Status (2026-10-11)
Partly delivered by the source-control work (#535): the pull request tools and the linking instruction block that Tcode injects are shown per thread under "What the agent sees" (desktop Pull requests tab and the phone sheet, #658, #689), the linking block is prepended visibly to every turn rather than hidden (#652), and a thread keeps the tool descriptions of the registration the model actually has (#683). Still open: the general inventory of every injection source (project instructions, tool definitions, MCP context, skills, runtime-added messages), the persisted "Show automatically supplied context" setting, and optional translation.
Status (2026-09-20)
This is a direct consequence of Principle 1: anything Tcode injects into a model must be visible to the user. The Show automatically supplied context setting defaulting to on is decided, not proposed. The constraints in the proposal below (exact payload vs discovered metadata, labelled translations that never replace the real model input, no credentials from launch environments) stand.
Proposal
Problem and existing behavior
Users cannot inspect all the context that tcode or its harness automatically supplies to a model. The existing “Orchestrate skill” disclosure is a useful presentation precedent: the send path records an injected prefix boundary, and ChatView renders that prefix separately from the user's message. This is currently a specific orchestration path, not a general inventory of automatically supplied context.
Proposed behavior
- Add a persisted Show automatically supplied context setting, enabled by default. Render context as expandable ChatView entries using the existing disclosure language.
- Inventory every observable injection source across tcode and provider adapters. AGENTS.md or equivalent project instructions, tool definitions, MCP server/tool context, skills, orchestration guidance and runtime-added messages are examples, not an exhaustive list. Include later injections and changes, not only the first send.
- Identify the source, when it became effective, and whether the displayed content is the exact supplied payload or only discovered metadata. Do not present a list of installed tools/skills as proof that all their contents were injected. Clearly represent context the harness does not expose.
- Keep the original content inspectable and add optional on-demand translation into the user's UI language through a configurable lower-cost model. Label translations, preserve the original, and never replace the actual model input with its translated display.
- Keep provider payloads and context records owned by the agent/runtime layers; UI owns localized presentation. Do not expose credentials from launch environments as context.
Acceptance criteria
The setting defaults on and survives restart. A first turn and a later context change show accurate, attributed disclosures; reopening a conversation preserves available records. Provider-specific visibility gaps are explicit. Translating, collapsing or hiding a disclosure does not resend or alter the task, change title generation, or count the same injected content twice. Validate the inventory with representative provider event traces and both UI languages.
Code context
Reviewed against main at b4c549787da6 (2026-09-08).
Status (2026-10-11)
Partly delivered by the source-control work (#535): the pull request tools and the linking instruction block that Tcode injects are shown per thread under "What the agent sees" (desktop Pull requests tab and the phone sheet, #658, #689), the linking block is prepended visibly to every turn rather than hidden (#652), and a thread keeps the tool descriptions of the registration the model actually has (#683). Still open: the general inventory of every injection source (project instructions, tool definitions, MCP context, skills, runtime-added messages), the persisted "Show automatically supplied context" setting, and optional translation.
Status (2026-09-20)
This is a direct consequence of Principle 1: anything Tcode injects into a model must be visible to the user. The Show automatically supplied context setting defaulting to on is decided, not proposed. The constraints in the proposal below (exact payload vs discovered metadata, labelled translations that never replace the real model input, no credentials from launch environments) stand.
Proposal
Problem and existing behavior
Users cannot inspect all the context that tcode or its harness automatically supplies to a model. The existing “Orchestrate skill” disclosure is a useful presentation precedent: the send path records an injected prefix boundary, and ChatView renders that prefix separately from the user's message. This is currently a specific orchestration path, not a general inventory of automatically supplied context.
Proposed behavior
Acceptance criteria
The setting defaults on and survives restart. A first turn and a later context change show accurate, attributed disclosures; reopening a conversation preserves available records. Provider-specific visibility gaps are explicit. Translating, collapsing or hiding a disclosure does not resend or alter the task, change title generation, or count the same injected content twice. Validate the inventory with representative provider event traces and both UI languages.
Code context
Reviewed against
mainatb4c549787da6(2026-09-08).