Prepare Acceptance

Prepare Acceptance

Prepare an implemented change for a separate human acceptance session. Exercise it as a human user or real client would, report clear defects, wait for CI once no defects remain, and produce concise evidence for the code currently checked out. Fixes and automated validation belong to separate caller-managed steps when used. PR and commit identity establish final CI and handoff alignment; they are not prerequisites for exercising the product.

Required inputs

Resolve these from the caller before acting:

• approved requirements, scenarios, design, test plan when one exists, and task artifacts;
• caller-supplied implementation summary or completion evidence identifying delivered behavior;
• evidence output directory;
• unresolved-assumptions ledger path, when one exists;
• verification scope:
  • full;
  • targeted, naming the affected flows, directly dependent flows, a concise impact rationale, and prior full-pass evidence to retain, plus the caller's attestation that this scope covers the tracked changes since that baseline; or
  • evidence-only, identifying prior flow evidence whose coverage revision exactly matches the current tracked contents and supplying the caller's attestation of that content match.

If source-of-truth artifacts or the evidence directory are missing, report what is missing and stop before flow execution. Do not infer product behavior from implementation alone. When an approved test plan exists, treat its required and activated conditional AT-* flows, evidence requirements, authorized effects, and permitted substitutes as authoritative. Use full when the caller does not supply a scope or no trustworthy full-pass baseline exists. Targeted and evidence-only reuse depend on the caller's impact or content attestation: reconcile it with the approved flow inventory, but do not independently inspect a diff. Use evidence-only only when the caller attests that no tracked product contents changed after the recorded coverage revision; PR alignment, pushing the already-tested commit, waiting for CI, or a Validator run that made no tracked change do not invalidate flow evidence. When no assumptions-ledger path is supplied, use <evidence-directory>/acceptance-assumptions.md and create it if needed. Keep this preparation autonomous: preserve product, scope, or design ambiguity for the later human acceptance session instead of asking the user here.

Procedure

1. Establish the test context

1. Read repository instructions, the approved artifacts, the unresolved-assumptions ledger, and the existing automated-validation evidence supplied by the caller.
2. Treat the code currently present in the checked-out worktree as the test target. Record local HEAD and tracked worktree status as context, but do not require a clean worktree or PR-head equality before exercising flows. Do not inspect or reconstruct a diff.
3. Use the project's normal setup, build, install, seed, and start commands when needed to expose the current worktree through its real public surface. These are test setup, not automated validation. Do not require an executable to embed a Git SHA, compare executable and commit timestamps, or infer source identity from a semantic version string.

2. Exercise the product like a user

When an approved test plan exists, use its required and activated conditional AT-* obligations as the executable flow inventory, regardless of how the plan structures them. Identify applicable HT-* obligations separately and carry them forward only as human-review instructions, never as flows for the agent to execute or support with passing evidence. Otherwise derive a concise list of representative user or client flows from the approved artifacts and the caller-supplied implementation evidence. Do not inspect a diff to discover behavior. A flow is a meaningful journey or operation through a public surface, not an individual requirement scenario, input permutation, internal branch, or automated-test case. Group equivalent variants within a flow and choose typical data, configuration, and roles that demonstrate the delivered behavior.

• For full scope, exercise every required or activated conditional AT-* flow from the approved test plan. When no test plan exists, exercise the representative flows that collectively cover the delivered user or client journeys and primary public surfaces. Do not turn every specification scenario into a separate manual case.
• For targeted scope, exercise only the affected flows and directly dependent flows named by the caller. Require an existing <evidence-directory>/acceptance-flow-evidence.md from a prior full pass, plus the caller's impact attestation and rationale bounding why flows outside the scope are unaffected; it may group them by surface or subsystem rather than enumerate them individually. Reconcile that scope with the approved flow inventory without inspecting a diff. Do not expand a targeted pass into a full pass merely because the revision changed. If the scope omits an obvious dependency or cannot be reconciled with the approved flow list, report the scope gap instead of silently guessing or rerunning everything.
• For evidence-only scope, do not exercise flows or replace screenshots. Require acceptance-flow-evidence.md to cover the representative flow inventory and identify a coverage revision, plus the caller's attestation that current tracked contents match that revision. Do not independently inspect a diff to establish the match. Continue directly to current-head CI and handoff collation. If the attestation is absent, reports changed contents, or coverage is incomplete, reject evidence-only reuse and require a full or targeted scope from the caller.

Exercise the selected flows through the same public surface a real user or client would use.

Do not omit a required or activated conditional AT-* flow because it is slow, expensive, needs credentials, or causes external effects when the approved test plan explicitly authorizes those conditions. Use a mock, dry run, sandbox, or other substitute only when that flow explicitly permits it. If its prerequisites are unavailable or it cannot be exercised safely within its approved conditions, record the impediment and do not claim the full pass or final acceptance evidence is complete.

• For a web, mobile, desktop, or terminal UI, navigate and operate it through its real interface.
• For an API, call it through its documented HTTP, SDK, or CLI surface as a client would. Record a sanitized request shape, response status and relevant response excerpt, plus observable side effects.
• For a CLI or library, invoke its public commands or API as a consumer would and retain representative input, output, and resulting state.
• For background behavior, trigger it through the normal entry point and inspect the user-visible or externally observable result.

Use typical representative data, configuration, and roles. Run only the setup and build commands needed to expose the checked-out product for hands-on use. Do not run automated unit, integration, or end-to-end suites, linters, or other automated validation commands; those belong to the caller. Do not fuzz, try bizarre inputs, attempt to break the product, or build an exhaustive edge-case matrix. Test an error, empty, boundary, or transient state only when it is an approved flow or necessary to complete a normal flow.

For each selected flow, record the action, observed outcome, concise evidence, and any limitation. Verify the resulting state instead of relying on task completion or agent assertions.

When one or more clear implementation defects appear, do not fix them. Continue through every remaining in-scope flow that can still be exercised safely and independently; one finding alone is not a reason to end the scoped pass. Stop early only when a defect makes the remaining in-scope flows unreachable, unsafe, or incapable of producing trustworthy evidence. Record every unexercised in-scope flow and the defect that blocked it.

After every pass, write or update <evidence-directory>/acceptance-flow-evidence.md. Record the representative executable flow inventory, including each applicable AT-* identifier when the test plan supplies one, the current verification scope and rationale, and for each flow its most recent tested SHA, outcome, evidence paths, and limitations. List applicable HT-* obligations separately as human-review instructions without execution outcomes. A full pass establishes the baseline. A targeted pass replaces evidence for its selected flows and retains prior evidence for flows the caller identified as unaffected, preserving the original SHA and provenance rather than relabeling it as current-head evidence. Label retained evidence as caller-scoped provenance, not as an independently reverified current-head observation. Record the revision whose tracked contents this combined coverage supports according to the caller's impact or content attestation.

Write <evidence-directory>/acceptance-findings.md with the tested SHA and all clear defects found, including each affected flow, expected and observed behavior, concise reproduction steps, and evidence paths. On a targeted pass, mark a prior finding resolved only when its affected flow is in scope and the observed behavior now passes; retain other prior findings with their original status and SHA. Do not wait for CI or write final acceptance evidence while a current finding remains. Report the aggregated findings clearly so the caller can run its fix and validation process.

Append product, scope, or design ambiguity to the unresolved-assumptions ledger with the decision needed and likely impact; never silently choose an interpretation or report ambiguity as a clear defect. After a fix, require the caller to provide a new impact scope. Use a full pass only when the change is broad, cross-cutting, or its impact cannot be bounded confidently.

3. Capture visual evidence

For any UI—including web, mobile, desktop, and TUI—store screenshots under <evidence-directory>/acceptance-screenshots/ as the flows are exercised. Capture at least one meaningful stable result for every in-scope UI flow, and more only when multiple states are needed to understand that flow. Retain prior screenshots for out-of-scope flows with their original SHA and provenance. Do not capture loading, empty, error, responsive, or before/after states unless they are part of the flow being tested.

Record for every screenshot:

• flow demonstrated;
• expected behavior and what the reviewer should notice;
• route or command, representative input, and user role when relevant;
• tested head SHA;
• concise text equivalent.

Screenshots support the observed interaction; they are not proof on their own. For non-visual surfaces, retain the corresponding client request/output evidence instead.

4. Wait for current-head CI

Before invoking codagent:wait-ci, require a clean tracked worktree, a committed and pushed local HEAD, and a matching PR head. If any publication or alignment condition is absent, stop and tell the caller to commit, push, or align the PR as needed; do not perform those actions in this skill.

Invoke codagent:wait-ci only after that alignment and after the required full or targeted pass finds no clear defects, or evidence-only reuse satisfies its caller attestation. After CI returns, immediately re-read local HEAD and the PR headRefOid. Require the returned head_sha, current local HEAD, and current PR head to be identical; otherwise report that final handoff evidence cannot yet be produced. If the caller reports tracked product-content changes after flow testing, require a new caller-supplied impact scope and run that verification before producing final evidence. A push that publishes already-tested contents, PR alignment, or a no-change Validator run does not invalidate the pass.

• passed: continue.
• failed or comments: do not write final acceptance-test or handoff evidence or claim readiness. When the failure or comment is clearly attributable to the implementation and actionable in the repository, add it to <evidence-directory>/acceptance-findings.md with the tested SHA, failing check or comment, relevant log or link, and concise expected-versus-observed behavior. Report it as a clear defect so the caller can run its fix and validation process. Otherwise, report the external, environmental, or infrastructure blocker without classifying it as an implementation defect.
• pending: do not claim readiness. Report the pending checks and stop for the caller to resume or retry later.
• no configured checks: continue only after recording that CI coverage is absent.

Treat skipped, neutral, cancelled, stale, and timed-out checks as explicit evidence states, never as proof that associated behavior passed.

5. Write acceptance-test evidence

Write acceptance-test.md in the evidence directory only after CI passes or is explicitly absent, the required verification makes no tracked changes, acceptance-flow-evidence.md covers the current tracked contents under the caller's supplied scope or content attestation, every required or activated conditional AT-* flow or derived representative flow has passing evidence from either the full baseline or a subsequent targeted pass, and any automated-validation evidence supplied by the caller applies to the current pushed head. Do not require automated-validation evidence when none was supplied; explicitly record its absence instead. Include:

• exact current local/PR head SHA;
• the representative flows, their outcomes, most recent tested SHA, and whether evidence came from the full baseline or a targeted post-fix pass; label retained observations as caller-scoped provenance rather than independently reverified current-head observations;
• the caller-supplied impact scope and rationale for retaining unaffected-flow evidence;
• clear-defect status and any prior fix evidence supplied by the caller;
• existing automated-validation and CI status without rerunning either, explicitly noting when automated-validation evidence was not supplied;
• screenshot metadata and text equivalents for every UI flow;
• representative API, CLI, library, or background-operation evidence for non-UI flows;
• unavailable flows, environment limitations, warnings, and residual risks;
• new unresolved assumptions added to the canonical ledger.

List any applicable approved HT-* obligations separately as human-review instructions. Do not perform or claim human-only testing.

Prefer concise conclusions with durable log paths over raw log dumps.

6. Prepare the human handoff

Collate existing implementation, acceptance-test, available automated-validation, and CI evidence. Do not run new acceptance tests, capture replacement screenshots, or fix issues during collation. If required evidence is missing or stale, stop and require the producing phase to run again. The explicit absence of caller-supplied automated-validation evidence is not missing evidence.

Write acceptance-handoff.md in the evidence directory with:

1. Decision brief — unresolved decisions, delivered behavior, overall validation status, known limitations, and suggested human review path.
2. Revision identity — repository, PR URL, current head SHA, generation time, tracked worktree status, full-baseline SHA, and subsequent targeted-pass SHAs.
3. Flow evidence — each supported flow, what was done, the observed result, its evidence, most recent tested SHA, baseline or targeted provenance, and any limitation; label retained observations as caller-scoped provenance rather than independently reverified current-head observations.
4. Automated validation and CI — concise status plus durable logs or links produced by separate validation and CI steps; explicitly say when automated-validation evidence was not supplied.
5. Visual and client evidence — screenshot metadata/text equivalents, API or CLI evidence, and practical human review instructions.
6. Residual risk — unavailable flows, environment limitations, warnings, and relevant omissions.
7. Assumption handoff — canonical ledger path plus a concise summary of every unresolved item.

7. Verify readiness

Before reporting readiness:

1. Require a clean tracked worktree.
2. Require local HEAD, PR head, acceptance-test.md, and acceptance-handoff.md to name the same current revision. Require acceptance-flow-evidence.md to account for the representative flow inventory under the caller's supplied scope or content attestation and preserve the actual SHA of every retained or refreshed observation.
3. Require CI evidence for that SHA to be passing, or explicitly record that no checks exist.
4. Verify every required or activated conditional AT-* flow is accounted for, every referenced evidence and screenshot path exists, and every UI flow has visual evidence from the baseline or an applicable targeted pass. Verify applicable HT-* obligations are listed only as human-review instructions.
5. Ensure persisted evidence does not expose secrets, credentials, tokens, private data, or unrelated user content.
6. Record the current PR state for the caller without changing it.

Use revision identity to show that final CI and handoff evidence describe the code under review. Preserved observations remain tied to their original revisions and claim only caller-scoped applicability to the current head. Do not require or inspect embedded executable revision metadata.

Report the exact ready-for-acceptance SHA, PR URL, handoff path, unresolved-decision count, CI status, and known limitations. Do not mark the PR ready or perform human acceptance. The caller owns any machine-readable status file, terminal marker, or control-flow protocol.

Out of scope

• Do not obtain human acceptance; leave that to the later human acceptance session.
• Do not fix implementation defects, modify tracked source or approved artifacts, commit, push, or alter PR metadata. Disposable or ignored setup and build outputs are allowed.
• Do not run automated test suites, linters, other automated checks, or automated validation; the caller owns validation and reinvocation. Setup and builds needed to use the product are allowed.
• Do not change approved requirements, product scope, or design to make tests pass.
• Do not archive the change, mark the PR ready, merge, or release.
• Do not perform fuzzing, adversarial testing, or exhaustive edge-case exploration.
• Do not treat successful task execution, agent summaries, or screenshots alone as proof of behavior.