Skip to content

THREAD_DESIGN.md

Latest commit

 

History

History
161 lines (112 loc) · 6.83 KB

THREAD_DESIGN.md

File metadata and controls

161 lines (112 loc) · 6.83 KB
Raw

Cognitive Thread v2 — Design Contract

Hypothesis-centric Research Thread for Paper_Rec. Differentiates from Anaxa (quest→manuscript audit), PaperSeek (generic discovery), YFR (topic monitoring/survey ops), and article-mcp (search MCP) by owning long-horizon personal research memory: claims, gaps, evidence gates, lit↔exp links.

See also: MCP.md. Competitor landscape notes are maintained privately outside this repository.

Non-goals (explicit)

  • Manuscript / LaTeX / citation-audit / PDF finalization (Anaxa)
  • Generic “best-in-class” search agent without thread context (PaperSeek)
  • Multi-tenant public survey / WeChat draft factory (YFR)
  • Duplicate literature-search MCP (compose article-mcp instead)

Layout

content/threads/<thread_id>/
  thread.json      # canonical structured state
  README.md        # human notes + mirror of title/status/hypothesis
  events.jsonl     # cognitive ledger (append-only)
  evidences.jsonl  # Claim–Evidence Map (quote/metric/figure/note)
  deltas/          # Watch/Delta markdown briefs (Phase B)

Phases (status)

Phase Scope Status
A Model, bridge, API, Wiki UI, Skill 1.5/2.5 done
B Watch/Delta, graph thread/exp nodes, claim gate updates done
C thread_* MCP (memory first) done
D Wiki exp Chart.js viz + /draw venue styles done
E Claim–Evidence Map + MCP zero-PYTHONPATH / wiki·exp tools done (2.9.0)
F Iterative retrieval (query_iter) + multi-run exp board done (2.10.0)
G Cognitive map graph + timeline; PDF-lite; BibTeX; Related Work outline done (2.12.0)
H Install scripts; evidence confidence; PDF upload pipeline; query auto; writing recommend; MCP tools done (2.14.0)
I configure MCP; paper_draft; prerank; evidence UX; citation-expand done (2.16.0)
J OA pdf-fetch; RRF; thread feedback; CSL-JSON; Docker done (2.19.0)
K Thread-Bench; optional Delta webhook done (2.20.0)
L Template marketplace; multi-channel Thread Bot done (2.21.0)
M Benchmark REPORT + LitSearch adapter (credibility P0/1A) done (2.22.0)
N CEBM-lite evidence_level (1a–5) supplement done (2.23.0)

Watch / Delta (Phase B)

CLI: thread-delta --mode auto|new_digest|diff_brief|gap_focus|exp_bridge
API: POST /api/threads/{id}/delta
Skill: /wiki thread delta

Claim proposals stay gate:suggested until thread-claim --accept or API /claims/accept.

Claim–Evidence Map (Phase E / 2.9.0+)

Evidence binds a claim_id to a paper quote, exp metric, figure, or note (evidences.jsonl).

Anaxa-parity spine (same grain):

claim_id → paper_path / exp_id → citation_key → quote + page → support_status → confidence + gate

Gates: suggested → accepted. Confidence ∈ [0,1] (subjective strength of this binding).

CEBM-lite supplement (JARVIS-inspired, optional, orthogonal):

Code Meaning
1a Systematic review of RCTs
1b Individual RCT
2a / 2b Cohort review / cohort
3a / 3b Case-control review / study
4 Case series / poor-quality
5 Expert opinion / mechanism / anecdote

Field: evidence_level (+ display evidence_level_label). Aliases: meta→1a, rct→1b, cohort→2b, anecdote→5, ….
Does not replace confidence or gate; use for design-level strength when writing Related Work / audits.

  • CLI: thread-evidence-add|list|gate (--evidence-level)
  • API: /api/threads/{id}/evidences, /evidence-map, gate endpoint; coverage includes cebm_histogram
  • Wiki: PageView「挂到主线」CEBM 下拉; ThreadDetail shows CEBM 1b …
  • MCP: thread_add_evidence

Does not auto-extract claims from full PDFs (human highlight + agent suggest only). Does not implement PRISMA system-review pipelines (out of scope vs JARVIS).

Iterative retrieval (Phase F / 2.10.0)

Skill Modules 2a/2b: multi-path queries + ≤1 refine wave when thread: or iterative. Ledger kind query_iter.

  • CLI: query-trace --thread … --json trace.json
  • API: POST /api/threads/{id}/query-trace
  • Report: retrieval_trace in JSON → auto-appended on sync-report --thread

Phase F+ live loop: deep-search CLI persists drafts/deep_search_*.md + query_iter

Multi-run curves (Phase F / 2.10.0)

metrics/curves.json + metrics/curves_<run>.json → API curve_runs[]. Wiki ExpDetail: run overlay, compare exp, primary-only, 5s poll.

Related Work outline (Phase G)

CLI: related-work --thread → content/threads/<id>/drafts/related_work_outline.md
API: POST /api/threads/{id}/related-work
Skill: /wiki thread related-work

PDF-lite + BibTeX (Phase G)

  • pdf-ingest → fulltext.md; claim-suggest --apply → suggested claims/evidences
  • bibtex-export; API GET /api/wiki/bibtex?paths=

Cognitive graph (Phase G)

CLI: thread-graph · API: GET /api/threads/{id}/graph (nodes/edges + day-grouped timeline)

MCP (Phase C + E)

Package: packages/thread-mcp — set PAPER_REC_ROOT only (auto-locates wiki-bridge).

Tools: thread_list|get|search_context|query_hint|score_papers|link_*|add_evidence|delta|claim_* · wiki_list_papers · exp_list|exp_get_metrics.

Venue styles (Phase D)

skill-draw/lib/venues.py + draw(..., venue="cvpr") / CLI --venue.

Template marketplace (Phase L / 2.21.0)

Reusable research-line packs (hypothesis / claims / gaps / seeds), not manuscript templates.

content/thread-templates/<template_id>/
  template.json   # catalog meta (title, tags, claims_n, builtin?)
  thread.json     # sanitized snapshot (no local paper_paths)
  README.md
  drafts/         # optional
  • CLI: thread-template-list [--seed] · thread-template-export · thread-template-import
  • API: GET /api/threads/templates · POST /api/threads/templates/import · POST /api/threads/{id}/export-template
  • Wiki: Threads 页「主线模板市场」；详情「导出为模板」
  • Builtins (seed): multimodal-alignment · rag-evaluation · code-agents

Multi-channel Bot (Phase L / 2.21.0)

Package packages/thread-bot — unified command router over Feishu / Telegram / WeCom / QQ(OneBot). See BOTS.md.

One-way Delta push (no command loop): WEBHOOK.md.

Bridge CLI

thread-create | thread-list | thread-show
thread-link-paper | thread-link-exp
thread-evidence-add | thread-evidence-list | thread-evidence-gate
thread-delta | thread-claim
thread-template-list | thread-template-export | thread-template-import
query-trace
sync-report --thread <id> [--auto-link]
sync-exp --thread <id>

API

/api/threads — CRUD + link + timeline + delta + context + score + claims + evidences + evidence-map + query-trace + templates + by-paper / by-exp. /api/exp/{id} — includes curves + curve_runs.