feat: show workspace MCP discovery status and warnings in chat context - #29453

Stack context

This stack replaces #29114 after review feedback that a single mcp_settled bool was underdeveloped and the change too large to review. It defines how workspace MCP discovery is communicated from the agent to chats: the agent publishes one discovery report inside the existing context snapshot (PushContextState), coderd persists it atomically with the resources, and chatd, find_tools, the create/start workspace tools, and the chat context UI all consume that same committed report.

The report has three parts. Completeness: MCPDiscovery.phase is PENDING until the agent's initial MCP reload has finished (success, failure, or zero servers all count) and COMPLETE afterwards for the rest of the process; later reloads update outcomes but never move it back. Per-source outcomes stay on the existing ContextResource rows: mcp_config rows carry structural and semantic config errors, mcp_server rows carry discovered servers (including zero tools), failed servers with a sanitized error, and a retained previous connection as OK plus a warning. Freshness: every snapshot carries the agent_run_id of the process that produced it, and the agent reports the same id in Startup, so a snapshot is known to describe the current process when the two ids match. Timestamps are deliberately not used for freshness because ready_at is agent clock and received_at is coderd clock.

Discovery completeness, per-source outcomes, turn eligibility, and the result of an actual tool invocation stay separate facts. Out of scope: continuous health monitoring, reconnect detection, automatic retries, and any retry or reload UI.

Layers, bottom to top: #29448 agent diagnostics, #29449 protocol and persistence, #29450 agent publication, #29451 chat projection, #29452 first-turn gate, #29453 context UI. Each layer builds and passes its tests on its own. The composed stack was validated end to end with remote UAT; the record is on #29453.

Why

The context usage indicator listed healthy servers and treated only non-ok rows as issues. With the discovery report it can now tell the user whether discovery is initializing, complete, complete but empty, usable with warnings, failed, or describes a previous agent run.

Fix

ContextUsageIndicator renders from chat.context.mcp_discovery and the existing resource rows: an initializing line while the phase is pending, "Discovery complete. No MCP servers discovered." for an empty complete catalog, discovered servers with zero tools marked "(no tools)", warnings on ok rows added to Issues without marking the server unavailable, failed servers labeled failed, and a stale banner when the rows come from a previous agent run and tools are withheld. Long error text wraps inside the popover. Copy never claims live connectivity. The existing context_dirty refetch already drives updates, so no new fetches are added. Storybook stories cover each state plus an in-place update.

Testing

Typecheck, biome, oxlint, the React Compiler check, and the Storybook tests for the component (17 stories, including the pre-first-snapshot pending state). The repository frontend-review checklist was applied to the diff.

Remote UAT record (composed stack)

UAT ran remotely through Coder Agents against the branch tip with a nested deployment built from the exact commit, a real model, and a disposable stdio MCP fixture with planted secret sentinels. Chat: https://dogfood.cdr.dev/agents/7b13b4d7-55d0-435d-b76b-a47d6e37608a.

• Round 1 on 8884f7f53cf (branch mike/workspace-mcp-discovery): endorsed FAIL. A hung server delayed a healthy sibling's tools past the bounded wait; a bound chat lost its whole context display after a workspace restart with no interaction (root cause: the new chat GET path failed when the bound agent row was soft-deleted by the rebuild, confirmed with DB and log forensics); long error text broke the popover layout; failed servers were labeled with the raw status word. The data plane (redaction, config errors, zero-tool servers, reconnect warnings, withholding, plan-mode exclusion, tool-result embedding) held.
• Round 2 on a3d761da266 (same branch, fixes applied): endorsed PASS. All four issues re-verified with browser evidence, plus regression checks (mixed working, broken, and unreachable servers with redaction; 12 s delayed sibling; 25 s sibling longer than the wait; always-error; removing .mcp.json while up; repeated identical reloads; create and start tool results; empty tools, no match, plan mode). Remaining findings were cosmetic; the stale banner wording was adjusted afterwards.

Heads at the first Codex round differed from a3d761da266 by: the autosquash of the fix commits into their owning layers (tree preserving), one copy-only change to the stale banner text in ContextUsageIndicator.tsx (Storybook tests re-run), and a rebase from 6967259af4d onto c29116d4bc6 (four unrelated main commits; sqlc regeneration produced no diff; targeted agent, agentapi, and chatd suites re-run green). Those changes were not exercised by live UAT. The Codex round 1 fixes (see each layer's review record) and the rebase onto 9d97337 are also unverified by live UAT; a follow-up UAT round is planned once the stack is otherwise clean.

Remote UAT record

• Round 1 (FAIL) and round 2 (PASS) ran on the pre-stack branch head a3d761d; evidence under .mux-uat/round-1 and .mux-uat/round-2 in the implementing workspace, remote chat https://dogfood.cdr.dev/agents/7b13b4d7-55d0-435d-b76b-a47d6e37608a.
• Round 3 ran on the composed head 190f232 (07:20 to 08:19 UTC): the redaction sweep, custom-config diagnostics, bounded diagnostics, hung sibling with two healthy servers, pre-first-snapshot state, rebuild rebind with the new banner wording, and the .mcp.json removal path were verified through the popover and the chat API with zero sentinel leaks; real-model turns (tool execution, create_workspace and start_workspace results) were blocked by a stale provider key on the test deployment (HTTP 403 from the AI gateway) and a re-run probe was still open when this record was written. Two observations for the author: an invalid entry in the same .mcp.json as a healthy server invalidates the whole file (pre-existing ParseConfig semantics), and a stdio server whose positional argument equals its own name has that name redacted inside its error text.
• Commits after 190f232 on this stack (Codex rounds 5 to 7 fixes: per-session and ambient secret redaction, one-sided run id rule in the wait, spent-attempt refresh and re-keying, bounded prompt note, header credential redaction) are not covered by live UAT.

Review record

• Codex review 1, requested 2026-09-17 04:29 UTC on 802f7fdd79: no findings (Codex Review: Didn't find any major issues). This layer then gained the context indicator sentence in extending-agents.md (moved from feat: gate first turns on current-run MCP discovery #29452) and one story.
• Codex review 2, requested 2026-09-17 05:37 UTC on 015fef82ea: no findings on 015fef82ea; the title check then failed because the moved docs sentence sits outside the (site) scope, so the PR title dropped its scope (head unchanged).
• Codex review 3, requested 2026-09-17 06:14 UTC on f9cd025a37: no findings on f9cd025a37.
• Codex review 4, requested 2026-09-17 06:42 UTC on a58845980b: no findings on a58845980b.
• Codex review 5, requested 2026-09-17 07:17 UTC on 190f2322ca: no findings on 190f2322ca.
• Codex review 6, requested 2026-09-17 07:54 UTC on c56e317bc0: no findings on c56e317bc0.
• Codex review 7, requested 2026-09-17 08:26 UTC on d9bad62950: no findings on d9bad62950.
• Codex review 8 requested on ce0c8110a3: pending.

This PR was authored by Xum on behalf of @ibetitsmike.