Instruction file imported from fabioc-aloha/circuit-breaker (
.github/instructions/tool-awareness.instructions.md). Copyright stays with the author.
Tool Awareness
Always-on rationale: deferred-tool resolution and external-ingest awareness apply to any tool-using turn regardless of file context. The search before calling discipline must fire on every deferred-tool need; scoping by file pattern would silence the protection exactly when the agent needs an unfamiliar tool.
Deferred Tools (VS Code 1.118+)
Many tools are deferred (lazy-loaded). They appear in availableDeferredTools but cannot be called directly. Load via tool_search first with a natural-language capability description.
Rules
- Search before calling. Calling a deferred tool without loading via
tool_searchfails silently. - Search once per tool. After load, the tool stays available for the session.
- Use broad queries. One broad search beats multiple narrow ones.
- No results means unavailable. Don't retry with synonyms.
For common deferred tool categories and search-query patterns, see tool-awareness-categories.instructions.md (scoped, loads on tool/MCP/GitHub work).
External Ingest (VS Code 1.119+)
In remote or virtual-filesystem workspaces (GitHub.dev, VS Code Remote, Codespaces), the editor provides codebase context automatically. semantic_search and file operations work transparently — no agent action needed.
VS Code 1.122–1.128 conveniences
| Release | Capability | What it changes for me |
|---|---|---|
| 1.122 | /models slash command |
Opens the model picker from chat input. Useful when the user asks to switch models mid-task without leaving chat. |
| 1.122 | BYOK air-gapped | Bring Your Own Key models work without GitHub authentication. Heirs in regulated/enterprise contexts can run Copilot Chat fully offline; the BYOK token counter (introduced 1.120) keeps working. |
| 1.122 | Local agent host default-on (Insiders only) | Watchpoint: when this reaches Stable, deferred-tool resolution may shift. No action until first observed behavior change. |
| 1.123 | Session sync + /chronicle |
Chat sessions auto-sync to the GitHub account (gated by chat.sessionSync.enabled, org-managed). The /chronicle slash command queries that history. The brain's own chronicle skill is local-only; the platform feature is an adjacent capability, not a replacement. |
| 1.123 | Sandbox network-retry | When a local-agent terminal command needs an unallowed domain, VS Code auto-retries inside the sandbox with unrestricted network before falling back to unsandboxed (chat.agent.sandbox.retryWithAllowNetworkRequests). Reduces spurious failures on git fetch / npm install. |
| 1.124 | Autopilot enabled by default | Autopilot Preview is now on by default; chat.permissions.default controls the per-workspace level. ACT's heir-workspace baseline pins default as the deliberate opt-out — see heir-workspace-settings-baseline.json. |
| 1.124 | Advanced Autopilot (opt-in) | chat.autopilot.advanced.enabled uses a utility model to judge when a task is truly done. Capped at 3 iterations. Off by default; opt-in only. |
| 1.124 | Enterprise Copilot plugin policies | chat.plugins.enabledPlugins / chat.plugins.extraMarketplaces / chat.plugins.strictMarketplaces let org admins allowlist plugin IDs and marketplaces. Heirs in regulated orgs may see Mall installs silently blocked or marketplaces tagged as policy-managed; surface a clear message rather than retrying when an install refuses. |
| 1.125 | extensions.autoUpdate simplified to on / off |
Old values (true / false / onlyEnabledExtensions / delayed) migrate automatically. Edition welcome-baseline.json pins the canonical on shape from this release forward; pre-1.125 heirs that have the old value still work via migration. |
| 1.125 | extensions.autoUpdateDelay configurable |
The 2-hour supply-chain delay introduced in 1.123 is now a configurable hour count. Edition does not pin a value — heirs ride the platform default — but the setting exists if a heir or org wants tighter / looser update windows. |
| 1.125 | Forwarded-port URL rewriting for agents | When an agent in a remote workspace requests a port that has been forwarded, VS Code rewrites the URL and notifies the agent of the change. Reduces spurious browser-tool failures in Codespaces / Remote-SSH heir setups. No action needed; surface in fetch/browser diagnostics if a heir reports a port mismatch. |
| 1.125 | Native MDM delivery for managed Copilot settings | On Windows/macOS, org admins can deliver Copilot settings via MDM channels in addition to the account-based enterprise file. Settings delivered via MDM appear as policy-enforced and cannot be overridden locally — heirs in MDM-managed orgs may see baseline welcome-baseline.json keys ignored if the MDM policy disagrees. |
| 1.126 | Edit mode deprecated → Agent mode | chat.editMode.hidden also removed. Reinforces welcome-baseline chat.agent.enabled: true pin — user policy and platform default converged. Heirs with agent mode policy-disabled still see legacy Edit mode. |
| 1.127 | macOS/Linux terminal sandboxing | Agent-invoked terminal commands run with network blocked + FS restricted; agent only prompts on elevation. Substantially reduces approval prompts on non-Windows heirs. Toggle in Permissions dropdown. Windows unaffected — Backtick Hazard + Output Capture rules still apply. |
| 1.127 | /troubleshoot on agent host sessions |
Adjacent to the brain's ACT: Diagnose Fetch pattern — platform-side diagnostic for agent behavior questions (custom instructions ignored, slow responses); not a replacement for our fetch diagnostics. |
| 1.127 | Browser tools GA (workbench.browser.enableChatTools) |
Agent can open pages, screenshot, click through to validate its own work. Org-managed setting. Enterprise policies BrowserChatTools + ChatAgentNetworkFilter (agent domain allow/deny lists) may block Mall/GitHub fetches for regulated heirs — surface a clear message rather than retrying when a fetch refuses. |
| 1.127 | File-based managed Copilot settings | Extends 1.125 MDM channel: managed-settings.json at %ProgramFiles%\GitHubCopilot\ (Win) / /Library/Application Support/GitHubCopilot/ (mac) / /etc/github-copilot/ (Linux). Same policy-lockout semantics as the MDM row — baseline keys may be ignored in policy-managed environments. |
| 1.127 | Built-in Ollama provider deprecated | Official Ollama VS Code extension replaces it. Heirs using BYOK Ollama should install the extension and remove the built-in provider. No brain action; documents why welcome-baseline.json doesn't pin an Ollama provider. |
| 1.128 | BYOK Claude via own Anthropic key | Heirs can run the Claude agent through their own Anthropic API credentials instead of consuming GitHub Copilot quota. Relevant to the Model Compatibility credit-economy discussion in Edition README. No baseline setting; user-opt-in via Claude harness auth. |
| 1.128 | Custom endpoint model options for BYOK | Enables BYOK against strict-schema providers (Moonshot, Kimi, etc.) that reject non-standard params. temperature and provider-specific options now configurable. Unblocks heirs who reported provider-rejection errors before this. |
| 1.128 | Claude agent → integrated browser DOM/tools | Feature parity with the Copilot agent's browser tools GA (1.127). Same BrowserChatTools + ChatAgentNetworkFilter enterprise policies apply. |
Skill Picker Surfacing (VS Code 1.118+)
In 1.118+, .github/skills/<name>/SKILL.md files with a non-empty description in their frontmatter ALSO surface in the chat slash-command picker (alongside .github/prompts/*.prompt.md). Controlled by the experimental setting github.copilot.chat.skillTool.enabled (default on).
Consequence for the brain
When a prompt and a skill share a base name (/meditate prompt + meditation skill), the picker shows both. This is not a brain defect — the verb-prompt / noun-skill pairing is intentional (prompts are workflow entry points, skills are knowledge bodies). The picker noise is a side effect of the platform surface postdating the brain's design.
Lever, not stripping
If picker noise is the problem, the lever is the user-level setting:
// settings.json (user scope)
"github.copilot.chat.skillTool.enabled": falseNever strip the SKILL.md description to declutter the picker. The description field has three consumers and the picker is the least important of them:
- Agent skill discovery (primary) — every session loads SKILL.md descriptions into the
<skills>block; this is how the parent agent decides whether to invoke the skill - Brain QA enforcement — where a brain-qa script exists (Supervisor ships one as
scripts/brain-qa.cjs), it hard-fails on missing/empty description - Chat picker tooltip — the surface visible to humans
Stripping (1) and (2) to fix (3) is a Type III error (right cost, wrong problem). The setting is the right scope.
Would Revise If
Revise if VS Code changes the deferred-tool mechanism (e.g. tool_search semantics change, deferred tools become directly callable, or external-ingest changes scope in remote workspaces), or if the "search before calling" rule produces no observed failures over a quarter (the rule is no longer load-bearing because the platform changed).
Skill picker section falsifier: revise by 2026-08-24 (90 days) or sooner if any of the following fires: (a) VS Code renames or removes github.copilot.chat.skillTool.enabled; (b) setting the flag to false does not reduce skill-name entries in the slash picker; (c) the brain restructures SKILL.md frontmatter such that description ceases to be the agent-discovery signal. First observed contradiction wins.