Make your computer a home for your AI partner. Avibe is the local-first Agent OS: Vibey, its built-in AI partner, works out of the box, and Claude Code, Codex & OpenCode run unchanged beside it — reach them from your browser, phone, or any chat app.
Python agentagent-osaiai-agentsai-partnerchatopsclaudeclaude-codecodexdevtoolsdiscord-botlark-botllmlocal-firstlong-term-memoryopencodeslack-botvibe-codingvibeywechat
docs(pr-delivery): separate gate evidence from user reports (#2684)
Call White4723deba40bb
diff --git a/.agents/skills/pr-delivery-loop/SKILL.md b/.agents/skills/pr-delivery-loop/SKILL.mdindex d8f7d1f57..0df942b07 100644--- a/.agents/skills/pr-delivery-loop/SKILL.md+++ b/.agents/skills/pr-delivery-loop/SKILL.md@@ -142,7 +142,8 @@ do not load them routinely. They introduce no additional delivery gates.their drafts, report the blocker, keep observation live, and continue safework. Retry the reply after they submit/discard it; do not hide the open thread.- Notify findings/errors promptly; keep half-success silent until CI/review- both pass. Fix reports include PR/head; no round-summary PR comments.+ both pass. Internal round notes name the PR/head, findings addressed, and+ remaining gates in one or two lines; do not add round-summary PR comments.## 4. Complete delivery@@ -155,8 +156,12 @@ Before closing out, verify all review gates together:evidence blocks the gate.3. Zero unresolved threads across the entire PR, including older/outdated heads.-The final report names the repository, PR URL, reviewed head, changes,-validation layers, and residual manual/E2E work.+Keep exact head-bound evidence in the internal gate record and lane reports.+For the user, briefly explain residual issues (or none known), what changed and+why it matters, and the next step/merge readiness. Include a clickable PR link.+SHAs, run/check/thread/Watch counts, and gate-state dumps belong in the evidence,+not the default reply; surface them only to explain a blocker or on request.+Do not turn these questions into a rigid report template.In a user-started orchestrator session, close out in the current turn: verifythe gates, remove the Watch, then send the final report to the user. Do not@@ -190,15 +195,10 @@ deployment. Neither merge nor close-out authorizes deployment or a restart.## 5. User-facing next actions-Only the orchestrator's final **user-facing** report offers action buttons,-never an internal lane report or GitHub comment. Name one repository/PR and its-reviewed head. Use actual destination capability and reply-enhancement settings:-when supported and enabled, append the trailing `---` and `[label]` syntax from-Avibe's `core/prompts/quick-replies.md`, outside code fences. Otherwise, including-unknown capability/settings, use ordinary visible prose with the PR URL in each-suggested reply, without a trailing separator/bracket row that could be stripped.--Use the conversation language, English when unknown, and honor explicit labels:+Only the orchestrator's user-facing delivery report offers PR action buttons,+not internal lane reports or GitHub comments. Use the conversation language and+honor requested labels. When buttons are unsupported, disabled, or unknown,+use ordinary visible prose with the target PR URL, not a strip-prone footer.| Action | Chinese | English || --- | --- | --- |@@ -206,63 +206,28 @@ Use the conversation language, English when unknown, and honor explicit labels:| Update local master regression | 更新回归 | Update master || Update and verify end to end | 更新并验证 | Update + E2E |-**Bind every action to its target before mutation.** The callback submits only-the label; adjacent prose and Markdown links add no hidden payload. Recover the-source report from actual source-message metadata when available. If unavailable-or not uniquely bound to a repository/PR, ask for the PR URL. Never select the-latest mention, current worktree, or only open PR. An explicit user target-resolves ambiguity without extra confirmation; do not change requested labels-or invent a payload format.--- Offer **Merge** only when the review, CI, and thread gates pass and the PR is- open, non-draft, and `CLEAN`. Offering it is not authorization. After binding a- click, re-fetch the head and gates; if the offered head changed or readiness- was lost, report that change and refresh the offer instead of using stale approval.-- After GitHub confirms the named PR is **MERGED**, replace Merge with both- regression choices and name the local target. Closed/unmerged is not merged.- State in the report that both choices update from master; keep labels short.- Update + E2E performs the same update, then relevant end-to-end verification.- Both are separate opt-ins, not implied by merge. For Chinese reports:-- ```text- ---- [更新回归] | [更新并验证]- ```+- When review, CI, and thread gates pass and the PR is open, non-draft, and+ `CLEAN`, offer Merge with the PR link unless the owner asked not to offer it.+ Lack of auto-merge authority does not hide this option; offering is not+ executing. A single Merge button is enough.+- Before offering a mutating action, persist the target PR and exact offered head+ in source-bound action metadata or the internal source report; the visible+ report may omit the SHA, but the callback must be able to compare that bound+ head with the freshly fetched head. If the destination cannot retain this+ binding, do not offer a label-only mutation button.+- Bind a click to its source report using actual source-message metadata, or to+ an explicit user-named target. Labels and adjacent links carry no hidden PR or+ SHA payload. If the target is ambiguous, ask for the PR URL; never guess from+ the latest mention, current worktree, or only open PR. Re-fetch the bound head+ and gates; if the offered head changed or readiness was lost, refresh the+ offer instead of using stale approval.+- After GitHub confirms **MERGED**, replace Merge with both regression choices+ only where that repository documents a local workflow. Name the local target;+ both update from master, and Update + E2E also verifies the affected flows.+ Neither runs without a separate opt-in.### Authorized local regression updates-For Avibe, bind the action and recheck `MERGED`, then follow-`docs/regression/README.md`. Other repos offer these choices only when their own-documented local workflow applies; never invent a target.--1. Fetch origin and fast-forward the **primary** checkout with `--ff-only`;- verify `master` equals the fetched `origin/master` SHA. Establish that the- actual `sync_source()` input, including ignored deployable files, matches that- Git tree under the runner's sender exclusions, and keep it unchanged through- the update. Git status or `dirty=false` alone is not proof. A mismatch or an- unprovable match blocks mutation; never stash, discard, or commit user edits- to bypass it. Invoke the runner from this verified checkout, not a task tree.-2. Confirm the persistent **local** master environment (`avr-master` /- `avibe-master` by default), product config, and runtime environment file exist- and are readable. Missing state is a blocker, not provisioning/re-seeding- authority. Use the documented local Lima route on macOS; preserve existing- bind/port settings explicitly when different from defaults.-3. For the existing target, run from the verified checkout:-- ```bash- python3 scripts/incus_regression.py up --target master --reset-mode none --clean --env-file /dev/null- ```-- `--env-file /dev/null` suppresses automatic env discovery and leaves- `/etc/avibe-regression.env` untouched; an empty regular file does not.- Stop before mutation if the checked-out runner cannot preserve this behavior.- `--clean` reconciles disposable source/build output, including same-size,- same-mtime stale bytes. Preserve credentials, runtime environment, product- state, pairing, agent homes, and sessions; neither choice authorizes a reset.-4. Let the runner rebuild assets. Verify the receipt and served source commit- match the intended SHA and the runtime environment is unchanged without- exposing values, before success or E2E checks. Report environment, deployed- commit, and actual test results; health alone is not an end-to-end pass.--Neither choice authorizes production, a remote tenant, or the running local Avibe-to be updated or restarted.+After the owner opts in, read [Local regression](references/local-regression.md)+before any update. It preserves the existing local target, source-verification,+and state-protection procedure; do not load it during routine PR review.diff --git a/.agents/skills/pr-delivery-loop/references/local-regression.md b/.agents/skills/pr-delivery-loop/references/local-regression.mdnew file mode 100644index 000000000..e3081ec8d--- /dev/null+++ b/.agents/skills/pr-delivery-loop/references/local-regression.md@@ -0,0 +1,35 @@+# Local regression++For Avibe, bind the action and recheck `MERGED`, then follow this procedure. Other repos offer these choices only when their own documented local workflow applies; never invent a target.++1. Fetch origin and fast-forward the **primary** checkout with `--ff-only`;+ verify `master` equals the fetched `origin/master` SHA. Establish that the+ actual `sync_source()` input, including ignored deployable files, matches that+ Git tree under the runner's sender exclusions, and keep it unchanged through+ the update. Git status or `dirty=false` alone is not proof. A mismatch or an+ unprovable match blocks mutation; never stash, discard, or commit user edits+ to bypass it. Invoke the runner from this verified checkout, not a task tree.+2. Confirm the persistent **local** master environment (`avr-master` /+ `avibe-master` by default), product config, and runtime environment file exist+ and are readable. Missing state is a blocker, not provisioning/re-seeding+ authority. Use the documented local Lima route on macOS; preserve existing+ bind/port settings explicitly when different from defaults.+3. For the existing target, run from the verified checkout:++ ```bash+ python3 scripts/incus_regression.py up --target master --reset-mode none --clean --env-file /dev/null+ ```++ `--env-file /dev/null` suppresses automatic env discovery and leaves+ `/etc/avibe-regression.env` untouched; an empty regular file does not.+ Stop before mutation if the checked-out runner cannot preserve this behavior.+ `--clean` reconciles disposable source/build output, including same-size,+ same-mtime stale bytes. Preserve credentials, runtime environment, product+ state, pairing, agent homes, and sessions; neither choice authorizes a reset.+4. Let the runner rebuild assets. Verify the receipt and served source commit+ match the intended SHA and the runtime environment is unchanged without+ exposing values, before success or E2E checks. Report environment, deployed+ commit, and actual test results; health alone is not an end-to-end pass.++Neither choice authorizes production, a remote tenant, or the running local Avibe+to be updated or restarted.diff --git a/.agents/skills/pr-delivery-loop/references/rationale.md b/.agents/skills/pr-delivery-loop/references/rationale.mdindex f4e9ce546..41d80b928 100644--- a/.agents/skills/pr-delivery-loop/references/rationale.md+++ b/.agents/skills/pr-delivery-loop/references/rationale.md@@ -40,6 +40,14 @@ A lane Run marked `succeeded` proves only that invocation ended successfully.It does not prove that the current-head review started, the Watch remains live,or the orchestrator received a report. Those are separate observations.+## Gate evidence and user reports are different surfaces++Exact head, run, thread, and Watch evidence protects the delivery gate, but it is+not the user-facing result. Keep that evidence in the internal record; the final+report should state residual issues, the change and its benefit, the next action,+and merge readiness. Expose identifiers only when they explain a blocker or the+user asks for evidence.+## A repeated finding calls for diagnosis, not automatic expansionIf a reviewer flags a searchable row's projected title, similar fields maydiff --git a/core/prompts/quick-replies.md b/core/prompts/quick-replies.mdindex 40673e6ea..e08a816c6 100644--- a/core/prompts/quick-replies.md+++ b/core/prompts/quick-replies.md@@ -1,10 +1,12 @@## Quick-reply buttons-At the very end of the message, add a `---` separator followed by `[button text]` to provide clickable quick replies. Example:+Offer buttons for useful next actions or decisions, not to fill every reply.+When none is useful, omit the footer. A single relevant action is enough.++When buttons are supported and enabled, put this exact multiline footer at the very end:+```text----[👌 Continue] | [✅ Submit PR] | [👀 Review first]-Rules:-- Think through the tacit knowledge behind the user's words, infer their deeper intent, and suggest likely next replies from the conversation context and the user's habits-- Do not add filler unrelated to the user's likely next intent, such as: got it, received, thanks-- They must appear at the very end of the message, after the `---` separator-- Wrap each button in `[text]` and separate them with `|`; you may start with emoji to improve clarity-- Use at most 2-4 buttons, each no longer than 20 characters+[button text] | [another action]+```+Use at most four short labels, each no longer than 20 characters, in the+conversation language. Do not invent filler choices or treat displaying an+action as authorization to execute it.diff --git a/tests/test_harness_skill_guidance.py b/tests/test_harness_skill_guidance.pyindex 72f33e903..d8d12e77e 100644--- a/tests/test_harness_skill_guidance.py+++ b/tests/test_harness_skill_guidance.py@@ -240,28 +240,45 @@ def test_background_watch_skill_uses_bundled_waiters_and_persists_managed_pr_staassert '--workflow lint' in body-def test_pr_delivery_loop_keeps_one_optional_reference_out_of_routine_loading() -> None:+def test_pr_delivery_loop_keeps_optional_references_out_of_routine_loading() -> None:directory = ROOT / ".agents/skills/pr-delivery-loop"body = (directory / "SKILL.md").read_text()flat = " ".join(body.split())references = re.findall(r"\]\((references/[^)]+\.md)\)", body)- assert references == ["references/rationale.md"]+ assert references == ["references/rationale.md", "references/local-regression.md"]assert {path.relative_to(directory).as_posix() for path in directory.rglob("*.md")} == {"SKILL.md", *references,}assert "are optional reading" in flat and "do not load them routinely" in flatassert "They introduce no additional delivery gates." in flatassert "sync the **whole maintained package**, including references, byte-for-byte" in flat+ assert "Keep exact head-bound evidence in the internal gate record" in flat+ assert "residual issues (or none known)" in flat+ assert "gate-state dumps belong in the evidence" in flat+ assert "offer Merge with the PR link" in flat+ assert "persist the target PR and exact offered head" in flat+ assert "visible report may omit the SHA" in flatskill = parse_skill_file(directory / "SKILL.md", priority=(1, 0, 1))assert skill is not Nonerendered = render_skill_content(skill)assert "## 4. Complete delivery" in renderedassert "### Authorized local regression updates" in rendered- reference = (directory / references[0]).read_text()- assert reference.startswith("# Delivery Rationale and Examples")- assert reference not in rendered+ for reference_path in references:+ reference = (directory / reference_path).read_text()+ assert reference not in rendered+++def test_quick_replies_are_action_driven_not_mandatory() -> None:+ body = _read("core/prompts/quick-replies.md")+ flat = " ".join(body.split())+ assert "Offer buttons for useful next actions or decisions" in body+ assert "When buttons are supported and enabled, put this exact multiline footer" in body+ assert "---\n[button text] | [another action]" in body+ assert "When none is useful, omit the footer" in body+ assert "Do not invent filler choices" in body+ assert "treat displaying an action as authorization" in flatdef test_agent_prompt_audit_examples_parse_against_the_real_cli() -> None:diff --git a/tests/test_prompt_render_export.py b/tests/test_prompt_render_export.pyindex 96fb31436..260934b42 100644--- a/tests/test_prompt_render_export.py+++ b/tests/test_prompt_render_export.py@@ -522,6 +522,7 @@ def test_approved_guidance_changes_preserve_all_other_injection_bytes(monkeypatc"skills-prompt","skills-manual-prompt","codex-skill-reuse",+ "quick-replies-prompt",}for backend, history, skill_mode in itertools.product(("claude", "codex", "opencode"), ("off", "managed", "self-managed"), ("empty", "manual", "pages"),@@ -536,10 +537,10 @@ def test_approved_guidance_changes_preserve_all_other_injection_bytes(monkeypatctext = text.replace(_FOLLOW_THROUGH_PRINCIPLE, _PREVIOUS_FOLLOW_THROUGH_PRINCIPLE)outputs.append(text.replace(_EVIDENCE_LED_PRINCIPLE, _PREVIOUS_VERIFICATION_PRINCIPLE))digest = hashlib.sha256(json.dumps(outputs, ensure_ascii=False).encode()).hexdigest()- # Captured before editing from 1e9ba96bc61b0e86027d4871553e68e8ed85c1ac,- # omitting only the approved Computer Use and Skill-loading guidance blocks- # (including the Codex-only reuse sentence) and restoring the previous- # verification principle and conversation-history guidance and Follow-through- # principle.+ # Captured before editing from b4eb5cca140955d7e6f7f4206c793892cf19ef59,+ # omitting only the approved quick-reply, Computer Use and Skill-loading+ # guidance blocks (including the Codex-only reuse sentence) and restoring the+ # previous verification principle and conversation-history guidance and+ # Follow-through principle.# Every other byte/order stays pinned, including all Claude/OpenCode output.- assert digest == "262c7cbabbbb0bfefa4a69511d805d28f76a3559bd511184006242f831b4836e"+ assert digest == "9475df69a6c14f6ddda396dc33d665f4ebda657517cd3ea3134d23d7f5c5c712"