Deep Architectural Research¶
Deep architectural research is a goal variant of the brainstorm workflow, not a separate Task kind. It walks a tatara-deep-architectural-research skill decision tree that adds a "survey the field" step to the ordinary brainstorm loop and frames output as a long-lived ADR-style artifact rather than a plain proposal issue. This is Phase 1 of a four-phase design (docs/2026-06-23-deeper-research-engine.md); only Phase 1 is deployed today.
No dedicated research Task kind
An earlier design considered a distinct architectural-research Task kind with its own egress-allowed pod profile. That was not adopted - the decision on record is to widen the existing brainstorm pods instead of introducing a separate kind. Everything below runs inside the brainstorm kind.
What Phase 1 actually ships¶
| Piece | Status | Where |
|---|---|---|
tatara-deep-architectural-research SKILL.md + ADR template | Live | tatara-agent-skills/skills/brainstorming/tatara-deep-architectural-research/ |
skip_research MCP tool | Live | tatara-cli MCP tools, registered in the brainstorm tool profile |
archfitness import-graph checker | Live, but currently synthetic-test-only | tatara-operator/internal/archfitness |
| ADR-framed brainstorm prompt | Live | brainstorm goal-prompt builder |
| Serena MCP env-gate plumbing | Plumbed end-to-end but inert | TATARA_SERENA_URL unset in helmfile; no Serena deployed |
| Dedicated research Task kind, web-search/arXiv/OpenAlex/Firecrawl MCP servers, egress policy, tournament-rank proposal admission, SCMProvider port refactor, skills-archive/self-improvement loop | Not built - design-doc only (Phases 2-4) | n/a |
skip_research¶
skip_research(reason) is an escape-hatch MCP tool available to brainstorm agents. It requires a non-empty reason and posts {"action": "none", "reason": ...} to the Task's brainstorm outcome endpoint - the same outcome shape as a normal no-yield brainstorm cycle. It exists so an agent that surveys a problem and finds nothing novel or shippable can terminate cheaply instead of running an expensive research fan-out for no result.
skip_research is a tool name, not a CRD field - there is no skipResearch field on the Project or Task spec.
The ADR artifact¶
Where a normal brainstorm proposal is a single issue body forbidden from listing open questions, a research-flavored proposal is framed as an ADR: problem statement with file:line evidence, 2-3 options with tradeoffs, a recommendation, and (unlike a normal proposal) open questions are explicitly allowed. It is filed the same way as any other proposal - via propose_issue, landing as an issue body - it is not written to an in-repo docs/adr/ directory. Cross-cycle championing (letting a research thread survive multiple brainstorm cycles, exempt from the one-issue-per-cycle / silence-over-noise defaults) is called out in the design doc as still deferred.
archfitness¶
archfitness (tatara-operator/internal/archfitness/archfitness.go) is an import-graph checker: it loads the module dependency graph and fails a check when core code imports a designated "banned" vendor SDK (e.g. a raw GitHub/GitLab client) outside an adapter package. It is the guardrail intended for a future strangler-pattern refactor (e.g. extracting an SCMProvider port) - today it has zero live violations to flag, since that refactor has not started.
Egress¶
Research still rides the existing spec.scm.cron.brainstorm.sources egress hook (see Brainstorm): setting internet in sources stamps tatara.io/egress: internet on the brainstorm Pod. No dedicated web-search, academic paper, or crawl MCP server is wired to consume that egress yet - that is Phase 2 scope.
Current capability (Phase 1)¶
As deployed, deep architectural research is: the existing brainstorm loop, plus a skill that asks the agent to frame findings as an ADR against the in-cluster knowledge graph and on-disk sources, plus a cheap early-exit tool (skip_research) and an as-yet-unenforced import-fitness check (archfitness). It is not a live, internet-connected, multi-source research agent: the internet-facing, multi-agent "survey the field" capability from the original design (Phases 2-4) is unimplemented. Enrolling internet in brainstorm.sources only stamps an egress label; no web-search, academic-paper, or crawl MCP server is wired to consume it yet.