release: fix mcp array params and hivemind memory flooding

[github-actions]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

swarm-tools@0.60.0

Minor Changes

• 2ad83e6 Thanks @joelhooks! - Overhaul hivemind memory hooks to stop flooding sessions with garbage context

  Auto-recall fixes:

  • Add 30s cooldown between recall queries (was firing ~608x per 8hr session)
  • Raise minScore from 0.3 to 0.55 (filter weak matches)
  • Skip system messages (watchdog, heartbeat, exec logs, telegrams, etc.)
  • Truncate query to first 200 chars instead of sending full prompt
  • Move console.log spam behind debug flag
  • Reduce maxRecallResults from 5 to 3

  Auto-capture fixes:

  • Only capture assistant responses (was also capturing user messages and system prompts)
  • Replace loose CAPTURE_PATTERNS with STRONG_CAPTURE_PATTERNS requiring actual knowledge signals
  • Add 19-pattern system message blocklist (heartbeat, watchdog, OUTCOME/DECISION blocks, etc.)
  • Raise min capture length from 50 to 80 chars
  • Long messages (300+) only captured if also entity-rich
  • Reduce captures per turn from 2 to 1
  • Increase truncation from 500 to 1000 chars (capture complete thoughts)

  Format improvements:

  • Increase recalled content from 300 to 600 chars
  • Drop emoji decay badges (waste tokens)
  • Drop "Relevant memories:" header and "Use naturally" instruction
  • Cleaner format: - (85%) [tags] content

  Tag detection:

  • Remove "task" tag (matched every system log)
  • Add "gotcha", "architecture", "config" tags

  Session handler:

  • Don't store "Session ended without summary" garbage (require 50+ char summary)
  • Better session start query, limit results from 5 to 3

  New config options: recallCooldownMs, maxCapturePerTurn, captureContentLimit, recallContentLimit

claude-code-swarm-plugin@0.63.4

Patch Changes

• 02c3ae9 Thanks @joelhooks! - fix(mcp): coerce string→array params flattened by MCP protocol

  The MCP protocol flattens all type: "array" JSON Schema params to type: "string",
  causing Claude to send JSON-encoded strings or pipe-delimited strings instead of actual
  arrays. This broke hive_create_epic (subtasks), swarmmail_reserve (paths), and 8
  other array params.

  Adds coerceArrayParams() that handles JSON strings, pipe-delimited, comma-separated,
  and single values. Relaxes Zod validation to accept both strings and arrays for these
  params.

  "Coerce objects into the roles we need them to play. Guard the borders, not the
  hinterlands." — Avdi Grimm, Confident Ruby

opencode-swarm-plugin@0.63.4

Patch Changes

• 02c3ae9 Thanks @joelhooks! - fix(mcp): coerce string→array params flattened by MCP protocol

  The MCP protocol flattens all type: "array" JSON Schema params to type: "string",
  causing Claude to send JSON-encoded strings or pipe-delimited strings instead of actual
  arrays. This broke hive_create_epic (subtasks), swarmmail_reserve (paths), and 8
  other array params.

  Adds coerceArrayParams() that handles JSON strings, pipe-delimited, comma-separated,
  and single values. Relaxes Zod validation to accept both strings and arrays for these
  params.

  "Coerce objects into the roles we need them to play. Guard the borders, not the
  hinterlands." — Avdi Grimm, Confident Ruby

• feb227e Thanks @joelhooks! - Fix hivemind memory CLI pointing at wrong database

  The swarm memory CLI commands (stats, find, store, etc.) were connecting to a per-project streams.db in /tmp/ instead of the global ~/.config/swarm-tools/swarm.db where all memories actually live. This caused swarm memory stats to show 0 and swarm memory find to return no results.

  Also fixes libSQL COUNT(*) returning 0 on tables with F32_BLOB vector columns — replaced with COUNT(id) across all memory-touching code paths.

• Updated dependencies [feb227e]:

  • swarm-mail@1.11.3

@swarmtools/evals@0.2.41

Patch Changes

• Updated dependencies [02c3ae9, feb227e]:
  • opencode-swarm-plugin@0.63.4
  • swarm-mail@1.11.3

swarm-mail@1.11.3

Patch Changes

• feb227e Thanks @joelhooks! - Fix hivemind memory CLI pointing at wrong database

  The swarm memory CLI commands (stats, find, store, etc.) were connecting to a per-project streams.db in /tmp/ instead of the global ~/.config/swarm-tools/swarm.db where all memories actually live. This caused swarm memory stats to show 0 and swarm memory find to return no results.

  Also fixes libSQL COUNT(*) returning 0 on tables with F32_BLOB vector columns — replaced with COUNT(id) across all memory-touching code paths.

swarm-dashboard@0.0.16

Patch Changes

• Updated dependencies [feb227e]:
  • swarm-mail@1.11.3

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
swarmtools-ai	Ready	Preview, Comment	May 21, 2026 5:57am

[socket-security]

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	@​types/​bun@​1.3.14
	evalite@​1.0.0-beta.15 ⏵ 1.0.0-beta.16	+1		+1	-3
	@​changesets/​cli@​2.29.8 ⏵ 2.31.0			+1	-1

View full report