plan: #3134

[PolyphonyRequiem]

---

requests_parent_change: false
ancestor_plan_generations:
root: 1

Plan for #3134 (root #3131)

Promotes plan/3131-3134 into plan/3131.

AB#3134

[PolyphonyRequiem]

🤖 Polyphony plan reviewer (advisory)

Summary: Tight, well-scoped issue plan for the sufficiency-judge + deep-tier escalation. It correctly inherits schemas from issue-2, names the cap default and override path, and lists harness scenarios that map 1:1 to the routing branches. Main concerns are around (a) where the per-topic escalation counter physically lives in a stateless sub-workflow, (b) cite-both-sources being asserted only on the deep path while AC #3 is phrased unconditionally, and (c) the absence of an explicit failure mode if deep-researcher tools are unavailable.

Strengths

• Clear disjoint-surface note vs issue-4, with an explicit rebase-not-merge instruction — exactly the kind of guidance the executor wave needs.
• Profile schema extension is minimal and additive: optional field, sensible default, 0-disables semantics called out, negative-value validation specified.
• Harness scenarios enumerate the four meaningful routing paths (sufficient / insufficient-under-cap / cap-exhausted / cap-zero) — good coverage of the state machine.
• Honors the parent's "no tier-pinned models" decision and defers rigor-bar work to issue-6, so the scope stays inside one PG.

Concerns

• Per-topic counter location is under-specified. Step 5 says "track escalation count keyed by topic id within a single sub-workflow invocation" but doesn't say where: workflow state.*, an agent-emitted field threaded through routes, or a script node. In a conductor sub-workflow this matters for re-entry semantics and for the route condition in step 2 to be expressible without a hard-coded state name. Pin the mechanism.
• AC #3 ("findings cite both archive and live sources when needed") is phrased as a global criterion but is only verifiable on the escalation path. Either scope it to the deep-tier output explicitly, or add the symmetric criterion that archive-sufficient outputs cite archive sources only — otherwise the test matrix is ambiguous.
• No failure-mode AC for the deep tier. What happens if web/GitHub/MCP tools are unreachable, rate-limited, or return empty? The plan should say whether that collapses to archive-only with a marker (mirroring cap-exhausted) or surfaces an error to the architect. Today the harness scenarios don't cover this.
• Verdict schema is named but not pinned. { sufficient, gaps, rationale } is good shape, but gaps element type is unspecified. Issue-2 defined the findings-document schema; do the same here so the deep-tier agent has a contract to consume.
• "Recorded 'cap reached' marker" in the cap-exhausted scenario is the only place this marker appears. If downstream consumers (architect in issue-5, archivist in issue-4) need to distinguish capped-archive-only from naturally-sufficient-archive-only, name the marker field in the findings-document schema extension, not just in the test description.

Suggested changes

• Add a one-line "Counter mechanism" bullet under Proposed Approach naming the conductor primitive (e.g. workflow state field, agent-output passthrough) used to carry the per-topic count, and reference it from the route condition in step 2.
• Split AC #3 into two ACs: "deep-tier output cites at least one live source in addition to any archive citations" and "archive-sufficient output is unchanged from issue-2's contract."
• Add an AC + harness scenario for deep-tier tool failure (e.g. tool throws / returns empty) — degrade-to-archive-with-marker is the simplest contract and matches cap-exhausted.
• Pin the verdict schema (gaps: string[] or richer) and the "cap reached" marker as a named field in the findings document, so issue-4/issue-5 have a stable contract to read.
• Confirm none of the new YAML/script introduces hard-coded ADO state name strings ("Done", "Active", etc.) — the routing here is intra-workflow so it should be naturally clean, but worth an explicit no-op check given the recent regression history.

This is an advisory comment from the polyphony plan_reviewer agent. It does not vote — human review is the merge gate.

[PolyphonyRequiem]

polyphony:approve 0f49d41