v1.md`.

CONTRIBUTING.md

@@ -2,28 +2,28 @@
 
 - Branch from `develop`. Never push to `develop` or `main`.
 - Push the branch to this repository. Do not open the pull request from a personal fork.
-- Open a draft pull request. Stay draft until the pull request is **done** (below). A human merges.
+- As soon as the first signed task commit exists, push and open a **draft** pull request immediately ([docs/pull-request-lifecycle.md](docs/pull-request-lifecycle.md)). Stay draft until **Ready for review** (below). A human merges; only then is the pull request completed.
 - Sign commits with the GitHub identity that owns the commits.
 - Public repository: English for commits and comments. The visible pull-request summary is an `EN:` block, optionally followed by a labeled `DE:` block.
 - Do not name private repositories, internal hostnames, or internal infrastructure.
 - Add or update tests in the same change.
-- Run `pytest` before you push. Tests need PostgreSQL (`AGENT_TEST_PG` or a local `initdb`).
-- Pytest (or any green local suite) is a **check**, not done.
+- Run `pytest` on the exact clean signed final head before Ready for review. Tests need PostgreSQL (`AGENT_TEST_PG` or a local `initdb`). Full pytest is not a gate for the first draft publication.
+- Pytest (or any green local suite) is a **check**, not Ready for review and not completion.
 
-## Pull request done
+## Ready for review
 
-A draft plus local tests is not done. Do not claim the pull request is finished at that point.
+A draft plus local tests is not done. Do not claim the pull request is finished, done, or completed at that point — including after leave-draft. Draft timing and CI ownership while the draft is open are defined in [docs/pull-request-lifecycle.md](docs/pull-request-lifecycle.md).
 
-Done is all of:
+Ready for review requires all of:
 
 1. Signed commits on a branch in this repository, based on `develop`.
 2. Four lane verdicts on **this** head, two vendor stages: grok quality and grok logic in parallel, then Codex quality and Codex logic. Quality/conformance reads this file first. The session that authored the diff does not sit those reviews.
 3. Codex runs only if both grok dimensions are approved. If a vendor cannot run, abort loudly; do not record `approved`; do not substitute another vendor.
 4. Zero findings only after an explicit complete pass. Empty, partial, timeout, or unavailable output is not zero findings. Iterate until all four lane verdicts on this head are approved.
 5. Inner implement/review rounds (`review-loop`) are not the PR reviews (`pr-review`).
 6. CI green on **this** head. This public repository uses GitHub Actions. `skipped` and `cancelled` are not green unless the workflow documents that skip. The local-CI comment schema for **private** product repositories is defined in [docs/local-ci-v1.md](docs/local-ci-v1.md) and verified by `agent local-ci verify`.
-7. Stay draft until the reviews and CI above hold on this head. Then one comment whose review-pass count is those four `approved` verdicts on this head, then mark the GitHub pull request ready. When spine and pr-review are attached, `agent allow --action pr-ready` only checks task state (`pushing` or `pr-review`); it is not the leave-draft verdict. Do not mark ready if it denies.
-8. A human merges. When spine is attached, `agent allow --action task-done` still needs the workflow checklist and both summary sentences.
+7. Stay draft until the reviews and CI above hold on this head. Then one comment whose review-pass count is those four `approved` verdicts on this head, then mark the GitHub pull request ready for review (`isDraft=false`). When spine and pr-review are attached, `agent allow --action pr-ready` only checks task state (`pushing` or `pr-review`); it is not the leave-draft verdict. Do not mark ready if it denies. Ready for review is still not merge and not completion.
+8. A human merges. Claim pull-request completion only after that merge is verified. When spine is attached, `agent allow --action task-done` still needs the workflow checklist and both summary sentences; that ledger state is not proof of pull-request completion.
 
 The AI inserts `pr.open` / `comment.post`; a rejected review gate inserts `review.post`. `agent github pending` performs GitHub HTTP. A retry reuses the existing draft.
 

DESIGN.md

@@ -44,7 +44,7 @@ The AI session talks **only** to the local database. Scripts perform every actio
 | Outside facts | Scripts notice GitHub (and other outside) state. The agent is not told by a human and does not poll GitHub. Example: a recorded PR merges → script writes `pr.merged` on that session and knocks. |
 | AI vs scripts | The AI inserts local intent. Scripts perform every side effect that leaves the machine. Model text is never a state transition. |
 | Checks and gates | A **check** records a fact (`agent check record`). A **gate** is a policy verdict over evidence (`agent gate record`). A model claim is neither. Confidence is not proof. |
-| Pull request done | A draft plus local tests is not done. CONTRIBUTING.md is the contract for this repository. When spine and pr-review are attached, grok then Codex on this head are the gates; `agent allow --action pr-ready` only checks task state. A human merges. |
+| Ready for review | Draft publication is immediate after the first signed task commit ([docs/pull-request-lifecycle.md](docs/pull-request-lifecycle.md)). A draft plus local tests is not done. CONTRIBUTING.md is the contract for this repository. When spine and pr-review are attached, grok then Codex on this head are the gates; `agent allow --action pr-ready` only checks task state. Leave-draft is Ready for review, not completion. A human merges; claim completion only after that merge is verified. |
 | Local CI report | Frozen comment schema `dfx-local-ci/v1` in [docs/local-ci-v1.md](docs/local-ci-v1.md). `agent local-ci verify` parses it and computes pass/fail. Private product repositories attach the block to the ready comment. This public client still uses GitHub Actions for its own PRs. |
 | A38 | The complete central process is [docs/a38.md](docs/a38.md) plus [docs/a38-guard.md](docs/a38-guard.md). An adopter keeps only `.github/a38.json`, its integration, and a contributing pointer; plugins use short central pointers. `agent a38` runs and validates the existing report format, while `agent pr-guard` checks live author evidence without executing PR code. Private trusted-base opt-in may provide local code-gate equivalence; public/default GitHub CI, independent required GitHub-only checks, review gates, technical merge restrictions, and human merge remain unchanged. |
 | Merge | The client never merges. A human merges. |
@@ -508,7 +508,7 @@ The rules below were already implied by §§1–17. They are now explicit so a l
 | Whether a command ran and what it returned | The process that ran it | `agent check record` (name, command, `pass`/`fail`/`skip`, output) |
 | Merge | Human | Never a client command |
 
-A worker report such as “analysis complete” or “tests passed” is **input**. It is not the transition. Opening a draft is not done.
+A worker report such as “analysis complete” or “tests passed” is **input**. It is not the transition. Opening a draft is not done. Leaving draft is Ready for review, not completion.
 
 No transition that needs deterministic evidence may be satisfied by model text alone. Malformed structured output is rejected (unknown `activity.type` → `execution_status=error`; empty, partial, timeout, or unavailable review output is not zero findings). A patch that does not apply is a failed check, not a debate.
 
@@ -549,13 +549,13 @@ Partial multi-step actions (push, then open a pull request) record each complete
 
 Write owner, hub role, opt-in skills, required session row, and the generic `activity` catalog stay as in §2. Spine task states stay the spine skill’s states. They are not replaced by a hub machine such as `CREATED` / `ANALYZING` / `READY_FOR_PR`.
 
-### 19.6 Pull request done
+### 19.6 Ready for review
 
-A draft plus local tests is not done. A check records the local suite. When spine and pr-review are attached, a gate records each vendor dimension on **this** head, and leave-draft is four lane verdicts (grok quality and grok logic, then Codex quality and Codex logic) on this head plus CI green on this head. Without those skills, the target repository’s written contributing rules apply. `agent allow --action pr-ready` only checks that a task is in `pushing` or `pr-review`; it is not the leave-draft verdict. `task-done` still needs the workflow checklist and both summary sentences.
+Draft publication timing is [docs/pull-request-lifecycle.md](docs/pull-request-lifecycle.md): open the draft as soon as the first signed task commit exists; full tests and reviews are Ready-for-review gates, not draft gates. A draft plus local tests is not done. A check records the local suite. When spine and pr-review are attached, a gate records each vendor dimension on **this** head, and leave-draft (Ready for review) is four lane verdicts (grok quality and grok logic, then Codex quality and Codex logic) on this head plus CI green on this head. Without those skills, the target repository’s written contributing rules apply. `agent allow --action pr-ready` only checks that a task is in `pushing` or `pr-review`; it is not the leave-draft verdict. `task-done` still needs the workflow checklist and both summary sentences; ledger `task-done` is not pull-request completion.
 
 Quality and logic of one vendor stage run together. Vendors are `grok`, then `codex`. Codex runs only after both grok dimensions are `approved`. The session that authored the diff does not sit those PR reviews. If a vendor cannot run, abort loudly; do not record `approved`; do not substitute another vendor. Empty, partial, timeout, or unavailable review output is not zero findings.
 
-CI on this head is a script-measured fact whose applicability comes from the target repository's written rules. The frozen `dfx-local-ci/v1` format and legacy `agent local-ci verify` behavior do not by themselves adopt A38 or grant permission to skip GitHub CI. For A38 adopters, follow the central [A38 standard](docs/a38.md) and [guard guide](docs/a38-guard.md): private visibility alone is not opt-in, and private local code-gate equivalence requires trusted-base opt-in through a valid A38 manifest, assessment against the canonical active policy, and a separate live join against the actual latest report-like GitHub comment by the PR author. Public A38 adopters publish and validate the author report in addition to retaining cumulative GitHub CI; private repositories without that trusted-base opt-in and non-A38 repositories retain their existing written CI rules. For applicable GitHub CI checks, `skipped` and `cancelled` are not green unless the workflow documents that skip. Independently required GitHub-only checks, technical merge restrictions, review gates, and human merge remain required. Stay draft until the applicable rules hold. One comment whose review-pass count is those four `approved` verdicts on this head, then ready. A retry reuses the existing draft. A human merges.
+CI on this head is a script-measured fact whose applicability comes from the target repository's written rules. The frozen `dfx-local-ci/v1` format and legacy `agent local-ci verify` behavior do not by themselves adopt A38 or grant permission to skip GitHub CI. For A38 adopters, follow the central [A38 standard](docs/a38.md) and [guard guide](docs/a38-guard.md): private visibility alone is not opt-in, and private local code-gate equivalence requires trusted-base opt-in through a valid A38 manifest, assessment against the canonical active policy, and a separate live join against the actual latest report-like GitHub comment by the PR author. Public A38 adopters publish and validate the author report in addition to retaining cumulative GitHub CI; private repositories without that trusted-base opt-in and non-A38 repositories retain their existing written CI rules. For applicable GitHub CI checks, `skipped` and `cancelled` are not green unless the workflow documents that skip. Independently required GitHub-only checks, technical merge restrictions, review gates, and human merge remain required. Stay draft until the applicable rules hold. One comment whose review-pass count is those four `approved` verdicts on this head, then mark Ready for review (`isDraft=false`). A retry reuses the existing draft. Ready for review is still not merge and not completion. A human merges; claim completion only after that merge is verified.
 
 ## 20. Refused: hub as a coding control plane
 
@@ -670,8 +670,8 @@ On `error.fix`:
 
 1. `agent task create --workflow implement --error-id <error.seen-id>` on this session (find-or-create). That copies `error_id` and `repo` from the `error.seen` row into the task payload.
 2. Isolated worktree of that task `payload.repo` at the allowed base revision. Git operations are scripts. `payload.repo` is already on the task because analysis refused `error.fix` when `repo` was missing. Never fall back to the origin checkout.
-3. Spine implement: mandatory checks must `pass`, then `pr.open` opens a **draft** (spine `pushed`). Title/body may be model-drafted; the GitHub API call is a script. A retry finds an existing draft for this fingerprint instead of opening a second one.
-4. pr-review gates run on that head after `pushed`.
+3. As soon as the first signed task commit exists, `pr.open` opens a **draft** through the existing executor flow ([docs/pull-request-lifecycle.md](docs/pull-request-lifecycle.md)). Title/body may be model-drafted; the GitHub API call is a script. A retry finds an existing draft for this fingerprint instead of opening a second one. Mandatory checks and the spine checklist key `pushed` remain final validated bookkeeping; they are not prerequisites for that first draft publication.
+4. pr-review gates and applicable full checks run on the final head before Ready; closing spine `pushed` records the validated push step, not the early draft open.
 5. A human merges. `pr.merged` knocks as today.
 
 The model never receives production credentials. Analysis that only reads the excerpt does not need write access to the origin branch.

docs/a38-guard.md

@@ -98,7 +98,7 @@ For each reviewer, the latest substantive submitted state controls authorization
 
 An authorized migration may introduce, remove or change workflows, and may also add, remove or change `.github/pr-guard.json`, but the complete current head inventory and author report still must satisfy the approved head policy. Proposed pr-guard configuration is never used for scope until it is merged to the trusted default-branch revision. The executable guard remains pinned and never runs head commands. The **base policy's enforcement mode stays active** for this PR, even if the proposed mode is `observe`. If no valid base policy exists, explicit approval permits bootstrap under `enforce`; initial adoption cannot silently bypass reporting.
 
-An unpublished migration proposal is first measured and verified against the still-active base policy and pushed at that same commit; that run is proposal publication evidence, not Ready evidence. Once the exact current head/base approval exists, rerun the full approved head policy and local verification, publish that newly generated report, and perform the live join. Bootstrap with a missing or invalid base policy instead follows the repository's existing pre-push checks to publish the proposal, then requires explicit approval and `enforce`; it must not invent a report or waive approval.
+Publish the migration proposal draft first per the [pull request lifecycle](pull-request-lifecycle.md). Measurement and verification against the still-active base policy may follow on that same signed commit; that run is proposal evidence, not Ready evidence. Once the exact current head/base approval exists, rerun the full approved head policy and local verification, publish that newly generated report, and perform the live join. Bootstrap with a missing or invalid base policy likewise publishes the draft first under the lifecycle (full pre-push test gates are not a draft blocker), then requires explicit approval and `enforce`; it must not invent a report or waive approval.
 
 The bot identifies the active policy revision in its comment. Download `.github/a38.json` from that exact revision before generating the report. For ordinary PRs, this is the base; for explicitly approved migrations, it is the head.
 
@@ -114,7 +114,7 @@ agent a38 run --repo . --repository OWNER/NAME \
 
 `--repository` identifies the target repository, especially when the checkout origin is a fork. Post the complete generated report as a PR comment using the **PR author's account**. Preserve its JSON and markers. The existing local-CI wire schema remains unchanged for compatibility.
 
-For the private opt-in process, the checkout must be clean at the final repository-required signed commit before measurement. Run the full active policy and locally verify it before push, record that verified report as `local_check_pass` evidence, then push the same SHA without an intervening commit. Any fix, amend, or rebase creates a new SHA and requires the complete run and verification again. Execution roles follow the repository's orchestration rules; reviewers remain read-only. Job adapter commands are catalogued in [A38 job adapters](a38-job-adapters.md), without duplicating their schemas here.
+For the private opt-in process, the checkout must be clean at the final repository-required signed commit before Ready measurement. The draft may already exist under the [pull request lifecycle](pull-request-lifecycle.md). Run the full active policy and locally verify it before recording `local_check_pass` evidence; that verified SHA must be on the open draft with no intervening commit after measurement. Any fix, amend, or rebase creates a new SHA and requires the complete run and verification again. Execution roles follow the repository's orchestration rules; reviewers remain read-only. Job adapter commands are catalogued in [A38 job adapters](a38-job-adapters.md), without duplicating their schemas here.
 
 The latest author report-like comment, ordered by `updated_at` and numeric comment ID, is authoritative. A newer malformed or failed report never falls back to an older success. Other authors' reports cannot satisfy the requirement. Matching repository, head, visibility, full job set, names, commands, timeouts and successful measured results are mandatory, including for public repositories.