Acceptance (Builder Self-Evidence)

You are the builder for a delivery. A separate review step judges it against a plan — checks you author, or a verify plan handed to this run. A check that declares requiredEvidence cannot pass on your text alone: a missing artifact marks it uncertain and holds the delivery.

author (or discover) the plan  →  pick the surface  →  capture evidence  →  publish the round  →  self-check coverage

Decide whether to execute before starting a round

Creating or updating a PR, marking it ready, or being asked to upload a report must not by itself start another verification run. First inspect the requested scope and the task's existing reports, evidence, and published acceptance links (from the conversation, PR, or local .acceptances/ directory).

Delivery state	Action
Documentation/instruction-only change, or pure refactor/tooling change with no product behavior change	Skip product acceptance and briefly state why. Keep any applicable quality checks.
Gitlink-only sync	Do not launch a fresh acceptance. Link the upstream change and its existing acceptance when available; disclose missing upstream evidence without claiming it passed. Cloud changes accompanying the sync are assessed separately.
Completed acceptance already published and still covers the delivery	Reuse its URL and coverage. Do not create a round or rerun cases just for the PR.
Completed acceptance report and evidence exist locally and still cover the delivery	Inspect coverage and artifacts, then upload that report using report.md. Preserve the original execution provenance; no product rerun, new plan, or repeated completed checker review is needed merely for upload.
The delivery was already exercised on the real product earlier in this session (observations and raw artifacts exist, no report yet)	Do not rerun, re-plan, or open a checker stage. Write plan[] and cases[] from the observations already made, attach the original artifacts (logs, command output, captures) with their original provenance, disclose any required medium that was never captured instead of recapturing it, and ingest.
Product behavior lacks valid evidence, or relevant behavior changed after verification	Execute only the missing or affected outcomes, retain unaffected evidence with its original provenance, and publish according to the round rules below.

Evidence is reusable when its criteria cover the requested behavior, its artifacts are available and support the observations, and subsequent code, dependency, configuration, or environment changes do not invalidate those observations. Compare the relevant changes; a different commit SHA, rebase, PR event, or report publication status alone is not a reason to rerun. Failed/blocked checks and missing required evidence are not passes: repair or supplement those specific gaps. An explicit user request for fresh verification still takes precedence.

The execution, environment setup, plan/checker, and capture sections below apply when executing acceptance. For reuse or upload only, inspect the existing report and evidence and complete the necessary publication/coverage steps; do not boot services or replay completed cases. Uploading does not change when or against which implementation the evidence was captured.

Execution ownership and independent review

For new execution, use one fresh-context acceptance-runner for environment preflight, planning, product interaction, capture, report assembly, publication, and cleanup. Explicitly select a lighter model with the required tools; do not inherit the builder's full conversation or default high-cost model. The primary keeps scope, product decisions, code repairs, and delivery accountability, and coordinates through brief reports. Reuse the runner for follow-ups.

Read delegation.md once to assign source ownership, model capability, budgets, and handoffs. Each unchanged source section, runtime probe, and evidence inspection has one owner: pass its acquired facts and provenance, rather than fetching it again in another agent. Workers load only their role instructions and needed references, not this entire entrypoint.

Use one acceptance-checker in the first round: at most two plan-feedback responses before execution, then exactly one final report/evidence review. It is the default owner of verdict-bearing media inspection; the runner captures and indexes that media without pre-reviewing it. Read acceptance-checker.md only in that owner. The checker is the plan gate; never ask the user for routine plan approval. The primary resolves findings and tells the runner when execution or publication may proceed. There is no per-case approval loop or repeated full review after fixes.

Follow-up rounds have no independent checker gate. Assign new artifact inspection to one capable owner and re-run only affected cases. If delegation is unavailable, use one bounded execution context and disclose missing independent review. If model selection is unavailable, record the actual or inherited model rather than claiming lower cost. Retain all evidence requirements in either fallback.

Read the project layer first

Before touching an environment, its execution owner checks for .agents/acceptance/. Reuse sections and facts already acquired for the same revision; other agents receive the relevant rules and provenance in their handoff instead of independently reading these files:

File	What it owns
PROJECT.md	Start/stop commands, ports, services, auth, surfaces, probes
PROCESS.md	The run process: plan gate, execution rules, teardown
common-mistakes.md	Project living log — what earlier rounds got wrong here
probe-mock-patterns.md	Project living log — how to force state on this product

The project layer owns how this repository is run; this skill owns what a valid round is (plan, evidence, report, immutable round, the hard rule). On running, the project layer wins; on what may be published, this skill wins. Never invent a start command, port, or auth flow PROJECT.md answers; fix a divergence in the adapter during the run instead of working around it. No .agents/acceptance/ → bootstrap one first: project-adapter.md.

Living logs — inject each by its own shape

The execution owner acquires the relevant generic and project sections once the target is known and records their versions in the shared handoff:

• common-mistakes.md — acquire its Checklist once. Assign applicable lines to the execution or inspection owner, and apply those retained rules before pass; do not re-read the file for each case. Pull an entry by id only when its detail is needed and not already acquired.
• probe-mock-patterns.md — read the heading index, then pull only the entries this round needs. Pick by meaning, not keyword; rg over the body is the fallback.

rg -n '^#{2,4} ' <file>          # the index, with line numbers
sed -n '<start>,<end>p' <file>   # one entry, in full

Record new project-specific learnings in the project layer only.

CLI discovery

Use the installed lh as the command reference. The assigned execution or publishing owner consults the relevant subcommand's --help only when syntax or capability is unknown, and records the learned usage by CLI version. Reuse that result across stages and agents; do not walk the whole help tree, repeat discovery for every case, or invoke a write command to probe support.

Prefer CLI-provided schemas, examples and non-publishing validation. Read a linked payload schema or legacy recipe only when the installed CLI lacks that capability. Help teaches syntax; the rules below retain subject/operation identity, immutable rounds, evidence coverage, language and safe recovery. Inspect structured errors and ingest warnings; successful JSON parsing or a zero exit code alone does not establish publication or complete coverage.

Two paths — no id is required

Every evidence command targets a round. The authored path is the default; you have an operation id only when the invocation names one. Never hunt the environment for one, and never report this skill inapplicable — a round without an operation is simply recorded as standalone.

You have	Path
No plan — you author the checks	Write result.json + assets/, publish with lh acceptance run ingest — report.md
An operation id you were given	Read the frozen plan, write the same round directory keyed by its check item ids, publish with ingest --operation — plan-format.md

The paths differ only in where the checks come from and which round the ingest lands in. Capture is identical: the surface guides, recordings with chapters, before/after pairs, screenshots inspected by their designated owner, proposal.md. A handed-down plan never licenses a lighter capture.

Pass --subject (task:<id> / topic:<id> / document:<id>) only when the caller named one; otherwise ingest attaches the round itself when it can and creates a standalone acceptance when it cannot. On the first ingest, always supply --requirement "<one-sentence business goal>" — the durable goal of the whole acceptance, not this round's scope; it is immutable once recorded.

Prerequisites: the execution or publishing owner runs lh acceptance preflight --json once before writing any plan or report text, or reuses its recorded result while server, scope, credentials, and language configuration are unchanged. An auth error means stop and surface it. Install only the UI driver the selected surface needs; probe before adding dependencies, and never substitute a private agent plugin. Its language.locale is the report language: write every human-readable field (title, plan, case names, observations, evidence descriptions, conclusion, report.md, proposal.md) in it, even when the prompt that invoked you is in another language — a delegating agent's prompt does not show which language the user reads. The publish commands reject text in the wrong script, so run lh acceptance run ingest <dir> --dry-run with the same target flags the publish will use as soon as the plan is in result.json, before executing it. Offline validate does not check account language or remote plan membership and never replaces this dry-run. Never publish merely to test whether a payload is accepted. When language.status is unset, use the language the user converses in. On a CLI without preflight, check auth with lh acceptance run list --json.

Optional user-journey flows

For a requirement with state transitions or independently reviewable journeys, read flows.md only when authoring or updating that graph. Ordinary checklist rounds do not load it.

HARD RULE — programmatic gates are NEVER acceptance checks

Every check MUST be an outcome a person decides about the delivery: what the user sees, hears, reads, or receives. These MUST NOT appear as a check, under any phrasing: unit / integration / regression / snapshot tests, coverage, type-check / tsc, lint / eslint, format, "compiles", "build passes", "CI is green". Run them, then report them as one line of narrative.

Enforced at ingest: every matching item (matched on title, category, AND method — "run bun run test" under a product-sounding title still matches) is dropped with a warning and summary recounted; a round of only such checks fails to publish. The line is the subject of the check, not who judged it: a CLI behavior asserted by a command is a fine check (verifier: "program"); "the suite is green" is not. Before writing any plan, ask of each draft check: would the user click accept/reject on this?

Rounds are immutable — repair means a NEW round

A published round is a permanent record. Never re-submit into a round after changing the code — publish the re-verification as the next round and let the acceptance page show the progression.

Before a repair round, read the aggregate with lh acceptance view <acceptanceId | type:id> --json. Omit checks whose latest userReview.action is accept; address non-stale rejects under their exact stable ids; when a check semantically replaces another, declare supersedes: ['old-id'] and repeat the full lineage in every later round that reuses the successor id. Pass --acceptance <acceptanceId> so the round joins the same history.

Rules you will be tempted to skip

Not judgment calls — the moves an agent under pressure makes and must not. Each excuse below was made in a real round.

Excuse	Reality
"Injection is hard; happy-path plus unit tests covers it"	The error state was the goal. Walk the probe ladder (probe-mock-patterns.md A) before calling it blocked.
"The branch name says what to verify" / "Loading the living logs first…"	The task lives in the user's words. Recover it, or confirm a labeled guess with one structured question — silently; never narrate setup.
"The black frame is probably display sleep / a permission"	Measure first: pixel brightness, the permission bit, an A/B with one variable toggled. Publish "confirmed by X" or "suspected", never a guess.
"Let me ask how they want it run" / "I'll click Sign in and you authorize" / "too small to screenshot"	Environment mechanics are yours: full isolated run, auth by direct injection (never an interactive login — it hijacks the user's browser), a screenshot for every user-facing change. Ask only about the product decision.
"One more config edit and the env will boot" / "I'll mock it" / "I'll drive the rest myself"	Timebox. Inventory running instances, probe for the real capability before mocking (a mock that records nothing is not in the path), re-delegate a dead subagent's remaining steps. Search the project layer (.agents/acceptance/, including references/) for the exact error before asking; if it has no fix, revert experiments and ask for the missing prerequisite.
"The fix is in and tests pass — verified"	Reproduce the failure's precondition first, then verify with it held. A run that cannot fail proves nothing; "reproduces sometimes" means an unnamed precondition. When the mocked seam is the suspect, drop the mock.

Pick the surface by the user-visible outcome

Match the requirement to the cheapest surface that can prove the complete outcome, not merely the layer containing the code change. A backend fix for missing cards, stale lists, navigation, or another visible behavior still requires the consuming UI, its actual data response, and inspected screenshots. Database assertions and passing tests support that evidence; they do not replace it.

What your task changed	Surface	Guide
Backend / CLI / library / data logic with no UI outcome	CLI — stdout as text, zero UI flakiness	surfaces/cli.md
Web app frontend / styles / interactions	Web (agent-browser → running web app)	surfaces/web.md
New/changed API plus the UI consuming it	Web, full-stack (agent-browser + network capture)	surfaces/web.md
Desktop-only behavior (native windows, IPC, packaged shell)	Electron (agent-browser --cdp)	surfaces/electron.md
Native macOS app / OS chrome agent-browser can't reach	Native (osascript + screencapture, local macOS)	surfaces/native.md
Native iOS behavior, gestures, device-size layout	iOS Simulator (sim-use/AXe + simctl)	surfaces/ios-simulator.md

• Use CLI alone only when the required outcome has no UI surface. If a visible outcome cannot be exercised, report that acceptance as incomplete instead of narrowing it to data checks. Use Electron only when the criterion depends on desktop-only code; iOS is driven by a Simulator HID/AX CLI, never host mouse — mark the case blocked if the CLI cannot express the gesture.
• Structured data uses native visualizations (cases[].datasets + cases[].visualizations; raw CSV/JSON stays as evidence), not a PNG — report.md. A deliverable the user hears needs audio — evidence.md.
• A change to UI users already see needs before/after evidence. When the delivery changes existing copy, labels, toasts, layout, or style, each such case requires a comparison pair of the same view on the pre-change and post-change code — report.md. An after-only shot of a toast reading 重命名遇到了问题，请重试 does not show it replaced Rename failed. Prove the before half really runs the old code (probe-mock-patterns.md A6); text that does not render visibly, such as an aria-label, also needs the DOM reading. UI with no prior state is exempt.
• Auth is a gate scoped to the surface: authenticate that surface first or every capture lands on the sign-in page. Web: auth-web.md.
• A UI round may price its interaction cost by recording KLM operator counts into interaction-trace.jsonl; optional, never hand-written — interaction-cost.md.

Every file submission MUST include a non-empty, reviewer-facing description (--desc for CLI submissions; description for tools and ingest entries). Identify what the file contains and what it demonstrates for this criterion. A filename, path, artifact id, or generic label such as "evidence" is not a sufficient description. This also applies when a text file is stored inline.

Shared rules for every artifact — media types, provenance, file vs inline, safety — are in evidence.md.

Keep checklist explanations brief

Write each check's observation and inline explanation in the report language, usually 1–3 short sentences: what was done, what happened, and any limitation needed to judge that outcome. Do not paste the execution report into the check. Omit repeated titles, verdict labels, SHA/port/ID headers, environment boilerplate, and round-history explanations. Put shared setup and revision details once in the round report; keep commands, traces, raw output, and detailed reasoning in separate evidence attachments. Briefly disclose a limitation in the check when it changes the verdict; concision must not hide missing verification.

Example: “转派后，新 Agent 收到原对话上下文并创建了独立话题。刷新后消息仍保留。” For a failure, name the unmet outcome directly, without recounting the debugging process. Keep required evidence complete; shorten its presentation, not the work.

Final handoff (mandatory)

Cloud browser links use https://lobehub.com. For all acceptance, round, cleanup, and upgrade URLs in this skill (including instructions below that say "verbatim"), normalize LobeHub Cloud origins to https://lobehub.com, preserving the path, query, and fragment. Cloud hosts are lobehub.com and its subdomains. Keep self-hosted and development origins unchanged. This changes display links, not the CLI's configured API server. Keep the Skill installation resource at https://app.lobehub.com/acceptance/skill.md.

Close every browser session this run opened (agent-browser --session <name> close, web teardown) before handing off; a session left open keeps a full browser running indefinitely. Stop this run's resource guard (resource-guard.sh stop --state-dir <run state dir>) as well, and state in the round report whether it reached yellow or red and what that stopped; a run that hit red must say which checks it left blocked instead of passing.

Before declaring the task done, prove coverage: for each check with requiredEvidence, every declared type is present at least once. Report it explicitly; a missing type holds the delivery at uncertain no matter how good the work is.

If publication returns a storage-quota failure, the publishing owner reads publication-recovery.md once. Preserve local evidence, disclose incomplete uploaded coverage, and give the returned scope-specific cleanup and upgrade options. Never delete user data or blindly retry the whole ingest.

The final response for a completed handoff MUST include the published acceptance URL together with the coverage result — never only a check-result id or a prose claim. Obtain the links from the path you actually executed:

• Authored round: copy acceptanceUrl returned by lh acceptance run ingest --json verbatim.
• Operation-plan round: copy acceptanceUrl returned by lh acceptance run ingest --operation <id> --json. When the plan needed no ingest, or the CLI predates plan-round ingest, follow the read-only plan handoff lookup instead. Do not ingest without --operation, create another acceptance, or resubmit evidence merely to obtain a link.

Never guess a host, acceptance ID, or round index. The documented plan lookup is the only reconstruction needed for CLIs whose submission output contains only an internal run URL. If the run has no acceptance association or the lookup fails, report the handoff as blocked and preserve the submitted evidence; do not declare delivery complete or fabricate a link. Put no images, local paths, local file links, or internal run-page paths in the chat reply.

Write the link as a plain-text line, never inside a fenced or inline code block — the chat client only linkifies plain text, and a code block makes it unclickable. Hand off only the acceptance URL: the acceptance page opens on its latest round, so a separate per-round link adds nothing for the reader. Replace the placeholder below with the URL from the selected path:

Acceptance: <acceptanceUrl, verbatim> Coverage: 2/2 criteria, all required evidence uploaded

Portability rules

• Engine-level capture over OS capture. agent-browser screenshot / dom / eval run headless; screencapture / osascript are macOS-only. iOS: xcrun simctl io over host-window capture. Rounds land under .acceptances/; allocate them with new_round (report.md), which seeds the .acceptances/.gitignore that keeps them out of git.
• Save as you go. Write each artifact into the round's assets/ and its case into result.json the moment it is captured, so a crash near the end loses nothing; then publish the round once.
• Don't invent evidence. Capture only the types a check declares.
• Chapter every recording. Log a mark for each step and each claim you verify on a frame, and disclose every anomaly you noticed as a flag; the reviewer seeks to them instead of watching the whole clip — video-chapters.md.

Reference map

Only the assigned owner reads the relevant reference or section, once per version; other roles use its acquired facts and instructions. The runner does not load the checker contract, and the primary does not re-open their documents or evidence.

Need	Reference
Execution delegation, models, context ownership	delegation.md
First-round plan and evidence review	acceptance-checker.md
Optional flow authoring or regrouping	flows.md
Storage-quota publication recovery	publication-recovery.md
Bounding a run's memory use	resource-guard.md
The project layer, bootstrapping an adapter	project-adapter.md
Mistakes checklist (acquire once per version)	common-mistakes.md
Forcing state, error injection, runtime probes	probe-mock-patterns.md
Authored rounds, result.json, ingest	report.md
Plan-driven rounds: plan, ingest, coverage	plan-format.md
Evidence media, provenance, submission, safety	evidence.md
Interaction cost overlay	interaction-cost.md
Web/Electron Chromium CLI commands	agent-browser.md
iOS Simulator driver CLI commands	sim-use.md (preferred), axe.md (fallback)
Bundled CDP screenshot and macOS capture preflight	screenshot-helpers.md
Authenticated Web session	auth-web.md
Native macOS / OS-owned step	computer-use.md
Video chapters: steps, checks, flags on a clip	video-chapters.md
Temporal evidence: Web/Electron, iOS, native	recording-cdp.md, recording-ios-simulator.md, recording-native-macos.md