qa-test-plan

QA Test Plan

Owns: QA plan construction. It does not execute tests, change code, resolve gates, decide completion/release status, or produce a Verification Record.

Use the taxonomy and identifiers from the canonical Verification Record. The caller records observations and verification-audit performs the sole semantic reduction.

Inputs

Accept one acquisition route:

• Runner handoff: normalized Ticket Envelope, acceptance criteria, frozen diff, CandidateRef, supplied invariants/boundaries, environments, gates, and observed checks.
• Standalone acquisition: acquire a PR, commit, local diff, or user-requested scope read-only; record the observed identity, requested behavior, repository rules, available environments, limitations, and checks.

Do not parse Markdown to manufacture a Ticket Envelope. Mark unknown inputs as draft limits or gates; never assume access or successful execution.

Volatile intake bound

• max_volatile_bytes: 103998 normalized UTF-8 bytes per invocation. This is the observed 96,393-byte candidate-diff high-water mark plus the 2,380-byte ticket-body and 5,225-byte review-handoff maxima. The corpus is the run's TK-01/TK-02/TK-05/TK-07/TK-08 normalized git diff --no-ext-diff --no-color observations, its nine ticket bodies, and compact leaf results.
• max_single_output_bytes: 32596, the observed TK-02 executable-code candidate diff.

Count every diff, raw file slice, pasted handoff, evidence body, and tool result after CRLF or lone-CR normalization to LF. Start from the changed-file manifest; truncate command output before it enters context and slice larger material by file or hunk. Prefer path plus SHA-256 references over pasted artifacts and load content only for a causal gap the plan must cover. If the next required read would exceed a cap, return a schema-3 partial plan with exact inspected/remaining scope and budget-exhausted. Never remove a causal case, evidence classification, invariant, boundary, or gate merely to stay within the bound.

Build the plan

First apply verification admission and cost per attempt. Require current identity and delivered packaging/artifact-graph results before admitting expensive execution. Inventory all prior attempts, including failed or superseded shards; keep original evidence identities and unknown costs. Select causal checks without waiving mandatory profiles or exact-head CI. A repeated full suite needs a specific reason and the current attempt's remaining budget, not just a new CandidateRef.

For each changed behavior:

1. State the causal chain from injection point to user/external observation.
2. Select the smallest test that crosses the changed mechanism.
3. Classify the intended observation as static, unit, integration, simulated, or live using the canonical reference.
4. Name exact setup, action, expected result, cleanup, and evidence to capture.
5. Map the case to ticket criteria, invariants, boundary items, and gate IDs.
6. Add negative/error paths, retries, ordering/idempotency checks, and regression coverage where relevant.

Do not label a mocked provider or fake browser as live. If a required environment, credential, approval, or device is unavailable, write a specific open gate and an authorized simulation plan; do not silently lower the requirement.

Output

For a runner handoff, return the caller's schema-3 qa-plan leaf result. Bind it to the exact CandidateRef and canonical phase contract. Its schema-1 quality payload names causal scope, content-addressed planned-evidence references, and limitations; a timeout or interruption returns complete: false, the last durable phase, and exact remaining phases. Do not discard a partial plan or mark it passing. Include normalized schema-3 execution with observed isolation; never infer delegation, separate context, parallelism, or authority from the requested plan.

For standalone acquisition, return:

# QA Plan

## Candidate
- CandidateRef:

## Admission and Cost
- Current identity, delivered packaging/artifact-graph results and evidence references.
- Prior attempts/shards with original identities, outcomes, durations and artifact hashes.
- Summed invocation time versus measured end-to-end elapsed time; unknowns and lower bounds.
- Selected/omitted checks with causal or mandatory-policy reason, full-repeat justification,
  estimated next cost, the current attempt's remaining budget and unresolved execution state.

## Automated Checks
- ID, command, layer, causal path, expected evidence.

## Manual / Environment Checks
- ID, setup, action, expected observation, evidence class, cleanup.

## Negative and Regression Paths
- ID, failure or preserved behavior, expected observation.

## Mapping
- QA ID -> acceptance criterion / invariant / boundary item / gate.

## Open Gates and Limits
- Gate ID, owner, unblock condition, and claim impact.

Planning a check is not evidence that it ran. Keep planned, executed, passed, failed, skipped, and blocked states distinct. A standalone draft cannot claim ticket completion or release.