Skip to content

fix(global-settings): drop dead mcpServers block, document MCP scopes (v2.4.1) - #700

Merged
WilcoLouwerse merged 2 commits into
mainfrom
hotfix/global-settings-drop-mcpservers
Sep 8, 2026
Merged

WilcoLouwerse merged 2 commits into
mainfrom
hotfix/global-settings-drop-mcpservers

Conversation

@WilcoLouwerse

Copy link
Copy Markdown
Contributor

Why

Claude Code does not read an mcpServers key from settings.json — see Debug your configuration → Check common causes: "settings.json does not read an mcpServers key. Define project servers in .mcp.json at the repository root, or run claude mcp add --scope user for user-scoped servers."

The 7-browser mcpServers block in global-settings/settings.json therefore never loaded anything. Verified on 2026-09-07: with the project .mcp.json temporarily disabled, claude mcp list showed zero browser servers even though this block was installed. The browsers have always come from the workspace's .mcp.json.

What

  • global-settings/settings.json — remove the inert mcpServers block (30 lines, no behaviour change)
  • global-settings/VERSION — 2.4.0 → 2.4.1 so installed copies get the update notice
  • docs/claude/global-claude-settings.md — §9 now explains the two supported scopes (project .mcp.json / user claude mcp add --scope user) and the file table no longer lists "optional mcpServers"
  • docs/claude/playwright-setup.md
    • new User scope (all projects on this machine) section with the claude mcp add --scope user recipe for all 7 browsers
    • the hydra/.mcp.json symlink convention (Hydra ships the pool in ConductionNL/hydra; workspaces that symlink its .claude/skills symlink .mcp.json too)
    • corrected pre-approval paragraph: no enableAllProjectMcpServers or mcp__browser-* allow-list is required — background subagents surface their permission prompts in the main session, and Hydra deliberately does not commit .claude/settings.json. The old "silently denied" claim is no longer accurate.

Verification

  • python3 -c 'json.load(...)' on the new settings.json → valid
  • /verify-global-settings-version → Case B, bump 2.4.0 → 2.4.1 ✅
  • Companion PR: ConductionNL/hydra hotfix/mcp-browsers-config (ships .mcp.json)

🤖 Generated with Claude Code

… (v2.4.1)

Claude Code does not read an `mcpServers` key from settings.json
(code.claude.com/docs/en/debug-your-config#check-common-causes), so the
7-browser block shipped since the first release never loaded anything.
Verified 2026-09-07: with the project .mcp.json disabled, `claude mcp list`
showed zero browser servers despite this block being installed.

- global-settings/settings.json: remove the inert `mcpServers` block
- global-settings/VERSION: 2.4.0 -> 2.4.1
- docs/claude/global-claude-settings.md: §9 now explains the two supported
  scopes (project `.mcp.json`, user `claude mcp add --scope user`) and
  drops the "optional mcpServers" mention from the file table
- docs/claude/playwright-setup.md: add the user-scope recipe, the
  hydra/.mcp.json symlink convention, and correct the pre-approval
  paragraph — no `enableAllProjectMcpServers` or `mcp__browser-*`
  allow-list is required; background subagents surface their prompts in
  the main session (code.claude.com/docs/en/sub-agents), and hydra
  deliberately does not commit `.claude/settings.json`

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

@WilcoLouwerse WilcoLouwerse left a comment

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Verdict: APPROVE (Quick) — self-review posted as COMMENT (GitHub blocks self-APPROVE)

Small, well-scoped hotfix — verified locally against the hotfix/global-settings-drop-mcpservers checkout in /home/wilco/.github:

  • global-settings/settings.json parses as valid JSON after the mcpServers block removal; no leftover trailing-comma or structural issues.
  • global-settings/VERSION bumped 2.4.0 → 2.4.1, matches the ^[0-9]+\.[0-9]+\.[0-9]+$ format check-settings-version.sh expects.
  • No dangling mcpServers references — the three remaining hits (playwright-setup.md, .mcp.json.example, and explanatory prose in global-claude-settings.md §9) are all intentional (the .mcp.json example and doc text describing where it does belong).
  • The new anchor link playwright-setup.md#user-scope-all-projects-on-this-machine resolves correctly against the new ## User scope (all projects on this machine) heading.
  • The for i in 1 2 3 4 5 7; do … done + separate browser-6 loop in the new user-scope recipe correctly mirrors the original block's headless/headed split (browser-6 was the only headed instance).
  • No CI failures on the current head SHA; the few pending/queued checks are unrelated to this docs+config change and none are in a failed state.

No findings to post — zero inline comments this pass.

#698 shipped check_flow_node_taxonomy.py, its three-arm fixture tree, its
unit suite and an expect.conf declaring gate 111 — but never added a call
site to run-hydra-gates.sh. The acceptance matrix ran the runner over the
planted and clean arms, found no `[gate-111]` verdict line, and reported
"emitted NO verdict line at all" for both. `main` has been red on the
"Package invariants" check since f808893; this PR only inherited it,
because the hydra-gates workflow deliberately carries no `paths:` filter
and so runs on every PR to main.

The unit suite passed throughout, because it invokes the checker directly
rather than through the runner — a gate can be fully implemented and fully
unit-tested and still not exist for any repository the runner is pointed at.

Wiring follows gate-83's exit-code shape (0 pass / 4 not applicable /
crash-without-terminal-summary as `wiring`), with one deliberate difference:
no `[ -d ... ]` guard around the call. The checker reports NOT APPLICABLE
itself, so repos without lib/Service/Flow/Nodes/ get a verdict line instead
of the silence that hid this gate in the first place.

Verified locally:
- test_gate_acceptance_matrix.sh    244 passed / 0 failed (was 241 / 2)
- test_gate111_flow_node_taxonomy.sh  all three arms behaved
- check_gate_numbers_unique.sh      91 gates, every number claimed once
- errexit-discipline, exit-code-semantics, empty-scope suites all pass

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
@WilcoLouwerse
WilcoLouwerse merged commit d2dad1f into main Sep 8, 2026
41 checks passed
@WilcoLouwerse
WilcoLouwerse deleted the hotfix/global-settings-drop-mcpservers branch September 8, 2026 09:45
rubenvdlinde added a commit that referenced this pull request Sep 8, 2026
…ed (#710)

#706 and #700 fixed the same defect, and both landed.

The sequence: at 29c827e gate-111 was unregistered, which is what I measured
and what #706 was written against. #700 then merged 50ee1ff, "wire gate-111
into the runner so it emits a verdict", which fixed it. #706 merged on top and
added a second block. I did not re-check between opening the PR and merging
it, and that is the whole cause.

Two blocks reporting one gate number is the exact failure check_gate_numbers_
unique.sh exists to prevent: one verdict overwrites the other in the report,
and the COVERAGE tally counts one slot for two gates, so a gate that never ran
is indistinguishable from one that passed. (The uniqueness check reads
`_pass|_fail|_skip <n>` occurrences rather than block headers, so it counted
91 gates and stayed green over the duplicate.)

This removes MY block and keeps 50ee1ff's, which landed first and is the
better of the two: it routes the crash check through the shared
`_helper_finished` helper rather than open-coding the grep.

A pure deletion -- 65 lines removed, zero added.

Verified: gate-111 registered once, `check_gate_numbers_unique.sh` reports 91
gates every number claimed once, `bash -n` clean, and
`test_gate_acceptance_matrix.sh` stays at 244 passed / 0 failed with gate-111
still FAILing on planted/ naming PlantedNode and PASSing on clean/.

Co-authored-by: Conduction Release Bot <release-bot@conduction.nl>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant