[BUG][platform: claude.ai web] Post-compaction tool manifest rebuilt from last tool_search load only — native tools (incl. tool_search itself) dropped for rest of session (recurring, 5 threads)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Scope note (honest up front): this occurs in claude.ai web chat with MCP connectors (Microsoft 365, DocuSign) — not in the Claude Code CLI. Filing here because this is Anthropic's only public issue tracker with documented staff responses, and tool-manifest/compaction handling may be shared platform infrastructure. If this is the wrong queue, please route to the claude.ai platform team rather than closing — a support ticket with the full postmortem has been filed in parallel.

In long claude.ai sessions doing high-volume connector work (Outlook mailbox pagination via outlook_email_search, limit=25, JSON payloads of ~1,500-2,500 tokens per call, hundreds of calls per session), the tool manifest served to the model abruptly collapses from the full set (~20+ tools) to a small connector-only subset. All filesystem tools (bash_tool, create_file, str_replace, view, present_files), platform tools (project_knowledge_search, conversation_search, web_search, memory_user_edits) and tool_search itself disappear for the rest of the session.

Key diagnostic (best-documented thread): the surviving subset matched exactly — in number and identity — the tools loaded by that thread's last tool_search call, per the thread's own transcript: "Loaded 5 tools: Microsoft 365 (3): outlook_email_search, chat_message_search, outlook_calendar_search. Docusign (2): getEnvelopes, createEnvelopeFromTemplate." Another MCP connected to the same project was NOT retained — what survived is literally what had passed through that one tool_search load. Since tool_search itself is also dropped, there is no path left to reload anything.

Additional observed properties (5 threads, 09-10 Jul 2026):

1. The collapse lands at the start of a turn following a long uninterrupted extraction sequence. In the best-diagnosed thread the session had already opened with a "This conversation was successfully compacted..." notice; the manifest survived exactly one turn after that notice and collapsed on the next — a turn triggered by a user message with no content of its own ("Continuar").
2. 2. The degraded subset fluctuates turn to turn (one turn: 5 tools; a later turn: a different 4-tool subset; another thread: zero at first, then 5). This rules out a static permission/feature-flag failure and points at per-turn context/manifest assembly.
3. 3. No error is ever surfaced — tools are simply not offered. There is no signal the user or model can rely on to detect it.
4. 4. Consequence: the session keeps extracting through the surviving connector but can no longer persist results to files; thread close-out flows that depend on file tools are blocked. Real data-loss risk (materialized once: a 25-record batch extracted post-loss had to be reconstructed).

Related but distinct report on a different platform: #52221 ("Native Anthropic tool tier silently disappears from advertised functions list on Claude for Windows 1.3883.0") describes a similar symptom — native/tool_search-loaded tools dropping after what looks like context handling — but on Claude Desktop for Windows, not claude.ai web, and is closed as stale/not-planned rather than confirmed or fixed. Mentioning it only to flag a possible cross-platform pattern, not claiming it's the same bug.

What Should Happen?

The full manifest the conversation started with (or at minimum: native filesystem tools + tool_search) remains available after compaction.

Hypothesis (explicitly unconfirmed, but consistent with all observed facts above): post-compaction manifest reconstruction rebuilds from the record of tools loaded via tool_search instead of (native set union loaded set). Supporting observation from a separate healthy session on the same account: connector tools are served as "deferred" and loaded on demand via tool_search — the platform already assembles the manifest dynamically.

Error Messages/Logs

Actual Behavior: Only the tools from the last tool_search load remain; every native tool, including tool_search, is gone for the rest of the session. The subset then fluctuates turn to turn. No error message or log is ever produced — this is precisely the problem: the degradation is silent.

Environment: Platform: claude.ai web (chat), paid plan. Connectors: Microsoft 365 (heavy legitimate use — e-discovery on our own mailboxes), DocuSign. Dates: 09-10 July 2026, 5 independent threads (first documented 09 Jul). Parallel reports already filed: thumbs-down on the exact turns + a support ticket with the full technical postmortem.

Steps to Reproduce

1. claude.ai web session with MCP connectors enabled (Microsoft 365 + one more, e.g. DocuSign); ideally a session already showing a "conversation was compacted" notice.
2. 2. Load a subset of connector tools via one tool_search call.
3. 3. Run a long uninterrupted mechanical extraction (hundreds of outlook_email_search calls, full JSON output, multi-hour session).
4. 4. Send a contentless continuation message ("Continue").
5. 5. Ask Claude to write results to a file.

Claude Model

None

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

N/A — this bug occurs in claude.ai web chat, not the Claude Code CLI; no CLI version is involved.

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

_No response_