Tiinex Traceable Provenance for VS Code

This package is the VS Code-specific extension surface for the ai-provenance repo.

It intentionally lives under ides/vscode because the repo itself is broader than one IDE.

Marketplace Release

0.2.0 is the first Marketplace-oriented minor release for the provenance-side TRACEABLE package.

It is aimed at operators who need bounded provenance work in VS Code rather than broad autonomous orchestration.

Install path:

1. open the VS Code Extensions view
2. search for Tiinex Traceable Provenance
3. install the extension published by tiinex

Release highlights in 0.2.0:

• bounded TRACEABLE child-lane execution through #runTraceableSubagent
• bounded role and model discovery through #listTraceableAgents and #listTraceableModels
• reconstructed .trace.md evidence inspection through #viewTraceableSubagent
• provenance-owned TRACEABLE viewer UX with reopen-to-source and reopen-to-markdown-preview flows
• persisted runtime-decision visibility for model selection and request routing in evidence artifacts

Quick Start

After you install the extension in VS Code, the shortest useful first flow is:

1. type # in chat and pick #listTraceableAgents or #listTraceableModels
2. run one narrow lane with #runTraceableSubagent
3. inspect the returned evidence file with #viewTraceableSubagent

The important operator detail is that chat usually exposes the prompt-reference names after #, not the raw internal tool ids. In practice you should expect #listTraceableAgents, not list_traceable_agents.

What To Expect

This package is built for bounded provenance-first TRACEABLE work.

• It helps you preflight roles and models before a run.
• It can export and reopen .trace.md evidence artifacts when the lane requests exportToFolder or when the user explicitly chooses export.
• It is designed for narrow investigation slices, not broad autonomous orchestration.
• In chat, the first-class invocation surface is usually # plus the prompt reference name.

Marketplace Fit

This extension is a good fit when you want to:

• run a bounded TRACEABLE lane against one narrow investigative slice
• export .trace.md evidence and inspect it without rerunning the child lane
• inspect runtime decisions, request contracts, lineage, and outcome surfaces from one evidence artifact
• keep provenance-oriented tooling separate from Local-chat session-store and delete tooling

This extension is not trying to be:

• a general autonomous agent framework
• a replacement for ai-vscode-tools Local-chat/store workflows
• a hidden tool-routing layer with opaque evidence output

Current status:

• buildable as a real VS Code extension
• ready for local main-host junction linking on Windows
• now carries the provenance-side TRACEABLE tool surface: list_traceable_agents, list_traceable_models, view_traceable_subagent, and run_traceable_subagent
• now also carries the reconstructed .trace.md evidence viewer UX with source/preview reopen commands on the provenance side
• now also carries the first host-independent TRACEABLE contract slice: request/result, request-envelope, payload extraction, result construction, full markdown rendering, and evidence-related types
• now carries release-check, VSIX packaging, and semantic-version scripts for Marketplace-oriented delivery

Current included surface:

• Tiinex: Inspect TRACEABLE Evidence parses the embedded Traceable State block from a .trace.md file and lets you choose a bounded surface without rerunning the child lane
• Open Reconstructed Traceable View opens a provenance-owned reconstructed viewer for a .trace.md artifact and can reopen back into source or markdown preview
• list_traceable_agents exposes the bounded workspace-supported traceable agent catalog from the provenance side
• list_traceable_models exposes the bounded runtime-discoverable traceable model catalog from the provenance side
• run_traceable_subagent runs the provenance-owned TRACEABLE child-lane runtime with optional evidence export support
• current bounded surfaces: rendered-output, request-summary, summary, outcome, tool-ledger, status-history, tool-summary, file-summary, and state-json
• a separate provenance LM tool namespace is now present through list_traceable_agents, list_traceable_models, view_traceable_subagent, and run_traceable_subagent
• provenance-specific settings now live under tiinex.aiProvenance.*

What it exposes in VS Code:

• display name: Tiinex Traceable Provenance
• LM tool surfaces: list_traceable_agents, list_traceable_models, view_traceable_subagent, run_traceable_subagent
• command namespace: tiinex.aiProvenance.*
• settings namespace: tiinex.aiProvenance.*
• TRACEABLE panel/status shell under the provenance namespace

Canonical tool usage:

Canonical prompt references in chat:

• #listTraceableAgents: use this first when you want a grounded role-backed run; copy the exact returned display name or file path into run_traceable_subagent.agentRole instead of guessing a role label.
• #listTraceableModels: use this first when you need explicit model control; prefer sendableOnly: true, narrow with query when useful, and treat entries marked Policy: blocked as non-selectable for run_traceable_subagent.
• #runTraceableSubagent: choose the input mode deliberately. OPERATIVE, EPISTEMIC, and NON_LEADING_EPISTEMIC use the classic userInput plus parentTask form. DIRECT uses only userInput while still allowing lineage and runtime overrides. RESUME requires parentTracePath and resumes without any fresh userInput, parentTask, or parentFrame.
• #viewTraceableSubagent: after a run returns an evidence file, inspect that artifact before rerunning the child lane; start with summary or outcome, then use tool-ledger or state-json only when deeper debugging is needed.

Input mode quick guide:

• OPERATIVE: requires userInput and parentTask; use this for bounded operational delegation, with optional parentTracePath for continuation and handover.
• EPISTEMIC: requires userInput and parentTask; use this for inquiry-shaped delegation where the parent still carries the bounded task contract.
• NON_LEADING_EPISTEMIC: requires userInput and parentTask; use this when the child should preserve a non-leading investigative contract and surface input-mode validation explicitly.
• DIRECT: requires only userInput; use this for a live-chat-like fresh turn, optionally with parentTracePath, but without inheriting or injecting parentTask or parentFrame.
• RESUME: requires parentTracePath; use this for strict prompt-free continuation. Do not pass fresh userInput, parentTask, or parentFrame here.

Canonical examples:

• Role-grounded preflight flow: #listTraceableAgents -> #runTraceableSubagent with agentRole -> #viewTraceableSubagent on the returned evidence file.
• Model-grounded preflight flow: #listTraceableModels -> copy one allowed exact model id -> #runTraceableSubagent with modelSelector.id -> #viewTraceableSubagent on the returned evidence file.
• Recovery flow: if a run already produced .trace.md through exportToFolder or explicit export, inspect it with #viewTraceableSubagent before launching another lane.

Example payloads:

• OPERATIVE:

    {
    	"inputMode": "OPERATIVE",
    	"userInput": "Read README.md and summarize the current validation gap.",
    	"parentTask": "Produce a bounded operational summary grounded in the named file.",
    	"allowedToolNames": ["copilot_readFile"],
    	"budgetPolicy": { "maxIterations": 2, "maxToolCalls": 2 }
    }
  

• DIRECT:

    {
    	"inputMode": "DIRECT",
    	"userInput": "What changed in the last trace and what should I inspect next?",
    	"parentTracePath": "ai-provenance/.topics/m3-lineage-chain/01-anchor.trace.md",
    	"modelSelector": { "id": "copilot/gpt-5-mini" }
    }
  

• RESUME:

    {
    	"inputMode": "RESUME",
    	"parentTracePath": "ai-provenance/.topics/m3-lineage-chain/01-anchor.trace.md",
    	"allowedToolNames": ["copilot_readFile", "view_traceable_subagent"],
    	"reveal": true
    }
  

The DIRECT and RESUME examples above omit budgetPolicy intentionally so the child sees live-like conditions and any undeclared runtime fail-safe stays internal. Add budgetPolicy only when you want the child to treat that budget as part of the explicit request contract.

When budgetPolicy is omitted, TRACEABLE falls back to the hidden runtime fail-safe settings tiinex.aiProvenance.traceableUndeclaredMaxIterations and tiinex.aiProvenance.traceableUndeclaredMaxToolCalls instead of surfacing a synthesized default budget to the child.

Local development loop:

• This section is for extension contributors rather than Marketplace-first users.
• npm test builds and runs the current validation slice
• npm run package:vsix produces a local VSIX for install testing
• npm run release:check is the release gate used before publishing

Release flow:

• npm test
• npm run package:vsix
• npm run release:check
• npm run release:patch, npm run release:minor, npm run release:major
• npm run publish:vsce

Non-goal for this package scaffold:

• no MCP server surface
• no extra agent runtime surface
• no claim of native runSubagent UX parity or of broader host-private agent enumeration beyond the bounded provenance traceable surfaces