Add the foundation for deploying and managing MCP servers on the VPS
broker, with per-peer credential vaults and visibility scopes.
Architecture:
- One Docker container per mesh with a Node supervisor
- Each MCP server runs as a child process with its own stdio pipe
- claudemesh launch installs native MCP entries in ~/.claude.json
- Mid-session deploys fall back to svc__* dynamic tools + list_changed
New components:
- DB: mesh.service + mesh.vault_entry tables, mesh.skill extensions
- Broker: 19 wire protocol types, 11 message handlers, service catalog
in hello_ack with scope filtering, service-manager.ts (775 lines)
- CLI: 13 tool definitions, 12 WS client methods, tool call handlers,
startServiceProxy() for native MCP proxy mode
- Launch: catalog fetch, native MCP entry install, stale sweep, cleanup,
MCP_TIMEOUT=30s, MAX_MCP_OUTPUT_TOKENS=50k
Security: path sanitization on service names, column whitelist on
upsertService, returning()-based delete checks, vault E2E encryption.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Persistent MCP servers (opt-in via `persistent: true`) survive host
disconnects — they appear as offline in mcp_list and auto-restore when
the host reconnects. Ephemeral servers (default) still clean up on
disconnect. Offline servers return a clear error on mcp_call with
time-since-disconnect info.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Save groups, profile, visibility, summary, display name, and cumulative
stats to a new mesh.peer_state table on disconnect. On reconnect (same
meshId + memberId), restore them automatically — hello groups take
precedence over stored groups if provided. Broadcast peer_returned
system event with last-seen time and summary to other peers.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Peers now receive [system] notifications when MCP servers join or
leave the mesh, with tool names and hosting peer info.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Peers report os.hostname() in the hello handshake. list_peers shows
[local] or [remote] tag per peer. MCP instructions teach AI to read
local peers' files directly via filesystem instead of relay.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add the webhook handler module (webhooks.ts) that verifies secrets
against the mesh.webhook table and broadcasts incoming HTTP POST
payloads to all connected mesh peers. This completes the webhook
feature whose schema, types, WS CRUD handlers, and CLI tools were
added in the previous commits.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds share_skill, get_skill, list_skills, and remove_skill across the full
stack (Drizzle schema, broker CRUD + WS handlers, CLI client methods, MCP
tools). Skills are mesh-scoped, unique by name, and searchable via ILIKE
on name/description/tags.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Add tamper-evident audit logging where each entry includes a SHA-256
hash of the previous entry, forming a verifiable chain per mesh.
Events tracked: peer_joined, peer_left, state_set, message_sent
(never logs message content). New WS handlers: audit_query for
paginated retrieval, audit_verify for chain integrity verification.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Peers self-report resource usage via set_stats; stats visible in
list_peers responses and the new mesh_stats MCP tool. CLI auto-reports
every 60s and tracks messagesIn/Out, toolCalls, uptime, and errors.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Peers can register MCP servers with the mesh and other peers can invoke
those tools through the existing claudemesh connection without restarting.
Broker: in-memory MCP registry with mcp_register/unregister/list/call
handlers, call forwarding to hosting peer with 30s timeout, and automatic
cleanup on peer disconnect.
CLI: mcpRegister/mcpUnregister/mcpList/mcpCall client methods, inbound
mcp_call_forward handler, and 4 new MCP tools (mesh_mcp_register,
mesh_mcp_list, mesh_tool_call, mesh_mcp_remove).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Replace in-memory-only setTimeout scheduling with a DB-backed system
that survives broker restarts. Adds:
- `scheduled_message` table in mesh schema (Drizzle + raw CREATE TABLE
for zero-downtime deploys)
- Minimal 5-field cron parser (no dependencies) with next-fire-time
calculation for recurring entries
- On broker boot, all non-cancelled entries are loaded from PostgreSQL
and timers re-armed automatically
- CLI `schedule_reminder` MCP tool accepts optional `cron` expression
- CLI `remind` command accepts `--cron` flag
- One-shot reminders remain backward compatible — no cron field = same
behavior as before
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The "0 */2 * * *" cron example inside a /** comment caused TSC to
parse */ as end-of-comment, producing syntax errors.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Extend the WS hello handshake with optional peerType, channel, and model
fields so peers can advertise what kind of client they are. The broker
stores these in-memory on PeerConn and returns them (along with cwd) in
the peers_list response. CLI peers command and MCP list_peers tool now
display the new metadata.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When a peer connects or disconnects, the broker now broadcasts a
system push (subtype: "system") to all other peers in the same mesh.
The CLI formats these as [system] channel notifications so AI sessions
can react to topology changes without polling.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Extract _reqId from incoming WS messages and include it in every direct
response sendToPeer call and sendError call. Clients can now match
responses to requests by ID instead of relying on FIFO ordering.
Old clients without _reqId are unaffected (field simply omitted).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- shareContext: adds optional memberId param; when provided, upserts on
(meshId, memberId) instead of (meshId, presenceId) — prevents stale
context rows accumulating on every reconnect. Falls back to presenceId
for legacy/anonymous connections. Also refreshes presenceId on update
so it stays current.
- schema: adds member_id column + unique index context_mesh_member_idx
on mesh.context table; new migration 0013_context-stable-member-key.sql.
- index.ts call site updated to pass conn.memberId as the stable key.
- createStream: replaces SELECT-then-INSERT TOCTOU race with atomic
INSERT ... ON CONFLICT DO NOTHING RETURNING, followed by SELECT on miss.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- claim_task/complete_task: send taskId not id
- graph_result: read msg.records not msg.rows
- message_status: try all mesh clients, not only first
- broker: omit state_result for set_state (fixes get_state cross-contamination)
- error handler: unblock first pending resolver on unmatched broker errors
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
broker: owner also fetches sealedKey from mesh.file_key (not skipped),
only non-owners are blocked when key is missing
cli: explicit error when encrypted file has no sealedKey (no silent raw download)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- parse x-encrypted/x-owner-pubkey/x-file-keys headers in handleUploadPost
- pass encrypted and ownerPubkey to uploadFile, call insertFileKeys after
- get_file: fetch sealedKey for non-owners, block if missing, include in response
- list_files: include encrypted field per file
- add grant_file_access WS handler so owners can seal keys for peers
- update types.ts with new message interfaces and union members
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
broker: expand member groups to ancestor paths at drain time (pull model)
- @flexicar message reaches peers in @flexicar/core, @flexicar/output, etc.
- Resolved at drainForMember — no DB changes, fully backward-compatible
- Any depth: flexicar/team/backend also matches @flexicar and @flexicar/team
cli: wire --role all the way through to session config + env
- Config.role field added
- launch.ts stores role in sessionConfig, passes CLAUDEMESH_ROLE env var
- mcp/server.ts includes role in identity string
- manager.ts auto-joins groups from config on WS connect (--groups flag now works)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Normalise tags to Array before Drizzle insert (PgArray mapper calls
.map() and throws if value is not a standard JS Array)
- Use uploadedByName instead of uploadedByMember FK — the X-Member-Id
header carries the mesh slug, not a mesh.member primary key
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
z.coerce.boolean() treats any non-empty string as true, so MINIO_USE_SSL="false" → true.
Switch to explicit enum+transform.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The sender exclusion filter (excludeSenderSessionPubkey) was blocking
delivery of ALL messages from the sender, including direct messages
to other peers. Now only excludes on broadcast (target_spec = '*').
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Files: MinIO-backed file sharing built into the broker.
share_file for persistent mesh files, send_message(file:) for
ephemeral attachments. Presigned URLs for download, access
tracking per peer.
Broker infra: MinIO in docker-compose, internal network.
HTTP POST /upload endpoint. WS handlers for get_file,
list_files, file_status, delete_file.
Multi-target: send_message(to:) accepts string or array.
Targets deduplicated before delivery.
Targeted views: MCP instructions teach Claude to send
tailored messages per audience instead of generic broadcasts.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Phase B + C + message delivery status.
State: shared key-value store per mesh. set_state pushes changes
to all peers. get_state/list_state for reads. Peers coordinate
through shared facts instead of messages.
Memory: persistent knowledge with full-text search (tsvector).
remember/recall/forget. New peers recall context from past sessions.
message_status: check delivery status with per-recipient detail
(delivered/held/disconnected).
Multicast fix: broadcast and @group messages now push directly to
all connected peers instead of racing through queue drain.
MCP instructions: dynamic identity injection (name, groups, role),
comprehensive tool reference, group coordination guide.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Phase A of the claudemesh spec. Peers can now join named groups
with roles, and messages route to @group targets.
Broker:
- @group routing in fan-out (matches peer group membership)
- @all alias for broadcast
- join_group/leave_group WS messages + DB persistence
- list_peers returns group metadata
- drainForMember matches @group targetSpecs in SQL
CLI:
- join_group/leave_group MCP tools
- send_message supports @group targets
- list_peers shows group membership
- PeerInfo includes groups array
- Peer name cache for push notifications
Launch:
- --role flag (optional peer role)
- --groups flag (comma-separated, e.g. "frontend:lead,reviewers")
- Interactive wizard for role + groups when flags omitted
- Groups written to session config for broker hello
Spec: SPEC.md added with full v0.2 vision (groups, state, memory)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
excludeSenderMemberId blocked delivery to ALL peers sharing the
same member_id (all sessions from one join). Replaced with
excludeSenderSessionPubkey which only excludes the sender's own
session — peers with different session pubkeys receive correctly.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Store sender's sessionPubkey on message_queue at send time.
drainForMember returns COALESCE(sender_session_pubkey, peer_pubkey)
so the recipient gets the correct sender key for decryption.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Each WS connection generates its own ed25519 keypair (sessionPubkey)
sent in the hello handshake. The broker stores it on the presence
row and uses it for message routing + list_peers. This gives every
`claudemesh launch` a unique crypto identity without burning invite
uses — member auth stays permanent, session identity is ephemeral.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove ALL Payload imports, withPayload wrapper, and (payload)
routes. Blog index + changelog are now static data arrays.
Blog post at /blog/peer-messaging-claude-code is static TSX.
Payload CMS stays as a dev dependency for future local admin
but has zero presence in the production build.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
WS handshake is now authenticated end-to-end. The broker proves that
every connected peer actually holds the secret key for the pubkey
they claim as identity — not just that they know the pubkey.
wire format change:
{type:"hello", meshId, memberId, pubkey, sessionId, pid, cwd,
timestamp, signature}
where signature = ed25519_sign(canonical, secretKey)
and canonical = `${meshId}|${memberId}|${pubkey}|${timestamp}`
broker verifies on every hello:
1. timestamp within ±60s of broker clock → else close(1008, timestamp_skew)
2. pubkey is 64 hex chars, signature is 128 hex chars → else malformed
3. crypto_sign_verify_detached(signature, canonical, pubkey) → else bad_signature
4. (existing) mesh.member row exists for (meshId, pubkey) → else unauthorized
All rejection paths close the WS with code 1008 + structured error
message + metrics counter increment (connections_rejected_total by
reason).
new modules:
- apps/broker/src/crypto.ts: canonicalHello, verifyHelloSignature,
HELLO_SKEW_MS constant
- apps/cli/src/crypto/hello-sig.ts: matching signHello helper
clients updated:
- apps/cli/src/ws/client.ts: signs hello before send
- apps/broker/scripts/{peer-a,peer-b}.ts (smoke-test): sign hellos
with seed-provided secret keys
new regression tests — tests/hello-signature.test.ts (7):
- valid signature accepted
- bad signature (signed with wrong key) rejected
- timestamp too old rejected (>60s)
- timestamp too far in future rejected (>60s)
- tampered canonical field (different meshId at verify time) rejected
- malformed hex pubkey rejected
- malformed signature length rejected
verified live:
- apps/broker/scripts/smoke-test.sh: full hello+ack+send+push flow
- apps/cli/scripts/roundtrip.ts: signed hello + encrypted message
- 55/55 tests pass
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Direct messages between peers are now end-to-end encrypted. The
broker only ever sees {nonce, ciphertext} — plaintext lives on the
two endpoints.
apps/cli/src/crypto/envelope.ts:
- encryptDirect(message, recipientPubkeyHex, senderSecretKeyHex)
→ {nonce, ciphertext} via crypto_box_easy, 24-byte fresh nonce
- decryptDirect(envelope, senderPubkeyHex, recipientSecretKeyHex)
→ plaintext or null (null on MAC failure / malformed input)
- ed25519 keys (from Step 17) are converted to X25519 on the fly via
crypto_sign_ed25519_{pk,sk}_to_curve25519 — one signing keypair
covers both signing + encryption roles.
BrokerClient.send():
- if targetSpec is a 64-hex pubkey → encrypt via crypto_box
- else (broadcast "*" or channel "#foo") → base64-wrapped plaintext
(shared-key encryption for channels lands in a later step)
InboundPush now carries:
- plaintext: string | null (decrypted body, null if decryption failed
OR it's a non-direct message)
- kind: "direct" | "broadcast" | "channel" | "unknown"
MCP check_messages formatter reads plaintext directly.
side-fixes pulled in during 18a:
- apps/broker/scripts/seed-test-mesh.ts now generates real ed25519
keypairs (the previous "aaaa…" / "bbbb…" fillers weren't valid
curve points, so crypto_sign_ed25519_pk_to_curve25519 rejected
them). Seed output now includes secretKey for each peer.
- apps/broker/src/broker.ts drainForMember wraps the atomic claim in
a CTE + outer ORDER BY so FIFO ordering is SQL-sourced, not
JS-sorted (Postgres microsecond timestamps collapse to the same
Date.getTime() milliseconds otherwise).
- vitest.config.ts fileParallelism: false — test files share
DB state via cleanupAllTestMeshes afterAll, so running them in
parallel caused one file's cleanup to race another's inserts.
- integration/health.test.ts "returns 200" now uses waitFullyHealthy
(a 200-only waiter) instead of waitHealthyOrAny — prevents a race
with the startup DB ping.
verified live:
- apps/cli/scripts/roundtrip.ts (direct A→B): ciphertext in DB is
opaque bytes (not base64-plaintext), decrypted correctly on arrival
- apps/cli/scripts/join-roundtrip.ts (full join → encrypted send):
PASSED
- 48/48 broker tests green
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
drainForMember previously ran SELECT undelivered rows, THEN UPDATE
delivered_at. Two concurrent callers (e.g. WS fan-out on send +
handleHello's own drain for the target) could both SELECT the same
row before either UPDATEd, pushing the same envelope twice.
now: single atomic UPDATE ... FROM member ... WHERE id IN (
SELECT id ... FOR UPDATE SKIP LOCKED
) RETURNING mq.*, m.peer_pubkey AS sender_pubkey.
FOR UPDATE SKIP LOCKED is the key primitive — concurrent callers
each claim DISJOINT sets, so a message can never be drained twice.
Union of all concurrent drains still covers every eligible row.
re-sorts RETURNING rows by created_at client-side (Postgres makes no
FIFO guarantee on the RETURNING clause's output order), and normalizes
created_at to Date since raw-sql results can come back as ISO strings.
regression: tests/dup-delivery.test.ts (4 tests)
- two concurrent drains produce disjoint result sets
- six concurrent drains partition cleanly (20 messages, each drained once)
- subsequent drain after success returns empty
- FIFO ordering preserved within a single drain
48/48 tests pass. Live round-trip no longer logs the double-push.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Single HTTP POST /join the CLI calls after parsing an invite link +
generating an ed25519 keypair client-side. Broker validates the mesh
exists + is not archived, inserts a mesh.member row (or returns the
existing id for idempotency), returns {ok, memberId, alreadyMember?}.
body: {mesh_id, peer_pubkey, display_name, role}
- peer_pubkey must be 64 hex chars (32 bytes)
- role is "admin" | "member"
v0.1.0 trusts the request — no invite-token validation, no ed25519
signature check. Both land in Step 18 alongside libsodium wrapping.
size cap enforced via MAX_MESSAGE_BYTES (shared with hook endpoint).
structured log line per enrollment with truncated pubkey + whether
it was a new member or re-enrolled existing one.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
broker-client: full WS client with hello handshake + ack, auto-reconnect
with exponential backoff (1s → 30s capped), in-memory outbound queue
(max 100) during reconnect, 500-entry push buffer for check_messages.
MCP tool integration:
- send_message: "slug:target" prefix or single-mesh fast path
- check_messages: drains push buffers across all clients
- set_status: fans manual override across all connected meshes
- set_summary: stubbed (broker protocol extension needed)
- list_peers: stubbed — lists connected mesh slugs + statuses
manager module holds Map<meshId, BrokerClient>, starts on MCP server
boot for every joined mesh in ~/.claudemesh/config.json.
new CLI command: seed-test-mesh injects a mesh row for dev testing.
also fixes a broker-side hello race: handleHello sent hello_ack before
the caller closure assigned presenceId, so clients sending right after
the ack hit the no_hello check. Fix: return presenceId, caller sets
closure var, THEN sends hello_ack. Queue drain is fire-and-forget now.
round-trip verified: two clients, A→B, push received with correct
senderPubkey + ciphertext. 44/44 broker tests still pass.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds the minimum ops surface area for a production broker without
over-engineering. All new config knobs are env-var driven with sane
defaults.
New modules:
- logger.ts: structured JSON logs (one line, stderr, ready for
Loki/Datadog ingestion without preprocessing)
- metrics.ts: in-process Prometheus counters + gauges, exposed at
GET /metrics. Tracks connections, messages, queue depth, TTL
sweeps, hook requests, DB health.
- rate-limit.ts: token-bucket rate limiter keyed by (pid, cwd).
Applied to POST /hook/set-status at 30/min default.
- db-health.ts: Postgres ping loop with exponential-backoff retry.
GET /health returns 503 while DB is down.
- build-info.ts: version + gitSha (from GIT_SHA env or `git rev-parse`
fallback) + uptime, surfaced on /health.
Behavior changes:
- Connection caps: MAX_CONNECTIONS_PER_MESH (default 100). Exceed →
close(1008, "capacity") + metric increment.
- Message size: MAX_MESSAGE_BYTES (default 65536). WS applies it via
`ws.maxPayload`. Hook POST bodies cap out with 413.
- Structured logs everywhere replacing the old `log()` helper.
- Env validation stricter: DATABASE_URL required + regex-checked for
postgres:// prefix.
New endpoints:
- GET /health → {status, db, version, gitSha, uptime}. 503 if DB down.
- GET /metrics → Prometheus text format.
Verified: 21/21 tests still pass. Hit /health + /metrics live —
gitSha resolves correctly via `git rev-parse --short HEAD` in dev.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds scripts/{seed-test-mesh,peer-a,peer-b,smoke-test}.ts|.sh that
prove an end-to-end message flow works against a real Postgres:
- seed-test-mesh.ts creates user+mesh+2 members with deterministic
hex pubkeys ("aa..aa", "bb..bb"), writes seed JSON to stdout
- peer-a.ts sends hello then a direct "send" message to peer B's
pubkey with fake ciphertext "hello-from-a"
- peer-b.ts sends hello, waits up to 5s for a push, asserts
senderPubkey matches peer A, exits 0/1
- smoke-test.sh wires the three together
Verified flow: hello registers presence row → send queues into
mesh.message_queue → fanout matches connected peer by pubkey →
drainForMember joins on mesh.member for senderPubkey → push lands
with ciphertext + correct sender attribution.
Also fixes a date-serialization bug that blocked the first run:
applyPendingHookStatus used `sql${col} >= ${jsDate}` which passed
JS Date.toString() to Postgres (failed to parse). Replaced raw
sql`` template with typed gte/desc/isNotNull operators from
drizzle-orm. Same fix applied in sweepPendingStatuses.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Port 7899 is used by claude-intercom's broker on dev machines (it's
the convention for that tool). claudemesh is a distinct product and
should have its own default port. 7900 is unreserved and unconflicted.
Prod deploys override via BROKER_PORT env var, so this only affects
local dev ergonomics.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The schema/index.ts barrel does `export * from "./mesh"` + `export *
from "./auth"`. Both modules exported a symbol named `member`, which
caused TypeScript to silently exclude the ambiguous re-export and
drizzle-kit's introspection couldn't see mesh.member — its generated
migration was missing that table entirely.
Fix: rename the TypeScript binding only. The DB table name stays
"member" inside pgSchema "mesh" (still mesh.member in SQL):
- `export const member = schema.table("member", ...)` →
`export const meshMember = schema.table("member", ...)`
- Internal references in mesh.ts updated (FK lambdas, relations,
Zod schemas, inferred TS types)
- apps/broker/src/broker.ts import updated to meshMember as memberTable
- migrations/0000_sloppy_stryfe.sql regenerated — now includes all 7
mesh.* tables (audit_log, invite, member, mesh, message_queue,
pending_status, presence)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Single-port refactor:
- Drop the BROKER_PORT+1 HTTP side-port. Use `ws` with noServer:true
and attach to a single node:http server via the 'upgrade' event.
- Clients connect to ws://host:PORT/ws
- Hook POSTs go to http://host:PORT/hook/set-status
- Health probe at http://host:PORT/health
- One port = one Traefik label, one cert, one deploy route. Matches
the Coolify/VPS operational constraints.
senderPubkey on push:
- drainForMember now joins mesh.message_queue → mesh.member to return
the sender's peerPubkey alongside each envelope. No extra round-trip,
no cache invalidation needed (option A from review).
- index.ts populates WSPushMessage.senderPubkey from the join result
instead of the empty-string placeholder.
- Receivers can now identify who sent a message directly from the push.
README updated with a routes table for the single-port layout.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
