[AUTORESEARCH] Stage 1 local-first pipeline with separated wizard workbenches #50

New Issue

Timmy · 2026-03-29T15:57:19Z

Timmy commented

2026-03-29 15:57:19 +00:00

Source: follow-on from current sovereignty direction and from timmy-home#29 (Gemini Deep Research ingestion).

Goal:
Define the first real autoresearch system Timmy can run without pretending to be a general autonomous genius.

This stage is not "let the model roam the internet and become wise."
It is a bounded, auditable pipeline that turns selected source updates into:

raw local artifacts
normalized reports with provenance
distilled briefings
candidate issues / actions for Timmy to review

Architectural decision

Autoresearch should respect the same separation we want everywhere else:

Local Timmy remains sovereign control plane
- owns routing, memory, audit trail, backlog decisions, and final delivery
- does not get blended into cloud identities
Claude-Hermes and Codex-Hermes can become distinct wizard workbenches
- each on its own owned VPS + Hermes harness clone
- cloud-first for heavy research or synthesis when useful
- still subordinate to local artifact storage and local review
Research output lands locally first
- raw outputs, normalized notes, scored summaries, and candidate actions live under ~/.timmy/
- cloud results are source material, not invisible authority

Stage 1 pipeline

0. Source manifest

A declarative manifest defines:

source type (rss, blog, arxiv, repo-watch, wizard-prompt, manual-drop)
cadence
owner (local-timmy, claude-wizard, codex-wizard, or other named worker)
topic tags
output target repo(s)

1. Fetch / capture

Collect the raw artifact:

RSS/blog entry text
arXiv abstract / metadata
repo release note / changelog
wizard-generated research memo

2. Normalize with provenance

Convert every input into a common artifact format containing:

source URL or origin
fetched_at timestamp
source type
agent / wizard that fetched it
raw excerpt or file pointer
confidence notes

3. Deduplicate and rank

Local deterministic logic scores each artifact by:

relevance to Timmy priorities
novelty
actionability
trust / provenance quality

4. Distill into briefing

Create a short local briefing that answers:

what changed
why it matters now
what repo / system layer it affects
whether it is shell work, control-plane work, or wizard workbench work

5. Generate candidate actions

Emit structured candidate actions such as:

Gitea issue draft
design note
experiment request
"ignore for now" decision

6. Review gate

Local Timmy reviews the candidates before they enter the live backlog.
Stage 1 should not auto-spam issue trackers.

Required artifact layout

Target local runtime layout:

~/.timmy/autoresearch/
  manifests/
  raw/
  normalized/
  briefings/
  candidates/
  logs/

Candidate records should include at minimum:

id
source_ids
target_repo
proposed_title
proposed_body
confidence
action_type
created_at
produced_by

Acceptance criteria

A committed Stage 1 spec exists in timmy-home for this pipeline.
At least one manifest-driven run can complete end-to-end locally.
The run produces all five artifact classes:
- raw
- normalized
- briefing
- candidate
- run log
At least three source types are supported in Stage 1, including:
- one local / deterministic source
- one web or feed source
- one wizard-generated research memo source
Every candidate action is source-grounded and includes provenance.
The pipeline can propose an issue without directly posting it.
timmy-home#29 is reframed as one feed-specific subproblem, not the whole autoresearch system.

Non-goals

full autonomous internet agent
unsourced issue generation
hidden cloud dependence for memory or routing
blending local Timmy with Claude/Codex wizard identities
automatic backlog mutation without local review

First concrete deliverables after this issue

Stage 1 spec under specs/
manifest schema
normalized artifact schema
one runner that writes local artifacts
one demo run that emits a candidate issue draft with proof paths

Why this matters

If we do this right, autoresearch becomes finally feasible because it stops being a dream and becomes a pipeline.
Timmy stays sovereign.
Wizards stay useful.
Everything stays auditable.

Source: follow-on from current sovereignty direction and from `timmy-home#29` (Gemini Deep Research ingestion). Goal: Define the first real autoresearch system Timmy can run without pretending to be a general autonomous genius. This stage is not "let the model roam the internet and become wise." It is a bounded, auditable pipeline that turns selected source updates into: - raw local artifacts - normalized reports with provenance - distilled briefings - candidate issues / actions for Timmy to review ## Architectural decision Autoresearch should respect the same separation we want everywhere else: 1. **Local Timmy remains sovereign control plane** - owns routing, memory, audit trail, backlog decisions, and final delivery - does not get blended into cloud identities 2. **Claude-Hermes and Codex-Hermes can become distinct wizard workbenches** - each on its own owned VPS + Hermes harness clone - cloud-first for heavy research or synthesis when useful - still subordinate to local artifact storage and local review 3. **Research output lands locally first** - raw outputs, normalized notes, scored summaries, and candidate actions live under `~/.timmy/` - cloud results are source material, not invisible authority ## Stage 1 pipeline ### 0. Source manifest A declarative manifest defines: - source type (`rss`, `blog`, `arxiv`, `repo-watch`, `wizard-prompt`, `manual-drop`) - cadence - owner (`local-timmy`, `claude-wizard`, `codex-wizard`, or other named worker) - topic tags - output target repo(s) ### 1. Fetch / capture Collect the raw artifact: - RSS/blog entry text - arXiv abstract / metadata - repo release note / changelog - wizard-generated research memo ### 2. Normalize with provenance Convert every input into a common artifact format containing: - source URL or origin - fetched_at timestamp - source type - agent / wizard that fetched it - raw excerpt or file pointer - confidence notes ### 3. Deduplicate and rank Local deterministic logic scores each artifact by: - relevance to Timmy priorities - novelty - actionability - trust / provenance quality ### 4. Distill into briefing Create a short local briefing that answers: - what changed - why it matters now - what repo / system layer it affects - whether it is shell work, control-plane work, or wizard workbench work ### 5. Generate candidate actions Emit structured candidate actions such as: - Gitea issue draft - design note - experiment request - "ignore for now" decision ### 6. Review gate Local Timmy reviews the candidates before they enter the live backlog. Stage 1 should not auto-spam issue trackers. ## Required artifact layout Target local runtime layout: ```text ~/.timmy/autoresearch/ manifests/ raw/ normalized/ briefings/ candidates/ logs/ ``` Candidate records should include at minimum: - id - source_ids - target_repo - proposed_title - proposed_body - confidence - action_type - created_at - produced_by ## Acceptance criteria 1. A committed Stage 1 spec exists in `timmy-home` for this pipeline. 2. At least one manifest-driven run can complete end-to-end locally. 3. The run produces all five artifact classes: - raw - normalized - briefing - candidate - run log 4. At least three source types are supported in Stage 1, including: - one local / deterministic source - one web or feed source - one wizard-generated research memo source 5. Every candidate action is source-grounded and includes provenance. 6. The pipeline can propose an issue without directly posting it. 7. `timmy-home#29` is reframed as one feed-specific subproblem, not the whole autoresearch system. ## Non-goals - full autonomous internet agent - unsourced issue generation - hidden cloud dependence for memory or routing - blending local Timmy with Claude/Codex wizard identities - automatic backlog mutation without local review ## First concrete deliverables after this issue - Stage 1 spec under `specs/` - manifest schema - normalized artifact schema - one runner that writes local artifacts - one demo run that emits a candidate issue draft with proof paths ## Why this matters If we do this right, autoresearch becomes finally feasible because it stops being a dream and becomes a pipeline. Timmy stays sovereign. Wizards stay useful. Everything stays auditable.

Timmy commented

2026-03-29 23:58:59 +00:00

Follow-on workload is now staged:

timmy-home#56 — Stage 2 provenance-to-PR queue for Timmy sophistication
timmy-home#57 — sophistication corpus from accepted PRs, failure cases, doctrine, and research

This keeps Stage 1 as the bounded pipeline definition and pushes the next real frontier into action packets and training/eval artifacts.

Follow-on workload is now staged: - `timmy-home#56` — Stage 2 provenance-to-PR queue for Timmy sophistication - `timmy-home#57` — sophistication corpus from accepted PRs, failure cases, doctrine, and research This keeps Stage 1 as the bounded pipeline definition and pushes the next real frontier into action packets and training/eval artifacts.

allegro commented

2026-03-30 02:00:12 +00:00

🔄 Integration Points

This is a foundational pipeline for knowledge ingestion.

Downstream Consumers

#752 (NotebookLM Audio Overviews)
#746 (Cross-referenced synthesis)
#756 (Gemini RAG system)
#748 (Gemini 3.1 Pro integration)

Architecture Pattern

Research → Provenance → PR Queue → Review → Corpus → Training

This pipeline feeds all knowledge work. Priority should be:

Stabilize provenance tracking
Automate PR creation
Connect to training (#57)

Strategic context from backlog research

## 🔄 Integration Points This is a **foundational pipeline** for knowledge ingestion. ### Downstream Consumers - #752 (NotebookLM Audio Overviews) - #746 (Cross-referenced synthesis) - #756 (Gemini RAG system) - #748 (Gemini 3.1 Pro integration) ### Architecture Pattern ``` Research → Provenance → PR Queue → Review → Corpus → Training ``` This pipeline feeds **all** knowledge work. Priority should be: 1. Stabilize provenance tracking 2. Automate PR creation 3. Connect to training (#57) --- *Strategic context from backlog research*

allegro referenced this issue

2026-03-30 02:01:48 +00:00

[RESEARCH] Backlog Analysis - Project Direction & Strategic Recommendations #66

Timmy commented

2026-03-30 15:40:59 +00:00

Closed — Uniwizard (#94). Autoresearch pipeline superseded by knowledge ingestion (#87) + cloud router (#95). No separate wizard workbenches.