## What changed, and why The system prompt's **"MCP Tools (live registry)"** section tells the model which `mcp_*` tools exist this turn. It was written by the prompt builder (`internal/agent/prompt.go`) *before* the tool roster existed, so it described intent, not fact. Above the disclosure threshold (#506) `buildFantasyTools` hides every MCP tool behind `tool_search` / `tool_describe` / `tool_call` — and the section kept saying **"Call exactly these names"** over a list none of which resolved. Found verifying Linear for #1006: with four hosted connectors enabled (GitHub 44 + Notion 42 + Drive 8 + Linear 65 = 159 tools) the model called `mcp_linear_list_issues` by name, the agent framework answered `tool not found … Available tools: … tool_search, tool_describe, tool_call`, and the model concluded Linear was unavailable — zero bridge calls, three turns in a row. Linear alone (65 tools, direct mode) worked. Pre-existing for any bundle over ~110 tools since #506; #1446 widened the exposure to hosted-heavy deployments. **The fix: the section is derived from the roster, so it cannot disagree with the tool list.** - `agentcore.buildFantasyToolsWithRoster` returns a `toolRoster` summary next to the tools (direct `mcp_*` names, or the deferred count per connector). `buildFantasyTools` is now a thin wrapper — its ten test call sites are unchanged. - `agentcore.Run` appends the section (`live_registry.go`) to the driver's base prompt **after** the roster is built, and re-derives it from the base on the MCP-dirty rebuild path — never appended twice. Direct mode: same heading and sentences as before, names sorted. Deferred mode: "N MCP tools are available this turn, but … NOT in your tool list by name", the three bridges explained, "a direct `mcp_*` call will fail … use `tool_call`", and the connectors listed with counts (an orientation, not a 159-line roster). - `internal/agent/prompt.go` no longer writes the section. It keeps the one thing the roster cannot know — hosted connections that could **not** be mounted this turn — under its own heading `## Hosted connectors not mounted this turn`. `hostedMCPRoster` shrinks to that. - One sanitizer: `agentcore.PromptSafeName` (exported); the agent package's `promptSafeName` delegates to it. - The Browserbase skill and the `browserbase_live_view` tool description told the model *not* to trust the section (a workaround for the pre-#1446 blindness). They now say it is authoritative and how to reach the connector when the roster is deferred. `docs/TOOL-DISCLOSURE.md`, `docs/PROMPT-CACHE-CONTRACT.md`, `docs/BROWSERBASE.md` updated. Why not predict the deferral in the prompt builder instead: the count depends on opt-ins, the allowlist and the persona filter, all applied inside the roster builder; any drift between the prediction and the real decision recreates exactly this bug at the boundary. **Behavior changes to know about** - Scheduled runs gain the section (they had no roster section at all, so headless tasks with hosted connectors were equally blind). - The section moves from mid-prompt to the end of the system prompt. Still a pure function of the roster (sorted names / sorted connector counts), so the prompt-cache prefix stays byte-stable across turns. - Sub-agent children receive the **base** prompt (`internal/agent/subagent.go`) and run through the same loop, so each appends its own roster's section — a child never inherits the parent's. - Every consumer of the prompt in `Run` — the agent, `setModelContextPrefix`, the finalize seam, terminal structured output — uses the augmented form; the three auxiliary `ModelContextBudgetStep` callers already budget exactly the string they send. ## How I verified it - `go test -tags fleet_host_executor` on `internal/agentcore`, `internal/agent`, `internal/tools`, `internal/clientconfig`, `internal/scheduledrun`, `internal/httpapi`, `cmd/fleet`: all pass. `make lint`: 0 issues. - New tests: `TestLiveRegistrySectionShapes` (none / direct / deferred wording, sorted connectors, singular noun, 32× byte-stability, hostile name reduced to the grammar), `TestWithLiveRegistryJoinIsNormalized`, `TestBuildFantasyToolsWithRosterSummary` (below/above threshold summaries), and a whole-turn `TestManagerRunTurn_SystemPromptDescribesDeferredTools`: a 130-tool hosted overlay through the real `Manager.RunTurn` against the fake LLM, asserting the request's tool list carries `tool_search`/`tool_call` and no `mcp_big_*`, and the captured system prompt says "NOT in your tool list by name" / "`tool_call {name, arguments}`" / "- `big` (130 tools)" and never "Call exactly these names". - The existing whole-turn test (`…NamesHostedTools`) passes unchanged: the hosted tool names now arrive via the appended section. - `TestRun_ModelContextBudgetGuardsEverySuccessiveInnerToolStep` had to learn that the provider receives the augmented prompt: its independent accounting builds its prefix from `withLiveRegistry(systemPrompt, toolRoster{})`, because the estimator excludes the leading system message only on a byte match. The engine's own reservation was already correct. - The three prompt-cache prefix guards (`TestToolPrefix*`) pass. - Live on the OAuth test rig (main + this branch): pending — the rig rebuild follows this PR; the repro is enabling four hosted connectors in one chat. ## Scope and deviations - No schema, API or data change. ~260 lines net, a third of it tests. - Not in this PR: routing a direct call to a deferred name — the "tool not found" comes from the agent framework's own dispatcher (`charm.land/fantasy` `agent.go`), which only knows the tools it was given, so fleet cannot intercept it. - CI will show the `Audit rampart-service dependencies` step red on the unfixable `adm-zip` advisory (GHSA-vwc7-r8mq-g2x9, no patched release), the same as #1464/#1465/#1466. Everything else in the gate is expected green. Part of the #1006 OAuth verification series (#1446, #1449, #1454, #1464, #1465). 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Kristian Yendrek <yendrek.kristian@gmail.com> Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>

Uh oh!

Test oauth flow with official MCPs #1006

Description

Activity

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions