[depalmar]

Stacked on #74. Base is docs/refresh-stale-claims, because the CLAUDE.md count edit here builds on that branch. Merge #74 first and this retargets to main automatically.

Why this entry

The host sweep flagged AIRT-0003 as the catalog's best correction candidate: ~/.copilot/ and config.json HIT while three high-confidence paths MISSED on the same machine — the pattern CLAUDE.md names as the query that finds real errors.

Installed Copilot CLI 1.0.80 (npm install -g @github/copilot) to settle it. Not authenticated, so nothing here depends on a signed-in session and the two credential directories were never created or inspected.

Result: no correction needed

Path	Basis
~/.copilot/mcp-config.json	copilot mcp --help prints User ~/.copilot/mcp-config.json verbatim; 4 occurrences in the shipped package
~/.copilot/mcp-secrets/	3 occurrences in the shipped package
~/.copilot/mcp-oauth-config/	3 occurrences in the shipped package

The vendor's own binary is the source — not a search summary, not an aggregator — so this clears the last_verified bar.

They're created on first MCP server add, so an absence on a host that has the CLI is expected, not suspicious. That now lives in the mcp block rather than waiting for the next sweep to rediscover it.

One caveat that changes what the directory proves

The install wrote nothing into ~/.copilot. The directory already existed — empty ide/, one log — created by the IDE integration on a host where the CLI had never been installed.

~/.copilot/ presence is not evidence the CLI is present. A responder reading it that way would be wrong, using this catalog as written.

Rows added

Reading the package turned up more than the verification did — 5 disk rows, 2 MCP mechanisms:

• <repo>/.mcp.json and <repo>/.github/mcp.json — workspace MCP scope, vendor-documented, previously uncatalogued. Travels in the repository, so a clone can introduce a server the user never configured.
• plugins/computer-use/ — computer-use-mcp.exe (3.4 MB), CopilotComputerUse.exe (2.4 MB), and a .mcp.json registering it as a stdio server. Vendor-shipped, so a supply-chain surface rather than a user-configuration surface.
• copilot.exe, and the bundled voice engine with pvrecorder.

Capability change

screen_capture: true. Basis is the binary: 52 occurrences of screenshot, SendInput and UIAutomation for control, and strings describing an engine that "cannot perceive or act" without accessibility permission.

A tool catalogued as coding-agent-ide ships desktop observation and input synthesis by default, with no separate install.

Two changes declined:

• browser_control — the engine drives a browser the way it drives any window; the flag denotes programmatic automation. Setting it would assert a mechanism the tool doesn't have.
• Recategorising to computer-use-agent — computer use is a bundled plugin, not the tool's identity, and it would skew the category counts.

Scope limit

Every new row is windows only, stamped 2026-08-21. The plugin path embeds copilot-win32-x64 and the binaries are .exe. The binary contains macOS strings suggesting the plugin ships there too — that's a lead, not a verified row, and it's left unstamped.

The two credential directories remain unverified for mode and permissions. That needs an authenticated session.

Verification

validate.py            51 entries + 12 sigma rules, 0 problems
readme_counts.py       counts are current
build_site.py --check  622 rows, unique anchors, 0 problems
validate_mappings.py   68 on disk = 68 in MAPPINGS, in sync
validate_techniques.py 0 problems

All five feeds regenerated in-commit, including gen_credential_targets.py. Counts 615 → 622; disk 207 → 212, mcp-config 64 → 66.

Full reasoning, including what was declined, is in artifacts/docs/VERIFICATION.md.

🤖 Generated with Claude Code