Investigate
Gather deterministic project evidence before making claims.
PROJECTFORGE · EARLY DEVELOPMENT
An AI engineering partner built around evidence and control.
ProjectForge is still in its infancy. It is being built as a local-first engineering assistant that can inspect a real software project, gather evidence, understand the relevant subsystem, propose bounded changes, and keep the person building the software in control of what actually happens.
LOCAL-FIRSTEXPLICIT APPROVALVERIFIED EXECUTION
CURRENT DIRECTION
Investigate → Understand → Assess → Propose → Approve → Execute → Verify
The AI reasons and proposes. Deterministic controls handle evidence, scope, authority, execution, and closure.
WHAT IT IS
Project-aware engineering instead of isolated prompts.
The goal is for ProjectForge to understand the project it is working on before it changes anything. It investigates the actual code and project state, follows verified relationships, and builds an evidence-backed view of the relevant subsystem.
It is designed for controlled engineering work: read first, reason from evidence, propose a bounded change, wait for explicit approval, execute only the authorized scope, and independently verify the result.
CONTROL MODEL
AI reasoning does not equal authority.
Propose
Define what should change, where, why, and how success will be verified.
Approve
The user explicitly authorizes a frozen, bounded scope.
Verify
Post-write verification independently checks reality before closure.
WEB OBSERVATORY
Observed evidence stays separate from AI interpretation.
ProjectForge can capture the rendered page the user is actually viewing. Browsing alone does not count as evidence. Capture begins only when the user explicitly asks ProjectForge to analyze the page, and the resulting observed evidence is kept distinct from the model's interpretation.
PROJECT README
Current engineering notes
ProjectForge v0.30.0 — Web Observatory
General-purpose observation-only webpage analysis for ProjectForge.
What this build adds
- New Web workspace in the main sidebar.
- Embedded Electron webview with persistent browser session so the user can navigate and sign in themselves.
- Explicit Analyze Page capture boundary. Merely browsing does not tell ProjectForge that it observed page contents.
- Evidence capture from the rendered webview: URL/title, visible text, headings, links, tables, image metadata, controls, and screenshot preview when available.
- SHA-256 over canonical captured evidence.
- Separate WEBPAGE OBSERVED and AI INTERPRETATION panels.
- Grounded web analysis through the local llama-server. If the model is offline, capture still works but interpretation is refused.
- Observation only: no AI click automation, form submission, item transfer, delete, purchase, or account-changing capability is introduced.
Security model
The webview runs as a guest; the normal ProjectForge renderer remains context-isolated and sandboxed. Password controls are excluded from captured form metadata, and input values are not captured.
Recommended test
- Start with npm run dev and verify identities align.
- Open Web.
- Browse to a public page and click Analyze Page.
- Confirm observed evidence includes URL, timestamp, SHA-256, capture counts, text, and screenshot evidence.
- With llama-server online, run grounded analysis.
- For authenticated apps, sign in manually first and then capture the rendered page.
v0.32–v0.39 — Grounded Engineering Control Pipeline
ProjectForge now has the validated foundation for a general evidence-backed software-engineering agent.
Investigate → Understand → Assess → Propose → Freeze Scope → Explicit Approval → Controlled Execution → Independent Verification → Close or Refuse Closure
- v0.32 — Grounded Investigation
- v0.33 — Verified Relationship Expansion
- v0.34 — Grounded Subsystem Understanding
- v0.35 — Grounded Engineering Assessment
- v0.36 — Grounded Change Proposal
- v0.37 — Frozen Scope + Explicit Approval Boundary
- v0.38 — Controlled Hands Scope Execution
- v0.39 — Post-write Verification / Closure
Authority model
The AI is not the authority. AI reasoning cannot grant itself permissions, manufacture evidence, expand approved scope, or claim successful execution without deterministic verification.
AI reasons and proposes → deterministic code verifies evidence and scope → the user explicitly approves → deterministic Controlled Hands executes only authorized work → deterministic post-write verification proves or rejects the result.
v0.40 — Persistent Engineering State & Audit Continuity
v0.40 makes the validated engineering-control chain durable across application restarts. Frozen scopes, execution authorizations/results, and post-write verification records are persisted in a deterministic engineering ledger rather than existing only in process memory.
Persistent control state
- Persists frozen scope records and their pending/approved/rejected states.
- Persists execution authorizations and whether they remain pending, were rejected, or were consumed.
- Persists post-write verification and closure records.
- Reconstructs engineering history after restart without creating new authority.
- Uses atomic temporary-file replacement when persisting the ledger.
Security property
Persistence remembers authority state; it does not create authority.