docs/solutions/skill-design/new-knowledge-source-re-derives-persona-gate-and-route.md

---
title: "Wiring a new knowledge source into a conditional persona re-derives its spawn gate and its synthesis route"
date: 2026-09-08
category: skill-design
module: compound-engineering / ce-code-review
problem_type: design_pattern
component: tooling
severity: high
applies_when:
  - "Adding a knowledge source (packs, standards files, a corpus) to a persona another gate already selects"
  - "A guide promises enforcement at a stage whose only carrier is a conditional persona"
  - "A fresh-agent eval of a gated stage passed, but the run forced or short-circuited the gate"
tags:
  - ce-code-review
  - compound-packs
  - spawn-gate
  - synthesis
  - cloud-agent-dogfood
---

# Wiring a new knowledge source into a conditional persona re-derives its spawn gate and its synthesis route

## Context

Compound Packs added pack roots to `ce-code-review`'s `learnings-researcher` search-root list, and `docs/guides/packs.md` promised that a diff violating a pack rule is flagged with a citation. The branch's own dogfood matrix marked that leg as passing. Cloud-agent dogfood on the merged branch, run strictly "as written, do not force a persona the gate does not select", found on three model families (Claude Sonnet 5, GPT-5.6 Sol, Grok 4.6) that review emitted no pack citation at all on a repo without `docs/solutions/`: the persona's spawn gate still required a plausible match in an existing solutions corpus, and `review-scope.py` only knew about `docs/solutions`. A second round found the persona's output had no stated route into the numbered finding set, so in `mode:agent` a violation would sit in the `learnings` array where `lfg` never applies it.

## Guidance

A knowledge source is not "added" to a persona when its roots join the persona's search list. Three things own the outcome, and each must be re-derived from the source's guarantee:

1. **The spawn gate.** State the condition the persona serves (here: "there is institutional knowledge to check the change against"), and make every source that satisfies it select the persona on its own terms. Packs are declared and read-everything, so declaring them selects the persona without a corpus pre-search; a learnings corpus still needs the cheap match. Give the mechanical helper a signal for the new fact (`declared_packs`) so selection is not left to memory.
2. **The size/fast paths.** A persona selected by a repo-level criteria source (standards paths, declared packs) is not a diff-content conditional; it rides the lite roster the way `project-standards` already does, or a three-line violation escapes enforcement.
3. **The synthesis route.** If the persona's output is unstructured, say which of its outputs are findings and how they enter the helper's finding set (a complete compact reviewer return, before the first helper run), and which stay notes. "Flagged with the citation" means a numbered, actionable finding, not an advisory paragraph.

Then evaluate the gate the way it will be hit in production: a fixture that has the new source and **lacks** the persona's old trigger, with the instruction to follow the skill literally and report what the gate decided. An eval that passes because a capable model dispatched the persona anyway has not tested the gate.

## Why This Matters

The original leg passed because agents read the guide, saw packs should be searched, and dispatched the persona regardless of the gate; the weaker the model or the stricter the instruction, the more the gate wins over the guide. A gate gap of this kind is silent: nothing errors, the review just never mentions the rule, and a team adopting packs before it has learnings is exactly the population the guide targets.

## When to Apply

- Any time a stage's promise ("review enforces X") is carried by a persona that some other condition selects.
- When a mechanical helper reports the facts a gate keys on and a new fact is being introduced.
- When a prior eval "passed" on a fixture that also satisfied the old trigger, or when the eval prompt did not forbid forcing the persona.

## Examples

Before (gate stated as the old case):

> `learnings-researcher` — `<root>/solutions/` exists and a cheap path/title search finds a plausible match.

After (gate stated as its condition, with the new fact named and the helper carrying it):

> `learnings-researcher` — there is institutional knowledge to check the change against: `<root>/solutions/` exists and a cheap path/title search finds a plausible match, or, in local scope, the repo's CE config declares Compound Packs (Stage 1b `declared_packs`). Declared packs need no pre-search; the persona matches their rules itself.

Re-verification fixture: a scratch repo with a declared pack, no `docs/solutions/`, one violating diff, one 1-line lite-eligible diff, and a control with the `packs:` key removed — pass means the persona is selected in the first two and not the third, with `(pack: <id>, <path>)` on the finding in `actionable_findings`.

## Related

- `skills/ce-code-review/references/persona-catalog.md`, `skills/ce-code-review/references/select-and-route.md`, `skills/ce-code-review/references/scope.md`, `skills/ce-code-review/references/finish-review.md`, `skills/ce-code-review/scripts/review-scope.py`
- `tests/skills/ce-packs-contract.test.ts` ("selects the learnings persona for declared packs", "a contradicted pack rule becomes a numbered finding"), `tests/ce-code-review-mechanics.test.ts` (declared-packs helper cases)
- `docs/solutions/skill-design/authored-eval-corpora-contain-the-happy-path.md` — the same failure shape at the eval-corpus level
- Found and fixed in PR #1656