docs: six-month Inspector roadmap (Aug 2026 → Feb 2027)

[cliffhall]

Closes #1980

Adds docs/inspector-roadmap-2026-h2.md — a six-month plan covering 2026-08-11 → 2027-02-11 (~26 weekly milestones, v2.2.0 → ~v2.27.0).

One new file, nothing else touched.

Status: Draft for WG review. Merging this starts the discussion; it does not settle it. §7 carries six open questions I'd like answered before anyone treats the sequencing as committed.

Why

Through v1 the Inspector was a follow-along project: the spec moved, we chased it, and whatever planning capacity remained went to keeping up rather than to the tool's own design. v2 meeting the 2026-07-28 spec across Web, CLI, and TUI lifts that constraint — but nothing written down said so, and we had no shared list of what the MCP roadmap is likely to demand of us next. Without one, spec work keeps arriving as a surprise and getting a bespoke panel, and UX work never gets scheduled because it never competes for a milestone.

Structure

Two explicitly-budgeted tracks, targeted at a rough 50/50 capacity split.

Track A — spec-following. Predicted Inspector features per MCP roadmap theme and WG charter, each tagged 🟢 build now / 🟡 design now, build on signal / 🔴 watch — so a speculative item gets a WG liaison rather than code, which is the mistake v1 made repeatedly. Covers transports and session lifecycle, Server Cards (SEP-2127), Tasks retry/expiry and the extension→core migration, enterprise audit + ID-JAG, triggers/events, result types, interceptors (SEP-1763), file inputs (SEP-2356), skills (SEP-2640), primitive grouping, and conformance.

Track B — the work we choose. Headlined by the zoomable protocol/network timeline: lanes, spans rather than points, zoom/pan, brush-to-filter, MRTR and task grouping, a pinned mini-timeline strip. Then session record/replay/share, a shared diff primitive, a ⌘K palette, saved-call collections feeding CI assertions, the argument-editor workstream, Connection Doctor, workspace/layout, and the plugin architecture.

The sequencing argument

Several Track B items are force multipliers for Track A. Each new protocol feature arrives with a rendering problem, and one general timeline plus one general diff is cheaper than one bespoke panel per SEP — tasks become timeline spans for free, and the payload differ serves interceptor before/after, Server-Card-vs-reality, and stateless-instance comparison from one component. So Phase 1 builds the general surfaces and the later SEP work renders into them.

Two consolidations worth acting on regardless

These came out of surveying the open board and hold whether or not this plan is adopted as written:

• Fix JSON parameter editor escaping while typing #1853, Backspace recursively escapes JSON tool inputs #1856, JSON editor corrupts null values with cascading escape characters on nullable parameters #1885, Nullable enum fields (anyOf with null branch) fall back to broken raw-JSON Textarea — regression from v1.x #1928, Resource template UI does not support RFC 6570 expansion #1919, Make complex _meta possible #1910 are one defect class, not six bugs — the argument editor is not schema-aware. Treating them individually has already produced one regression from v1.x (Nullable enum fields (anyOf with null branch) fall back to broken raw-JSON Textarea — regression from v1.x #1928). A schema-aware editor closes the class and unblocks file inputs and strict validation.
• Improve DX for protocol version negotiation errors #962, inspector v2 fails to connect to MCP server at https://localhost #1936, Not working in dev containers #1951, Add support for *.localhost domains #1944, Bug: Failed manual connection leaves un-modifiable ghost server entry #1914 are one missing feature — an ordered connection diagnostic that reports which step failed (DNS · TCP · TLS · .well-known · negotiation · auth) and what to do about it. First-connection success is the entire first impression of the tool.

Notes for reviewers

• Sourcing. Built from the published roadmap at modelcontextprotocol.io/development/roadmap (last updated 2026-03-05) plus the nine WG and six IG charters, all cited in §8. The doc flags that Track A should be revised if internal planning carries themes the public page omits.
• Worth a decision: the Interceptors WG lists "CLI client for interceptor invocation and testing" as Ideating and unowned. It describes our CLI, and @olaservo co-leads both groups. Raised as an open question (§7.2) rather than assumed — if we take it, it moves from Phase 4 to Phase 3.
• Also worth a decision (§7.3): the conformance runner and CI assertions together point at "the Inspector tells you whether your server is correct," which is a real expansion of mission and possibly a charter amendment.
• Index updates deferred. The docs-maintenance rule would normally have this PR also add the file to the root README.md guide list and the AGENTS.md docs/ tree. Both are deliberately left out to keep this diff to the document itself, and will follow separately.

Testing

Markdown only — no source files touched. The change is outside every format and typecheck glob (SOURCE_EXTENSIONS in verify-format-coverage.mjs is ts/tsx/mts/cts/js/jsx/mjs/cjs, with no md), so no gate is affected and there is no behavior to smoke.

npm run ci was not run: this is a dedicated worktree with no install, and the full gate would require the postinstall cascade across all four clients to validate a documentation-only diff. GitHub CI will run it on this branch. No screenshots — no UI change.

🤖 Generated with Claude Code

https://claude.ai/code/session_013CaLPZQzfWDHoPxyL1vbYL