Data-analysis ledger (analysis_reports)

Analysis conclusions reached inside a channel can be persisted as first-class records: a dedicated analysis_reports table organized by line × recipe × time window, with key conclusions, metrics, data sources and actor attribution. The /analysis page provides multi-dimensional retrieval with live updates; the external analysis service (IDD) plugs in through diag_run as a background task, and its reports land in the same ledger.

Agent-native tools

Tool	Purpose	Key arguments
analysis_save	Persist an analysis record	title/summary required; conclusions (array), report (full text), metrics (object), recipe_id, window_from/to, data_sources; line resolution = explicit line_id → channel-bound line fallback
analysis_query	Query the ledger	report_id fetches the full record; otherwise list by line_id / recipe_id / category / q / from/to

The write gate mirrors line-data reads: an agent can only write records for lines it can read. The ledger is append-only for agents; every write is audited via recordOps (analysis.save) and streamed to the frontend over WS.

Human side (/analysis page)

• Multi-dimensional filters: line dropdown / category / time presets (1h · 24h · 7d · 30d) / keyword (title + summary + conclusions).
• Row expansion: full report body + metrics + data sources + attribution (actor / channel).
• Revision: users with operate line access can edit (PATCH); editor attribution (updated_by*) refreshes on the row. Agents always see the latest revision.
• Delete: admin only (data hygiene), audited as well.

Permission isolation (consistent platform-wide)

The ledger filters by line-level grants (visibleLineIds): without a grant, lists come back empty, details return 403/404, and HITL snapshots and ops-audit logs are equally invisible — data never leaves the server. readonly grants can look but not touch; operate grants can revise. See Line-level permissions.

IDD integration (diag-bridge)

diag_run starts a background analysis task and returns a task_id immediately — the channel never waits:

diag_run({ line, analysis_intent: "stability-report", question: "…" })
→ { task_id }
diag_status({ run_id: task_id })   // track progress by taskId
→ running (with stage) → completed + report file path

analysis_intent selects the IDD-side skill protocol (full intent catalog; executed by IDD's harness, decoupled from the platform):

Intent	Purpose
deep-diagnosis (default)	nine-stage deep root-cause pipeline (heaviest)
stability-report	line stability report: score + grade + advice (no root cause)
quick-screen	fast screening: over-limit/drift/freeze anomaly list + HTML dashboard (fastest)
doe-analysis	DOE analysis: factor effects / significance / optimal settings
experience-mining	tuning-experience mining: structured entries for the knowledge base
optimizer-campaign	optimization campaign: iterative search + recommended settings (heaviest)

Each intent carries a scenario-specific default question (business semantics injected by the bridge, executed by IDD). On completion the platform auto-delivers the result to the originating channel (score / verdict / summary / absolute report path), and the agent persists the conclusions (with run_id provenance) via analysis_save.

RAG mirrors this: kb_agent supports async tasks — mode:"async" returns a task_id immediately while the knowledge base ingests in the background, with kb_agent_status tracking by taskId; retrieval Q&A uses mode:"sync" for an immediate cited answer.

Retention

retention.analysis_days (default 365) controls ledger retention and participates in the unified retention sweep.

Results library

The diag-bridge plugin now integrates IDD's results library (index / detail / metadata editing). It is off by default — the visible result set equals the diagnosis API token owner's visibility, and results are not line-scoped, so an admin must enable it explicitly.

Agent tool	Purpose
diag_results	search the result index (q/intent/category_id/tag/scene/harness/min_score/sort/limit/offset; brief=true by default to save tokens, brief=false includes key_numbers)
diag_result_get	fetch detail by result_id (summary + traceable key numbers + absolute artifact/HTML URLs; optional include_md, truncated by limit)
diag_result_annotate	edit metadata (description/tags/title; description=null clears it; lead-only + plugin switch; ownership enforced by IDD)

Settings (Plugins → diag-bridge → Results library): results_enabled (master switch, default off), results_allow_annotate (editing, default off), results_allow_worker (worker read access, default off = lead only), results_default_limit (default 20), results_md_max_chars (report truncation limit, default 4000). The /plugins page ships a Results library panel: search/filter → detail → edit description & tags (same-source routes GET /api/plugins/diag-bridge/results, /results/detail?id=, PATCH /results/edit?id=).

Typical agent flow: diag_results(q="film thickness") → pick result_id → diag_result_get for conclusions and key numbers → diag_result_annotate to record what the result means and when to reuse it.

The ledger records conclusions; the EDA studio explores data — for interactive analysis (charts / pipeline / SPC / golden) see Data workbench & EDA studio.