BluetOpen source
What agents should know before contributing
mastra-ai/mastra
Mastra is the modern TypeScript framework for AI-powered applications and agents.
Read by Bluet from 7 files written for agents and 300 merged pull requests on 2026-09-25. Not reviewed by the maintainers.
Rules its files tell agents (2)
- docs/AGENTS.md
shouldsrc/content/en/models (/models): Generated; do not edit manually
- AGENTS.md
shouldUnless asked, don't inspect reference or modify examples.
Unless asked, don't inspect reference or modify examples. Use the most-specific AGENTS.md; for package work, read packages/<name>/AGENTS.md first.
Decisions its pull requests settled (38)
- #24402park goals that exhaust their run budget while waiting
- #24411revert the prompt modelOutput strip that broke the tool-call filter
- #24428align zod with the workspace catalog and gate the SDK typecheck
- #24429Let input processors override per-step modelSettings and maxRetries
- #24450validate external workspace subpaths
- #24464reject truncated structured output
- #24467ignore redelivered idle signals on the claimed-owner wake path
- #24486Stop huge command output from crashing the process
- #24503cap implicit error-processor retries at 3 and warn
- #24566enforce a trusted tenant scope across trace queries
- #24510keep cross-agent threads reachable after /new, and stop /new aborting a peer's run
- #24519set resourceId on manually triggered workflow runs
- #24576restore review checkout trust before prompt creation
- #24607emit each score once in delta polling after retried inserts
- #24613stop a persisted idle signal from starving the event loop
- #24636honor configured timeout when reconnecting to or resuming a sandbox
- #24649place GitHub cards by label on arrival, reconcile, and PR review
- #24638prevent background process completion failures from crashing the host
- #24658keep filled buttons opaque so board pills stack right
- #24679Count and trim the provider prompt in TokenLimiterProcessor
- #24691keep every ButtonsGroup segment on one size rung
- #24707run on @mastra/mcp 2.x while still serving legacy hosts
- #24696give ButtonsGroup a size, and let it own the rung
- #24741make Inngest function-level retries configurable
- #24758keep InngestAgent durable across __fork and resume/tool approval
- #24742keep internal spans internal after rebuildSpan()
- #24763run tool calls in parallel when a function approval policy returns false
- #24842stop function-level retries from re-running failed steps
- #24858show your own sessions in the Factory sidebar by default
- #24907show sub-agent tool on delegated channel approval cards
- #24906name declared next stages in invalid_transition rejections
- #24909run notification and signal wakes as the Factory session's user
- #24923stop concurrent callers from creating duplicate observational memory generations
- #24914only show channel approval cards when a person needs to decide
- #25000make feedback review updates fast again
- #25026download built-in models from Hugging Face instead of the legacy Qdrant GCS bucket
- #25005start tools as soon as their tool-call args complete
- #25027run the shared auth pipeline and serve registerApiRoute routes