…rage

Pull the pinned-vs-history selection out of prepareGeneration into
resolveTurnWorkspaceContext so the dispatch is unit-testable: the agent
binding refresh stays in prepareGeneration, and the helper takes the
resolved agent plus prompt rows. This also drops the unconditional
skillsFromParts scan, which the pinned path discarded.

Add table tests covering pinned precedence, history fallback when the
experiment is off, the empty case, and read-error propagation.

…iment

The pinned workspace-context path was gated behind ExperimentChatContextPin.
Replace the gate with presence-based selection: when a chat has pinned
context rows (the agent reported context through the push system), build the
prompt from the pin; otherwise fall back to the per-turn history path. Older
agents that never report context get legacy compatibility automatically,
with no experiment to enable.

Removes ExperimentChatContextPin and the Server/Config experiments field that
only fed the gate, and regenerates the API docs and types.

Extend the workspace-context feature so a chat shows the context it is
actually pinned to, flags when that context has drifted from the agent's
latest snapshot, and lets the user inspect the edits and refresh.

Backend
- codersdk.ChatContext gains `resources` (pinned instruction files +
  skills, metadata only) and `changes` (per-source diff vs the agent's
  latest snapshot: added/removed/modified, with sanitized old/new bodies
  for instruction files capped at 64KiB/side; name+description for
  skills).
- chatd.Server.ContextDetail computes both on read; the single-chat GET
  handler attaches them like files/diff_status, while list and watch
  payloads stay lightweight. Nothing is persisted. Body decoders are
  factored out of contextResourcesToPrompt so reported resources match
  what the prompt injects.

Frontend
- The context ring shows a yellow drift triangle (distinct error
  treatment), drives its file/skill list from the pinned resources
  (falling back to last_injected_context), and adds a "Refresh context"
  action plus a "View changes" dialog that renders instruction-file
  edits via the existing DiffViewer and lists skill changes.
- experimental.refreshChatContext + a refreshChatContext mutation;
  mergeWatchedChatSummary now carries context so context_dirty watch
  events update the cache, and the open chat is refetched to pull the
  change set.

Tests: diff/resource computation and ContextDetail gating (Go unit),
the GET resources/changes end to end (integration), the context_dirty
cache merge (unit), and indicator stories (clean/dirty/error play).

…sources

coder exp chat context add/show/remove now resolve a relative source path (e.g. ./) to an absolute path before sending it to the agent, which requires canonical absolute paths. A leading ~ is preserved for the agent to expand against its own home.

chatd now includes mcp_config and mcp_server rows in a chat's pinned context list and in the drift diff, instead of filtering to instruction files and skills only. MCP changes carry only source, kind, and status (no body diff).

The context popover gains an MCP section (config by file basename, server by name) and the changes dialog lists MCP changes alongside skills.

…apshot

Wire the MCP manager's live tool cache into the agentcontext resolver via a new ManagerOptions.MCP provider, so each connected MCP server becomes a KindMCPServer resource (with its tools) in the snapshot alongside instruction files and skills. Previously Resolver.MCP was never set, so only the .mcp.json file appeared (as an mcp_config resource) and servers/tools never reached coderd or the UI.

agent/x/agentmcp.Manager: export CachedTools (a non-blocking copy of the tool cache) and add SetOnToolsChanged, fired after a reload writes the cache so the agentcontext manager re-resolves when the live tool set changes.

Only servers exposing at least one tool are surfaced; a connected server with no tools, or a failed connection, is not shown. The wire/proto and coderd storage path for KindMCPServer already existed; this change only produces the resources.

Decode the pinned mcp_server body and report its tools (with the agent's "<server>__" name prefix stripped) on the single-chat GET, via a new codersdk.ChatContextMCPTool list on ChatContextResource.

The context popover now lists each MCP server's tools beneath it (wrench icon, description tooltip), mirroring how skills are shown. Tool decoding is defensive: a missing or malformed body yields no tools rather than breaking the pinned-context list.

pinnedContextResources now reports non-OK resources (invalid, unreadable, oversize, excluded) with their Status and Error instead of dropping them silently, and stamps Status on healthy resources. codersdk.ChatContextResource gains Status and Error (ChatContextResourceStatus mirrors the agent resolver's per-resource status, already stored end-to-end).

The context popover lists problem resources in a new Issues section (warning icon, kind and status, and the agent's error message). For example, a skill rejected because its front-matter name does not match its directory now shows that reason instead of silently vanishing from the list and read_skill.

…text issues

…the agent token