Lisa Verify Prd

PRD-level Verification: $ARGUMENTS

/lisa:verify-prd <prd> is the initiative-level acceptance gate. It runs after a PRD is shipped (all generated top-level work terminal, per prd-lifecycle-rollup's rollup phase) and proves the shipped product actually matches the PRD — not merely that every ticket is closed. shipped is necessary but not sufficient for verified: a PRD can have every child ticket closed while still missing a requirement, diverging from its acceptance criteria, or failing a real user workflow.

This is distinct from /lisa:verify, which empirically verifies a single work item (a ticket / story / sub-task) in its target environment as part of that item's Build/Fix/Improve flow. /lisa:verify drives a build ticket to done at the leaf/build level; it does not read the PRD or judge initiative-level acceptance. /lisa:verify-prd operates one level up, over the whole initiative. See the prd-lifecycle-rollup rule's "PRD-level verification vs ticket verification" section for the full distinction.

$ARGUMENTS is a single PRD reference: a GitHub issue URL (or <org>/<repo>#<n>), a Linear project/issue URL, a Notion page URL, a Confluence page URL, or a JIRA issue key/URL.

Confirmation policy

Do not re-prompt once invoked. Like the *-prd-intake skills, the caller has already authorized the run by invoking the skill; asking the user to confirm before reading or before applying the guard defeats the purpose of a batchable acceptance gate. Run the front-half (resolve → read child set → guard) to completion and report.

Scope of this skill

This skill covers the read/guard front-half, both verdict branches (PASS and FAIL), and the idempotency of PRD-level verification:

1. Resolve the PRD ref and detect its source vendor.
2. Read the PRD body and its generated top-level child work set via the prd-lifecycle-rollup contract.
3. Apply the per-vendor terminal predicate to the generated top-level work and run the terminal-child guard: if any required top-level child is non-terminal, report the incomplete set and STOP — do not run verification, do not transition the PRD.
4. Spec conformance — when the guard passes, invoke the spec-conformance skill with the PRD as the spec source to build a section-by-section coverage matrix against the shipped product (never reimplementing the matrix).
5. Empirical verification — invoke verification-lifecycle to run empirical checks appropriate to the PRD's surface (browser/computer-use, API, CLI, DB, screenshots, logs), honoring the verification rule. Quality gates (test/typecheck/lint) are not verification.
6. PASS transition + evidence — when spec conformance is CONFORMS and every applicable empirical check passes, transition the PRD lifecycle from the resolved shipped role to the resolved verified role (vendor-neutral via config-resolution) and post verification evidence (the coverage matrix + empirical proof artifacts) back on the PRD.
7. FAIL — re-open as ticketed + build-ready fix tickets (self-healing, never blocked) — when spec conformance is PARTIAL/DIVERGES, or any applicable empirical check fails (or a required surface is unavailable), move the PRD from the resolved shipped role back to the resolved ticketed role, create build-ready fix tickets via tracker-write (build_ready: true) for each missing/incorrect/divergent behavior — added to the PRD's generated top-level work — and post a product-readable failure report (with the verification-round count). PRD verification never moves the PRD to blocked. The fix tickets auto-build, rollup (*-prd-intake Phase 3f) re-ships the PRD once they are terminal, and a later intake cycle (Phase 3g) re-verifies — the loop closes itself.
8. Idempotency — every write in Phases 6 and 7 is safe to re-run: evidence/failure-report comments carry a stable sentinel marker and are regenerated in place (never appended), fix tickets are deduped by a stable PRD-ref + requirement marker (referenced/updated, never duplicated), and the lifecycle transition is a no-op when the PRD already carries the target role (exactly one lifecycle label/status remains). See Phase 8 — Idempotency for the per-write guards.

When verification passes, this skill runs the PASS branch of Phase 6 (shipped → verified). When it does not pass — spec conformance not CONFORMS, or any empirical check failing — this skill runs the FAIL branch of Phase 7 (shipped → ticketed + build-ready fix tickets + failure report); it does not leave the PRD at shipped and never uses blocked (PRD verification is self-healing — the fix tickets auto-build and the PRD re-ships and re-verifies). Re-running the skill against the same PRD produces no duplicate evidence comments, no duplicate fix tickets, and no duplicate lifecycle labels/statuses — the Phase 8 guards make each Phase 6/7 write idempotent, exactly as prd-backlink regenerates its ## Tickets section in place and github-prd-intake no-ops a rollup on an already-shipped PRD.

Phase 1 — Resolve the PRD ref and detect the source vendor

Detect the vendor from $ARGUMENTS the same way prd-ticket-coverage / prd-backlink do — from the host (or key shape for JIRA), never by guessing:

State the operation you need, never a vendor tool name: the access skills own substrate selection, and a hardcoded MCP tool name is not portable. The same server registers under different prefixes depending on install path (mcp__atlassian__*, mcp__claude_ai_Atlassian__*, mcp__plugin_atlassian_atlassian__*), and its data tools register only once OAuth completes — so a named tool fails outright in a context that has the CLI or token substrate available and would otherwise have worked. See the integration-access-layer rule.

Read the PRD body via the vendor-appropriate surface. The vendor that owns the PRD source is what Phase 2 reads the child set from; it is independent of which tracker hosts the generated tickets (a Notion PRD can own JIRA tickets — the cross-vendor case is handled by the documented generated-work section in Phase 2).

Where the PRD lives in the same vendor as the configured tracker (.lisa.config.json), prefer reading the PRD ticket through lisa-tracker-read so this skill stays agnostic of which tracker hosts the work — tracker-read dispatches to lisa-github-read-issue / lisa-jira-read-ticket / lisa-linear-read-issue per config and returns the consolidated context bundle (PRD body, native children, links). For PRD sources with no tracker counterpart (Notion / Confluence / cross-vendor), read the PRD body directly via the vendor surface above.

Phase 2 — Read the generated top-level child work set

Do NOT reimplement child enumeration. Consume the PRD→generated-top-level-work relationship recorded by the merged child-linking work (prd-backlink native linking + its always-written machine-readable generated-work section), exactly as github-prd-intake's rollup phase consumes it. Read the generated top-level work only — the PRD's created Epics and any top-level Story created directly under it — and exclude leaf Sub-tasks and any Story nested under a generated Epic, per the prd-lifecycle-rollup rule's generated-top-level-work contract.

Use two sources, native first, deduped by child-ref identity:

1. Native hierarchy (primary). Read the PRD's direct native children — these are its top-level children:

   • GitHub — the PRD issue's direct subIssues nodes (the GraphQL subIssues query lisa-github-read-issue Phase 3 uses; capture number title state url stateReason labels).
   • Linear — for a PRD Project, its member Issues (list_issues({project})); for a PRD Issue, its sub-Issues (get_issue children). Capture each child's state (workflow state + category).
   • JIRA — the PRD's children via the epic-link/parent JQL lisa-jira-read-ticket Phase 5 uses ("Epic Link" = <PRD-KEY> or parent = <PRD-KEY>). Capture each child's statusCategory.
   • Notion / Confluence — no native issue hierarchy; use the documented section below.

2. Documented generated-work section (fallback / cross-vendor). When native hierarchy is unavailable (older host, cross-vendor PRD→tracker, or the relationship was only recorded in the PRD body), parse the machine-readable generated-work section prd-backlink writes to the PRD body (## Tickets, alias ## Generated Work). Enumerate the <!-- lisa:gw ref=… url=… type=… parent=… --> tokens; the generated top-level child set is every token whose parent is empty (top-level), exactly as prd-backlink documents. Tokens with a non-empty parent are descendants — skip them. For GitHub, this is the same awk extraction github-prd-intake Phase 3f.2 uses.

Dedupe by child-ref identity (owner/repo#number for GitHub, the issue/project identifier for Linear, the issue key for JIRA, the recorded ref for Notion/Confluence — the prd-lifecycle-rollup idempotency dedupe key) so a child appearing in both the native graph and the documented section is counted once. Match by stable ref, never by title.

If neither source yields any generated top-level child (the PRD generated nothing, or the relationship was never recorded), report no generated top-level children — cannot verify an empty PRD and STOP. Do not transition the PRD.

Phase 3 — Terminal-child guard

Apply the per-vendor terminal-state predicate from the prd-lifecycle-rollup rule to every generated top-level work item (cite the rule by slug — do not restate its predicate table here). In summary, a top-level child is:

• Terminal — it has reached its source/tracker's done/shipped state (GitHub: closed + the resolved build done role label where used; Linear: a done-category completed state; JIRA: statusCategory.key == "done"; Notion/Confluence: the documented generated-work entry marked done). A generated Epic is terminal only when it has rolled up to its own terminal state per leaf-only-lifecycle — read the child's own resolved state; do not re-derive it from its leaves.
• Terminal-but-dropped — closed-as-not-planned (GitHub stateReason == "not_planned") / canceled or duplicate (Linear — both are terminal state.types; duplicate is distinct from canceled) / won't-do. It does not hold the PRD open and is excluded from the required set.
• Incomplete / blocked — anything else (still open, or closed without the done role). It holds the PRD open.

The required set is the top-level children minus the terminal-but-dropped ones. Branch:

Any required top-level child is non-terminal (the "Given a PRD has generated child work that is not terminal" scenario):

1. STOP. Do not run empirical verification or spec-conformance.
2. Leave the PRD lifecycle untouched — it stays at shipped. Do not transition it to verified or blocked; do not close or archive it.
3. Report the incomplete child set — list each non-terminal required top-level child as - <ref> "<title>" — <state>, so product can see exactly what is blocking PRD-level verification.

This guard exists because PRD-level acceptance is only meaningful once the work graph is actually complete. shipped is the precondition; verifying a PRD whose generated work is still in flight would produce a false PASS or FAIL against an incomplete product.

All required top-level children are terminal (at least one required child exists): the guard passes. Proceed to Phase 4 — Spec conformance and the rest of the PASS path below. The guard is the precondition for verification; only once the work graph is complete is it meaningful to check the shipped product against the PRD.

Phase 4 — Spec conformance against the PRD

The guard proves the work graph is complete; spec conformance proves the shipped product matches what the PRD asked for, section by section. This is the "accountant lens" — did the initiative ship exactly the PRD's requirements, nothing silently dropped, nothing scope-crept?

Do NOT reimplement the coverage matrix. Invoke the existing spec-conformance skill with the PRD as the spec source:

/spec-conformance <PRD ref/URL>

Pass the same $ARGUMENTS PRD reference resolved in Phase 1. spec-conformance Phase 1 already accepts a GitHub issue URL / Linear identifier / JIRA key / Notion / Confluence PRD as its spec source and loads the full PRD body (via tracker-read or the vendor surface), so the PRD is the spec here — not a plan file or a leaf ticket. It then:

• extracts every PRD requirement into a structured list (acceptance criteria, Out of Scope, technical commitments, Validation Journey assertions, deliverables);
• inspects the shipped product (the merged work across the generated top-level children) for evidence of each;
• builds a section-by-section coverage matrix mapping each requirement to evidence;
• flags scope creep (Out-of-Scope violations) and untraceable changes separately from misses;
• returns a verdict: CONFORMS, PARTIAL, or DIVERGES.

Capture the coverage matrix and verdict verbatim — both feed the evidence comment in Phase 6 and gate the PASS branch.

Spec conformance at the PRD level differs from the leaf-ticket conformance run during a single item's Build/Fix/Improve flow: the spec is the PRD itself (the whole initiative's requirements), and the shipped work is the union of all generated top-level children, not one branch's diff. The spec-conformance skill handles both because its spec source is parameterized; this skill simply hands it the PRD.

Branch on the verdict:

• CONFORMS — continue to Phase 5 (empirical verification).
• PARTIAL or DIVERGES — verification has not passed. Do not run Phase 5 (empirical verification adds nothing once the spec already diverges) and do not enter the PASS path. Record the verdict and the matrix, then go to Phase 7 — FAIL with a CONFORMANCE_FAILED cause: the failed/missing/scope-crept requirements the matrix flagged become the failure report's findings and the seeds for the linked fix issues.

Phase 5 — Empirical verification of the shipped surface

Spec conformance reads the diff and the requirements; empirical verification runs the actual shipped system and observes results. Quality gates (tests, typecheck, lint) are prerequisites, not verification — a green test suite never substitutes for exercising the shipped product (see the verification rule).

The verification surface is PRD-dependent. A PRD that shipped a UI flow is verified through the browser; one that shipped an API through request/response captures; a CLI through command runs; a schema change through database queries; a background job through logs and queue inspection. Do not assume a fixed surface — classify it from what the PRD actually delivered.

Do NOT reinvent the verification machinery. Invoke the verification-lifecycle skill and follow its mandatory sequence (confirm quality gates → classify empirical types → check tooling → fail-fast → plan → execute → codify → loop) against the shipped initiative:

/verification-lifecycle

1. Classify the empirical verification types that apply to the PRD's shipped surface, per the Verification Types table in the verification rule.
2. Discover tooling for each type (browser/computer-use MCP, HTTP client, DB client, CLI, log/metrics access) via the lifecycle's Tool Discovery Process. For a UI surface, reuse the product-walkthrough skill to drive the live product through a real browser and ground verification (and the eventual evidence comment) in what actually renders.
3. Plan each check: the exact tool/command, the expected pass outcome, and any prerequisites (running service, seeded data, auth). A plan that lists only test/typecheck/lint is not a verification plan.
4. Execute each check and collect proof artifacts (screenshots, request/response captures, query outputs, log excerpts with correlation IDs) per the lifecycle's Proof Artifacts Requirements.
5. Codify — for each passing empirical verification, the lifecycle invokes codify-verification to encode it as a regression test (Playwright for UI, integration test for API/DB, benchmark for performance) so the PRD's behavior cannot silently regress. Honor that step; it is mandatory for every empirical type except the inherently non-behavioral set the lifecycle exempts.

If a required surface or its tooling is unavailable, follow the lifecycle's Escalation Protocol — declare the verification level (PARTIALLY VERIFIED / UNVERIFIED) and surface the gap rather than declaring a false PASS.

Branch on the result:

• Every applicable empirical check passes (and is codified where the lifecycle requires) — continue to Phase 6 (PASS).
• Any applicable empirical check fails, or a required surface is unavailable — verification has not passed. Record what failed/was-blocked (the check, the tool/command, observed vs expected, and any proof artifacts captured), then go to Phase 7 — FAIL with an EMPIRICAL_FAILED cause: each failed check (or unavailable required surface) becomes a failure-report finding and a seed for a linked fix issue.

Single-environment note. In a single-environment project (main/production only, no dev/staging), the shipped surface is whatever production exposes. A project with no deployed application, sign-in, or end-to-end environment variables (Lisa itself) verifies on its CLI/dry-run surface — running the documentation/skill build and drift check — which is the empirical surface the PRD's Validation Journey declares. The surface is always PRD-dependent: read the PRD's Empirical Verification Plan and verify what it says ships.

Phase 6 — PASS: transition shipped → verified and post evidence

Reach this phase only when both are true:

• Phase 4 spec conformance returned CONFORMS, and
• Phase 5 every applicable empirical check passed (and was codified where required).

If either is false, do not enter this phase — record the verdict and route to Phase 7 — FAIL, which owns the shipped → ticketed (re-open + build-ready fix tickets) path.

6.1 — Resolve the verified and shipped roles

Resolve the PRD-lifecycle roles from .lisa.config.json (then .lisa.config.local.json override) per the config-resolution rule — the same role-resolution the *-prd-intake skills use. Cite config-resolution for the role vocabulary; do not hardcode label strings except as the documented defaults. Resolution per vendor:

For GitHub, resolve with the same helper github-prd-intake uses:

read_role() {  # role default → resolved value (local override wins)
  local role="$1" default="$2" local_v global_v
  local_v=$(jq -r ".github.labels.prd.${role} // empty" .lisa.config.local.json 2>/dev/null)
  global_v=$(jq -r ".github.labels.prd.${role} // empty" .lisa.config.json 2>/dev/null)
  echo "${local_v:-${global_v:-$default}}"
}
SHIPPED=$(read_role shipped  prd-shipped)
VERIFIED=$(read_role verified prd-verified)

6.2 — Transition the PRD shipped → verified

Idempotency guard (no-op if already verified). Before transitioning, read the PRD's current lifecycle role. If the PRD already carries $VERIFIED, the transition is a no-op — do not re-add the label/status, do not re-remove $SHIPPED (already gone). Record it as already verified (no-op) and proceed to 6.3 (the evidence comment is still refreshed in place, per Phase 8). This mirrors github-prd-intake Phase 3f.1's "no-op if already shipped" guard and the prd-lifecycle-rollup "rollup is keyed by the PRD's current state" rule (cite both by slug).

Otherwise apply the vendor-appropriate transition. This is the shipped → verified PASS hop the prd-lifecycle-rollup rule defines (cite it by slug; this skill is its PASS-path implementation, not a second source of truth):

• GitHub / Linear — remove the shipped label and add the verified label. For GitHub:

  gh issue edit <prd-num> --repo <org>/<repo> --remove-label "$SHIPPED" --add-label "$VERIFIED"
  

  Verify exactly one PRD-lifecycle label remains afterward (the single-label invariant github-prd-intake enforces) — a re-run must never leave both $SHIPPED and $VERIFIED, nor two copies of $VERIFIED. For GitHub, close the PRD issue immediately after the label transition with gh issue close <prd-num> --repo <org>/<repo> --reason completed; if it is already closed, record that native closure was already satisfied. For Linear, set the project/issue label equivalently, then archive/complete the project or issue using the vendor's native terminal mechanism where available.
• Notion — set the PRD page's notion.statusProperty (default Status) to the resolved verified value (default Verified). A status property holds exactly one value, so re-setting the same value is inherently a no-op. Then archive the page through lisa-notion-access where supported; if the page is already archived, record that native archival was already satisfied.
• Confluence — move the PRD page's parentId to confluence.parents.verified (the parent-page-based lifecycle; Atlassian scoped tokens cannot write labels — see config-resolution). A page has exactly one parent, so re-parenting to the same parent is a no-op. Then archive the page where the deployment supports archival; if archival is unavailable, report the capability-aware no-op or setup gap.
• JIRA — transition the PRD issue to the configured verified status. An issue holds exactly one status; if already verified, skip the transition. Then resolve/close the issue using the configured terminal workflow transition where supported.

verified is the terminal, product-owned PRD state; this skill is the only automated writer of it (intake/rollup never set it). After the verified transition, close, archive, or complete the PRD natively where the source vendor supports it. This close-out is mandatory and idempotent; do not introduce a configuration flag to skip it.

6.3 — Post verification evidence on the PRD

Post a verification-evidence comment back on the PRD, in the spirit of tracker-evidence (the vendor-neutral evidence poster). Because the evidence lands on the PRD source — which may be Notion or Confluence, not a tracker ticket — post via the vendor surface that owns the PRD: gh issue comment for GitHub, the Linear comment API, the Notion/Confluence page comment surface, or a JIRA comment. Where the PRD lives in the configured tracker, you may dispatch through tracker-evidence; for Notion/Confluence/cross-vendor PRDs, comment on the PRD page directly.

Idempotent — regenerate the evidence comment in place, never append. Lead the comment body with the stable sentinel marker <!-- lisa:verify-prd-evidence -->. Before posting, look up an existing evidence comment authored by this skill on the PRD whose body contains that sentinel (the same regenerate-don't-append discipline prd-backlink uses for its ## Tickets section; match by the marker, never by comment text or position). If one exists, edit it in place with the freshly regenerated body; only create a new comment when none exists. Per vendor:

• GitHub — gh issue view <prd-num> --repo <org>/<repo> --json comments and select the comment whose body contains <!-- lisa:verify-prd-evidence -->; update it with gh api -X PATCH /repos/<org>/<repo>/issues/comments/<comment-id> -f body=@evidence.md. Only gh issue comment <prd-num> --body-file evidence.md when no marked comment exists.
• Linear — find the existing comment carrying the sentinel and update it via the Linear comment-update API; create only if absent.
• Notion / Confluence — update the existing marked page comment in place; create only if absent.
• JIRA — update the existing marked comment in place; create only if absent.

The marked comment is the single canonical evidence comment for the PRD — a re-run refreshes it, never stacking a second one. The evidence comment MUST include:

1. Sentinel marker — the literal <!-- lisa:verify-prd-evidence --> as the first line, so the next run finds and regenerates this exact comment.
2. AI disclosure — lead with "PRD-level verification by Claude (AI agent, not a human)."
3. Verdict line — shipped → verified — PASS.
4. Spec-conformance coverage matrix — the section-by-section matrix from Phase 4 verbatim, with the CONFORMS verdict.
5. Empirical proof artifacts — the Phase 5 artifacts per surface: screenshots (upload via gh release upload pr-assets <files> --clobber and reference as plain URLs, per the tracker-evidence UI Evidence Checklist), request/response captures, query outputs, log excerpts, and the codified regression test(s) added.
6. What was verified — which PRD acceptance criteria each artifact covers, and the verification surface used.

Then emit the PASS output block (below).

Phase 7 — FAIL: re-open as ticketed, create build-ready fix tickets, post a failure report (never blocked)

Reach this phase when verification did not pass — i.e. either is true:

• Phase 4 spec conformance returned PARTIAL or DIVERGES (CONFORMANCE_FAILED cause), or
• Phase 5 had any applicable empirical check fail or a required surface unavailable (EMPIRICAL_FAILED cause).

PRD verification failure is self-healing, not a dead end. Instead of parking the PRD in blocked for a human, this phase: (1) moves the PRD back to ticketed (work in flight again), (2) creates build-ready fix tickets for the gaps so the build queue picks them up with no human promotion, and (3) posts a failure report. The fix tickets are added to the PRD's generated top-level work, so the existing machinery closes the loop on its own: the fix tickets build → reach terminal → the *-prd-intake rollup (Phase 3f) re-ships the PRD ticketed → shipped → the next intake cycle (Phase 3g) re-dispatches /lisa:verify-prd → PASS gives verified, FAIL runs this phase again with another round of fix tickets. This is the FAIL counterpart of the Phase 6 PASS hop and is the shipped → ticketed FAIL hop the prd-lifecycle-rollup rule's "Closing the loop" section defines (cite it by slug; this skill is its FAIL-path implementation, not a second source of truth).

PRD verification never moves the PRD to blocked. blocked is the intake (ready-stage validation) failure state, not the verification failure state — there is no prd-verifying / prd-verification-failed state either; the lifecycle stays small. The loop never auto-halts; the failure report carries a verification-round count so a human can spot a PRD stuck across repeated rounds, but the skill keeps creating fix tickets and re-verifying.

Carry forward the verdict cause and the concrete findings that produced it: from CONFORMANCE_FAILED, the matrix's missed/divergent/scope-crept requirements; from EMPIRICAL_FAILED, each failing check (the requirement/AC it exercised, the tool/command, observed vs expected, and any captured artifacts). These findings drive both the failure report (7.3) and the fix issues (7.4).

7.1 — Resolve the shipped and ticketed roles

Resolve the PRD-lifecycle roles from .lisa.config.json (then .lisa.config.local.json override) per the config-resolution rule — the same role-resolution Phase 6.1 and the *-prd-intake skills use. Cite config-resolution for the role vocabulary; do not hardcode label strings except as the documented defaults. Resolution per vendor:

For GitHub, resolve with the same helper Phase 6.1 / github-prd-intake use:

read_role() {  # role default → resolved value (local override wins)
  local role="$1" default="$2" local_v global_v
  local_v=$(jq -r ".github.labels.prd.${role} // empty" .lisa.config.local.json 2>/dev/null)
  global_v=$(jq -r ".github.labels.prd.${role} // empty" .lisa.config.json 2>/dev/null)
  echo "${local_v:-${global_v:-$default}}"
}
SHIPPED=$(read_role shipped prd-shipped)
TICKETED=$(read_role ticketed prd-ticketed)

7.2 — Re-open the PRD shipped → ticketed

Idempotency guard (no-op if already ticketed). Before transitioning, read the PRD's current lifecycle role. If the PRD already carries $TICKETED (the common re-run case — a prior failed round already re-opened it and its fix tickets are still in flight), the transition is a no-op — do not re-add the label/status, do not re-remove $SHIPPED (already gone). Record it as already ticketed (no-op) and proceed to 7.3, where the existing failure report is updated in place (not stacked, round count incremented) and 7.4, where existing fix tickets are referenced rather than re-created. This mirrors github-prd-intake Phase 3f.1's "no-op if already shipped" guard and the prd-lifecycle-rollup "rollup is keyed by the PRD's current state" rule (cite both by slug).

Otherwise apply the vendor-appropriate transition. This is the shipped → ticketed FAIL hop from the prd-lifecycle-rollup rule's "Closing the loop" section (cite it by slug) — a deliberate backward hop that puts the PRD back into "work in flight" so its new fix tickets are the in-flight work the rollup waits on:

• GitHub / Linear — remove the shipped label and add the ticketed label. For GitHub:

  gh issue edit <prd-num> --repo <org>/<repo> --remove-label "$SHIPPED" --add-label "$TICKETED"
  

  Verify exactly one PRD-lifecycle label remains afterward (the single-label invariant github-prd-intake enforces) — a re-run must never leave both $SHIPPED and $TICKETED, nor two copies of $TICKETED. For Linear, set the project/issue label equivalently.
• Notion — set the PRD page's notion.statusProperty (default Status) to the resolved ticketed value (default Ticketed). A status property holds exactly one value, so re-setting the same value is inherently a no-op.
• Confluence — move the PRD page's parentId to confluence.parents.ticketed (the parent-page-based lifecycle; Atlassian scoped tokens cannot write labels — see config-resolution). A page has exactly one parent, so re-parenting to the same parent is a no-op.
• JIRA — transition the PRD issue to the configured ticketed status. An issue holds exactly one status; if already ticketed, skip the transition.

Do not close or archive the PRD here, and never move it to blocked — ticketed signals "verification found gaps; fix work is in flight," and the PRD re-ships and re-verifies automatically once that work lands.

7.3 — Post a product-readable failure report on the PRD

Post a failure report comment back on the PRD, via the same vendor surface Phase 6.3 uses for evidence (gh issue comment for GitHub, the Linear comment API, the Notion/Confluence page comment surface, or a JIRA comment; dispatch through tracker-evidence where the PRD lives in the configured tracker). The report is written for a non-engineer product owner — plain language, no stack traces dumped raw. Capture its URL/anchor so the fix issues in 7.4 can back-link to it.

Idempotent — regenerate the failure report in place, never append. Lead the comment body with the stable sentinel marker <!-- lisa:verify-prd-failure-report -->. Before posting, look up an existing failure-report comment on the PRD whose body contains that sentinel (match by the marker, never by comment text or position — the same regenerate-don't-append discipline as Phase 6.3). If one exists, edit it in place with the freshly regenerated findings (so a re-run-after-a-previous-failure refreshes the same report rather than stacking a second one); only create a new comment when none exists. The GitHub mechanics are identical to Phase 6.3 (gh issue view --json comments to find the marked comment, gh api -X PATCH .../issues/comments/<id> to update, gh issue comment only when absent). It MUST include:

1. Sentinel marker — the literal <!-- lisa:verify-prd-failure-report --> as the first line, so the next run finds and regenerates this exact comment.
2. AI disclosure — lead with "PRD-level verification by Claude (AI agent, not a human)."
3. Verdict line + round — shipped → ticketed — FAIL (re-opened for fixes), the cause (CONFORMANCE_FAILED or EMPIRICAL_FAILED), and Round: N — the count of failed verification rounds for this PRD. Read the prior in-place failure report's round and increment (start at 1). The loop never auto-halts on a high count, but surfacing it lets a human notice a PRD stuck across repeated fix-and-re-verify rounds.
4. What failed, in plain language — for each finding, name the specific PRD requirement / acceptance criterion that was not met, then what was expected vs what was observed (the empirical evidence: what was checked, what the shipped product did instead). One bullet per finding so product can follow each independently.
5. Spec-conformance coverage matrix — for a CONFORMANCE_FAILED cause, the section-by-section matrix from Phase 4 verbatim with the PARTIAL/DIVERGES verdict, so the missed/divergent/scope-crept rows are visible.
6. Proof artifacts — any captured empirical artifacts (screenshots uploaded via gh release upload pr-assets <files> --clobber and referenced as plain URLs per the tracker-evidence UI Evidence Checklist, request/response captures, query outputs, log excerpts).
7. Fix issues — a list of the fix issues created/referenced in 7.4 (filled in after 7.4 runs, or posted as a brief follow-up edit), so the report is the single product-facing index of "what's wrong and where it's being fixed."

7.4 — Create linked fix issues for the missing/incorrect behavior

For each failed/missing/incorrect/divergent finding, create a build-ready fix ticket via tracker-write with build_ready: true (the vendor-neutral writer) — never by hand-rolling gh issue create, so each ticket passes the same quality gates (tracker-validate) every Lisa ticket does: three-audience description, Gherkin acceptance criteria, labels, and explicit relationship discovery. build_ready: true makes the build queue (lisa-intake build side / *-build-intake) auto-claim it with no human promotion — that is what makes the loop self-healing. Group findings that share one root cause into one fix ticket; do not fan out one ticket per matrix cell when several rows are the same defect.

Idempotent — dedupe fix issues by a stable marker; reference/update, never duplicate. This is the "re-run after a previous failure with the same missing behavior" scenario: the prior run already opened a fix issue for that requirement, so the re-run must find and reuse it, not create a second one. Apply the prd-lifecycle-rollup idempotency dedupe key discipline (match by a stable ref, never by title):

1. Compute a stable dedupe key per finding — the PRD ref plus a stable requirement/AC identity (e.g. the AC's heading/number or a slug of the requirement), independent of any mutable wording. Encode it in the fix-issue body as the marker <!-- lisa:verify-prd-fix prd=<prd-ref> req=<stable-req-id> -->.
2. Look up an existing OPEN fix issue carrying that exact marker before creating anything. On GitHub, search the repo for open issues whose body contains the marker (gh issue list --repo <org>/<repo> --state open --search '"<!-- lisa:verify-prd-fix prd=<prd-ref> req=<stable-req-id> -->"' --json number,url,body — or fetch and grep the marker); on Linear/JIRA, query by the marker stored in the body/a custom field. Match on the marker, never on the issue title (a title may have been edited; the marker is the stable identity).
3. If a matching open fix issue exists, reference/update it — do not create a duplicate. Refresh its captured evidence (the latest observed-vs-expected) and re-affirm the back-links to the PRD and the regenerated failure report, then fold its existing ref into the failure report's Fix issues list. A closed prior fix issue does not suppress a new one — if the requirement is failing again after the fix was closed, that is a regression and a fresh fix issue is correct.
4. Only when no matching open fix issue exists, create a new one via tracker-write.

Each fix issue (whether freshly created or referenced/updated) MUST:

1. Carry the dedupe marker — <!-- lisa:verify-prd-fix prd=<prd-ref> req=<stable-req-id> --> in its body, so the next run finds and reuses it.
2. Reference the specific failed requirement/AC — quote or cite the exact PRD requirement / acceptance criterion the finding violated, so the fix is scoped to a real gap (not a vague "make it work").
3. Carry the captured evidence — the observed-vs-expected from the failure report (what was checked, what was expected, what the shipped product did), so an implementer can reproduce without re-deriving it.
4. Back-link to the PRD and the failure report — link to the PRD (so the fix rolls back up to the initiative) and to the failure-report comment from 7.3 (so the full context is one click away). On GitHub, reference the PRD issue number and the failure-report comment URL in the body and, where supported, as a sub-issue/Relates to link; on Linear, set the relation; on JIRA, add the issue link and remote link.
5. Have acceptance criteria — Gherkin ACs describing the corrected behavior (what "fixed" looks like), enforced by tracker-write → the vendor *-validate-issue gate.
6. Be build-ready and counted as PRD work — created via tracker-write with build_ready: true, and registered in the PRD's generated top-level work: refresh the PRD's ## Tickets /

…(truncated)