| name | research-pipeline |
| description | MISHKAN shared research pipeline. Use when any agent faces an unknown and must research it. Orchestrates Jakin (clarify) → Ezra (formulate) → Caleb (web research) → Shaphan (summarise) → Shemaiah (evaluate) → Baruch (report). Produces a research-log.json entry and, on resolve, a Cognee node. Invokable by any agent at any time. |
Research Pipeline
The shared research workflow. Any agent that hits an unknown invokes this rather
than guessing. The pipeline is six stages, run in order, each a separate subagent.
When to use
- An agent faces a problem it cannot resolve from its own knowledge or the
curated library.
- Y4NN explicitly requests research.
Stages
- Jakin (clarificator) — raw query → clarified intent + open questions.
If intent is unclear, resolve open questions with the caller before proceeding.
- Ezra (formulator) — clarified intent → research brief. Checks the curated
library / Cognee first; if matched, short-circuit and skip to step 6 with
curated_library_match: true.
- Caleb (web researcher) — brief → raw findings with sources. Plans first
if the brief is multi-source.
- Shaphan (summariser) — raw findings → tight summary, sources preserved.
- Shemaiah (evaluator) — summary → verdict (resolved/partial/blocked) +
confidence + gaps + curated-library agreement.
- Baruch (reporter) — emits the
research-log.json entry and, on resolve
with cross-harness blast radius, writes a Cognee node.
How to run it
Invoke each stage as a subagent via the Task tool, passing the prior stage's
output as input. Do not collapse stages — generation and evaluation stay separate
(Shaphan summarises, Shemaiah judges, Baruch records; none of them do each
other's job).
Output contract
The pipeline returns one research-log.json object validating against
~/.claude/mishkan/templates/research-log.schema.json, plus a cognee_node_id
when a node was written. The calling agent attaches this log to its task; the
Team Reporter collects it at milestone.
Model tiers
Jakin, Ezra, Caleb run on Sonnet. Shaphan, Shemaiah, Baruch run on Haiku. No
local models — Claude Code routes tiers natively per each agent's frontmatter.
Constraints
Every stage obeys the universal constraints: stateful operations hard stop,
scope boundary, no fabricated facts, English output. The pipeline never executes
stateful operations — it only researches.
