v7.45.0

Paired with engine-service-v0.1.118. Not additive: the tuple size limits below
narrow what the engine accepts, so the deploy relay was armed before the client
tag (docs/release-arming/, nexus-h0fo3).

Changed

• The tuple space enforces size limits (nexus-r7xao, RDR-205 amendment). The
  tuple space holds coordination metadata, not values. The engine now refuses,
  with a new typed error TooLarge (HTTP 413): a body over 4096 bytes UTF-8
  (a template may declare a lower max_body_bytes; the ledger template's is 0,
  so ledger rows carry no body); a keys, dims or keys_pattern value over
  256 bytes; a subspace over 256 bytes; a nonce, claimant or claim_id
  over 128 bytes; and a whole request body over 8192 bytes on any /v1/tuples
  route, refused before it is parsed. Limits count bytes, not characters. The
  refusal names the field and the byte counts, never the value. A reply written
  by ack is held to the same limits and refused before the transaction opens,
  so the request stays claimed. Before this release the engine enforced none of
  these: a body between 4096 bytes and the edge's 8 KB request cap was accepted
  and is now refused. HttpTupleStore checks the same limits before sending and
  raises TooLargeError, which is both a TupleError and the existing
  RequestTooLargeError, so a handler for either still catches it. nx tuple,
  the tuple_* MCP tools and the plugin's two tuple-writing hooks apply the same
  numbers. TooLarge is the tenth typed tuple error. Put longer content in T2 or
  T3 and carry a reference in the tuple.
• The v0.1.118 migration deletes existing tuple rows whose body is over 4096
  bytes (nexus-r7xao). Changeset tuples-003-body-size-limit.xml logs the
  count of such rows, deletes them (their claim-log rows stay, with tuple_id
  set to NULL), then adds and validates a CHECK constraint on
  nexus.tuples.body. A census of the live store on 2026-09-13 found 2,150
  ledger rows, all with empty bodies, and 5 mailbox rows with a largest body of
  3,285 bytes.
• An acked tuple's body is cleared (nexus-8zoyp). ack sets body to NULL
  in the statement that consumes the row, with or without a reply; a reply is its
  own row and keeps its body. A consumed row was already unreadable through rd
  and in. Changeset tuples-004-1 clears the body of rows consumed before this
  release. Nacked, lapsed and dead-lettered rows keep their bodies.
• The store read tools share search's knowledge scope (nexus-mgqs8).
  search and query read corpus="knowledge" as every live knowledge__*
  collection, but store_get, store_get_many and store_list read their
  collection="knowledge" default as the single knowledge__knowledge
  placeholder, so a note search found came back "Not found" from store_get.
  They now use the same scope. store_get finds an id across it and a title in
  every knowledge collection, and names the collections when a title matches in
  more than one instead of guessing. store_list with the bare name lists the
  subject collections and their entry counts. A full collection name narrows any
  of them to one collection; the placeholder stays reachable by its full name.
  Defaults and parameters are unchanged.
• MCP tool descriptions have one fixed shape (nexus-cnzei.5). Thirteen
  descriptions ran past the 2048-character cut Claude Code applies (nx_answer
  ran to 15,881 characters), and no parameter carried a schema description, so
  parameter docs past the cut never reached the model. Every tool on both servers
  now states its purpose, when to use it instead of a named sibling, what it
  returns, and two to four constraints, in about 1,200 characters; each
  parameter's doc is in the wire schema. Corrected claims: the scoped-search tools
  and query do not fail in local mode (the real guard only rejects a test double
  that is not HttpVectorClient-backed); search's cluster_by default is
  empty; traverse always returns an empty ids. nx_answer names the text
  markers of a degraded answer, [budget exhausted ...] and
  [non-answer: <shape>].
• One entry point per plugin name (nexus-cnzei.4, nexus-cnzei.6). Slash
  commands that shared a name with a skill are removed, the three stub agents
  that only redirected to an MCP tool (knowledge-tidier, plan-auditor,
  plan-enricher) are deleted, and seven skills are renamed. The conexus
  changelog lists every name.
• The plugin no longer auto-approves the daemon_uninstall MCP tool (it can
  delete the nexus config directory); every call now asks (nexus-cnzei.5).

Added

• nx init and nx upgrade manage a user-level beads PRIME.md
  (nexus-cnzei.8). When bd or the beads Claude Code plugin is detected, they
  install a short, generic, conexus-managed PRIME.md at the path bd prime
  falls back to when a repo has no .beads/PRIME.md (macOS
  ~/Library/Application Support/beads/PRIME.md; Linux
  $XDG_CONFIG_HOME/beads/PRIME.md or ~/.config/beads/PRIME.md; Windows
  %AppData%/beads/PRIME.md). It is machine-wide and replaces bd's bundled text
  for every beads repo without its own file; a repo-level .beads/PRIME.md still
  wins. The marker records a hash of the installed body, so a hand-authored or
  hand-edited file is never overwritten; replacing a stale managed file backs it
  up to PRIME.md.bak. Opt out with --no-beads-prime for one run or
  nx config set beads_prime.manage false for good. nx doctor reports the
  file's state and the undo steps; nx uninstall removes it only while the hash
  still matches. beads 1.2.x still appends its own bd remember memories after
  this file.
• nx doctor row tuples.watch_permission (nexus-rml7o): whether the user
  ~/.claude/settings.json (honouring CLAUDE_CONFIG_DIR), the project
  .claude/settings.json or .claude/settings.local.json carries a
  permissions.allow rule covering nx tuple watch (the documented entry, a
  broader Bash(nx tuple:*) or Bash(nx:*), or bare Bash), naming the file. A
  matching permissions.deny in any of them wins and is reported as denied. A
  missing or denied rule is a soft warning, never a failure; on a box where
  Claude Code has never run, the row reports not applicable.
• nx doctor --check-quotas names a small-window model whose tokenizer the
  client cannot find, since chunks in its collections are then not checked
  against the token window (nexus-ajvjx). Default nx doctor output is
  unchanged.
• The ledger template accepts optional commit, t2_ref and verify
  (present or absent) dims (nexus-d9k5h, engine). The plugin's ledger
  projection fills them from an agent's closing VERIFY block, and
  scripts/check_agent_verify_claims.py confirms a claimed commit exists and
  touches the named paths and that a t2_ref exists in T2 (nexus-cnzei.6).
• The peer-messaging skill: messaging other Claude sessions and dispatched
  agents, and sharing one machine (nexus-tacsg).

Fixed

• The nexus-catalog MCP search tool was registered to the wrong function
  (nexus-cnzei.1). In 7.19.0 through 7.44.0,
  mcp__plugin_conexus_nexus-catalog__search exposed a private helper,
  _file_path_matches(entry_path, wanted), which returns true or false, instead
  of catalog metadata search. Commit 4b756c8 (nexus-fhim9, first released in
  7.19.0) inserted the helper between the @mcp.tool(name="search", ...)
  decorator and catalog_search, and Python applied the decorator to the helper.
  A call with the documented parameters (query, content_type, author,
  corpus, owner, file_path, limit, offset) failed argument validation,
  because the tool required entry_path and wanted; the tool's advertised
  description was the helper's docstring. Catalog metadata search was not
  reachable through MCP in those releases. The other nine catalog tools and the
  core server were not affected. search now takes query and the filters above
  and returns catalog entries. A committed wire-schema snapshot of both servers,
  a check that no registered tool resolves to a private function, and a
  subprocess round trip that calls search over MCP now fail on this class.
• A /clear then /resume back into a session no longer leaves T1 on a dead
  token (nexus-r0d37). The MCP scratch tool returned 401 and the CLI reported
  no session lease: handing off away from a session left its lease file in
  place, and the handoff back borrowed that orphaned lease, which nothing was
  renewing. The server now deletes the lease it owns when it hands off away,
  under the mint lock and only while the file still holds its own token, so a
  sibling's fresh lease survives and the return mints fresh. A resume into a new
  process is not covered.
• The 401 data-token re-mint is bounded (nexus-r0d37, nexus-umue1). Seven
  concurrent scratch calls on a stale T1 session drew
  "data-token mint failed (429) ... rate limit exceeded" from the managed edge,
  because every 401 invalidated and re-minted the data token with no guard,
  although no data token can satisfy a stale session id. The re-mint is now
  single-flighted on the bearer the failed request sent, and skipped while an
  earlier re-mint healed the bearer and the retry still returned 401. T2 and T3
  share the guard with a 60-second expiry, since a re-mint is normally the right
  remedy there; T2 also marks the right endpoint when its base URL changes
  mid-retry.
• Markdown indexing no longer drops the end of long paragraphs (nexus-2s91y).
  The markdown chunker cut any non-code paragraph longer than about 1,690
  characters and never stored the rest, in every mode, so search could not find
  that text. It now splits the paragraph across chunks. The 12 KB byte cap in the
  markdown and PDF chunkers now splits an oversized chunk instead of truncating
  it, which had been dropping the end of large code blocks. Collections indexed
  before this release keep the truncated text until reindexed with --force
  through whichever command built them: nx index repo, nx index md or
  nx index rdr.
• Chunks fit their embedding model's token window (nexus-spujb).
  bge-base-en-v1.5 reads 512 tokens and MiniLM 256; the rest of a longer chunk
  was silently left out of its vector, so search could not find it. Indexing now
  splits any chunk over its collection's model window, counted with that model's
  own tokenizer. In such a collection, store_put and nx store put store a note
  longer than the window as several chunks under one catalog document;
  store_get returns it whole by any piece's id or by title, and nx store get
  by any piece's id. The 16,384-byte limit is unchanged. Voyage collections are
  unaffected: the 12 KB chunk cap keeps every chunk inside their 32,000-token
  window. nx doctor --check-quotas reports the windows the chunkers enforce.
  Collections indexed earlier keep their truncated vectors until reindexed with
  --force.
• Pre-1.0 dependencies are capped at their next minor version, since a 0.x
  minor release can break. conexus[local] caps fastembed below 0.8: the
  extra had no upper bound, so a fresh install resolved fastembed 0.8.0, which no
  gate had exercised. tokenizers, tree-sitter-language-pack,
  llama-index-core and voyageai are capped the same way, with floors raised
  to the versions this release is tested with; no locked version moved. The
  dependency-bounds lint and the drift watch now apply the rule to the extras and
  to 0.x minors (nexus-oqh4s).
• nx tuple watch's ping line names the MCP drain (tuple_in, then
  tuple_ack or tuple_nack) instead of an nx tuple in command with no ack
  step (nexus-dyfg8).
• The SessionStart note about stale MCP servers counts only this session's own
  MCP siblings, not every nx-mcp host on the machine, and asks the user to run
  /mcp (nexus-cnzei.2).
• The engine's scheduled tuple sweep runs its three arms independently and
  releases lapsed claims row by row, so one failing row no longer stops the sweep
  for a tenant (nexus-r7xao).

Changed (tooling and documentation)

• Release tooling: scripts/check_release_workflow_shape.py runs native-smoke's
  client probes from a checkout and the release's Maven build with Docker
  unreachable, and runs automatically inside --shakeout (nexus-xihsm);
  native-smoke.sh gains a /v1/tuples real-client probe and --shakeout
  Phase B an nx tuple verb matrix (nexus-6flt7); --candidate-migration
  populates the tuple space, an over-cap body included, before the walk
  (nexus-58vc9); the tracker leg names the production-write guard refusal
  (nexus-jzyt3); stage_native fails on a .so beside the candidate
  (nexus-og52j); local-service-gate.sh holds the build lease across stamp,
  package and restore (nexus-56qvf); the heartbeat census names its denominator
  (nexus-39etc).
• Documentation: storage-tiers.md and sibling pages brought current with the
  permanent-by-default memory_put ttl and nine T2 stores (nexus-x5kfm); flag
  defaults, doctor checks and MCP tool counts corrected (nexus-04wmv); a
  retrieval-tool selection guide in mcp-servers.md (nexus-cnzei.5); two tuple
  walkthrough diagrams render on GitHub again (nexus-4mxua).
• More release tooling: concurrent gates in one checkout no longer leave
  release.properties stamped (nexus-iexvl); the migration-rehearsal scripts no
  longer write the user's global git config and refuse to run outside their
  container, and lints refuse a tracked script that writes global git config or
  a rehearsal script without the guard (nexus-oqh4s); every shell test suite in the repository now
  runs under pytest (nexus-fcjt7, nexus-q1upi); the MCP description lint checks
  documented return keys against the code (nexus-w9jxf); five tests that failed
  only under full-suite load or in isolation are fixed (nexus-rqji3,
  nexus-dee61, nexus-ujdha, nexus-y4dmz, and the real-config guard's
  dropped_writes.jsonl handling).
• Documentation: MCP tool, agent and skill counts, the v0.1.118 tuple-space
  migration notes, note splitting in nx store and the store tools, and several
  stale claims corrected (nexus-oqh4s).