[Tracking] llmXive agentic pipeline — phase + agent map

[jeremymanning]

[Tracking] llmXive agentic pipeline — phase + agent map

This is the umbrella tracking issue for the llmXive agentic pipeline.
Each phase has its own issue with sub-issues per agent, complete with
descriptions, dependencies, acceptance criteria, and manual test plans.

See notes/2026-05-04-pipeline-status.md
for the most recent status, including the verdict trend across the 5-cycle
PROJ-024 run and the structural bug that prevented convergence to all-accept.

Phases

Research-side

• [Phase 1] Idea Lifecycle #45 [Phase 1] Idea Lifecycle (brainstorm, flesh_out, research_question_validator, idea_selector) — DONE via spec 003
• [Phase 2] Project Bootstrap #46 [Phase 2] Project Bootstrap (project_initializer)
• [Phase 3] Spec Kit: Specify → Clarify #47 [Phase 3] Spec Kit: Specify → Clarify (specifier, clarifier)
• [Phase 4] Spec Kit: Plan → Tasks (with Analyze loop) #48 [Phase 4] Spec Kit: Plan → Tasks (with Analyze loop) (planner, tasker)
• [Phase 5] Implementation #49 [Phase 5] Implementation (implementer) — KNOWN STRUCTURAL BUG
• [Phase 6] Research Review (8-reviewer panel) #50 [Phase 6] Research Review (8-reviewer panel)
• [Phase 7] Advancement & Verdict Routing #51 [Phase 7] Advancement & Verdict Routing

Paper-side

• [Phase 8] Paper Bootstrap #52 [Phase 8] Paper Bootstrap (paper_initializer)
• [Phase 9] Paper Spec → Clarify #53 [Phase 9] Paper Spec → Clarify (paper_specifier, paper_clarifier)
• [Phase 10] Paper Plan → Tasks #54 [Phase 10] Paper Plan → Tasks (paper_planner, paper_tasker)
• [Phase 11] Paper Implement (dispatcher + sub-agents) #55 [Phase 11] Paper Implement (dispatcher + 7 sub-agents)
• [Phase 12] Paper Review (12-reviewer panel) #56 [Phase 12] Paper Review (12-reviewer panel)

Cross-cutting

• [Phase 13] Cross-cutting: Atomization & Joining #57 [Phase 13] Atomization & Joining (task_atomizer, task_joiner)
• [Phase 14] Posting, Reporting & Hygiene #58 [Phase 14] Posting, Reporting & Hygiene (status_reporter, repository_hygiene)

Top priority

#49 (Phase 5 — Implementation) is the #1 blocker for end-to-end
convergence. Until the Implementer re-verifies literal task assertions
against the filesystem, no project will reliably reach research_accepted
without 5+ revision cycles per project.

How to use these issues

1. Pick the phase (or specific agent) you want to work on.
2. Read the description, dependencies, and acceptance criteria.
3. Run the manual test plan to confirm current behavior.
4. Implement / patch.
5. Re-run the manual test plan; tick off the acceptance-criteria checkboxes.
6. Close the agent sub-issue when its checklist is complete.
7. Close the phase issue when all its agent sub-issues are closed AND the
   phase-level smoke test passes.

[jeremymanning]

Phase 1 (#45) — DONE

Spec 003 (003-phase1-idea-lifecycle-testing) merged to main. Phase 1 is now demonstrably working end-to-end on real backends with real projects.

Architectural change: Phase 1 expanded from 3 agents to 4 with the addition of research_question_validator between flesh_out and project_initializer. The new agent runs a four-check audit (phenomenon-vs-method, circularity, triviality, question-narrowing) and emits one of three verdicts (validated / validator_revise / validator_rejected) with sentinel-file routing.

Sub-issues closed:

• [Agent · Phase 1] brainstorm #59 brainstorm — DONE (FR-003a scope content v1.0.0 → v1.1.0)
• [Agent · Phase 1] flesh_out #60 flesh_out — DONE (literature-gap-as-feature v1.0.0 → v1.1.0; research-question-quality rules v1.1.0 → v1.2.0)
• [Agent · Phase 1] idea_selector #61 idea_selector — superseded by validator (D11 wiring gap noted; class remains for future use)

Carry-forward to Phase 2 / spec 004: 2 projects at project_initialized:

• PROJ-261-evaluating-the-impact-of-code-duplicatio (clean validated path)
• PROJ-262-predicting-molecular-dipole-moments-with (full revise loop exercised)

See carry-forward.yaml and the diagnostic report. 16 defects identified during the spec; 12 resolved in-PR, 4 deferred (D1, D5, D11, D16) with severity tags + follow-up rationale.

[jeremymanning]

Spec 004 PR open: #109 — Phase 2 (project_initializer) verified end-to-end on iter3 siblings of PROJ-261 + PROJ-262. Carry-forward manifest at specs/004-phase2-project-bootstrap-testing/carry-forward.yaml names both iter3 siblings as substrate for spec 005 (Phase 3 testing).

[jeremymanning]

Spec 005 (librarian agent) PR opened: #110

Aggregate verdict: PASS (12/12 success criteria).

What the librarian consolidates (Constitution Principle I):

• Single canonical search-and-verify agent at src/llmxive/librarian/ + src/llmxive/agents/librarian.py
• Replaces 3 duplicate implementations (lit_search, citation_fetcher, citation_resolver) — soft-deprecated with banners pointing to the librarian
• Backends: Semantic Scholar Graph API + arXiv API only (Q1)
• Verification: 3-check (URL resolves + title-token-overlap ≥0.7 + summary-grounded ≥0.5) + ≥10% adaptive PDF sample (Q2)
• Multi-step expansion: LLM brainstorm 10–20 ranked alt phrasings, iterate ≥10 distinct queries until ≥5 verified or exhausted (Q3)
• 600s wall-clock budget per invocation (Q4)
• sha256-keyed cache at state/librarian-cache/ with 30d/7d/90d TTLs

Re-validation findings (PROJ-261 + PROJ-262):

• Both canonicals revalidate cleanly: judgment: verified
• Validator returns validated (4/4 sub-checks) on both under the new librarian-backed evidence
• New ## Search trail subsection populated in both ideas (5 verified citations each)
• Both proceed to spec 006 unchanged at project_initialized

Cross-domain coverage (8/8 fields PASS): biology, chemistry, computer science, materials science, neuroscience, physics, psychology, statistics — 72 verified citations total, max duration 380s, 4 fields fired expansion.

Defects: 7 found, 7 fixed in-PR. No CRITICAL, no deferred-CRITICAL.

Diagnostic report: notes/2026-05-07-spec-005-librarian-diagnostic.md