Skip to content

PocketContext Crew 0.7.0

Choose a tag to compare

@github-actions github-actions released this 07 Nov 22:39
Immutable release. Only release title and notes can be modified.

Crew 0.7.0

Crew wakes now cost far less context. A wake run reads a compact execution subset
of the capabilities, carries protocol and digest references in its payload and
follows a shorter managed block. Agents find work through filtered, projected
reads and ranked search, and answer reporting questions with bounded read-only SQL.
Owner browser comments can wake the assignee, and a per-task limit stops bot
handoff loops.

Owner-comment wakes and the handoff limit

An integration can opt in to the comment subscription. An owner comment from the
browser then wakes the assignee of an eligible task; the payload names the comment
event, which the recipient reads through Crew as owner input. A comment during live
execution is deferred until execution.complete. Comments written through MCP
never wake anyone, so bots cannot wake each other by commenting.

Consecutive shared-MCP reassignments are counted per task. Past
max_bot_handoffs (default 4, owner-only, 1-20) the write is applied, but the task
is set Waiting with a handoff_limited marker and an audit event, and no wake is
sent. Shared MCP cannot set it Ready again; the owner reviews it and sets Ready,
which resets the count.

Compact wake protocol

crew_read {collection: "capabilities", id: "execution"} returns only the
operations used by routine, parent, dispatcher and recipient turns, with the
protocol version, write envelope, data._crew rule, limits and the same
capabilities_digest as the full document. Every task.wake and webhook.test
payload carries protocol_version and capabilities_digest; a turn that already
holds that digest skips the read, and the dispatcher forwards it to the recipient.

The managed block keeps protocol 3 and every earlier rule in fewer bytes. Wake
runs take bridge and execution revisions from the previous result and reread once
only on revision_conflict; the server still checks every revision. Native task
wakes no longer read webhook_configs. A later instruction revision adds two
retrieval rules: protocol steps never use search, schema or SQL, and task work
reads compactly. A 409 tool error now suggests a reread only for a genuine
revision conflict; refusals such as handoff_limited or execution_required say
they must not be retried unchanged.

Compact reads, ranked search and read-only SQL

crew_read accepts limit, fields projection, task filters (assigned_agent,
status, readiness, label) and changed_since. Projection rejects unknown or
redacted fields. crew_search ranks notes and tasks with SQLite FTS5 BM25 and
returns short hits without full bodies; edits, restores and archive changes are
visible immediately.

crew_schema lists SQL views and crew_query runs one read-only statement on a
separate read-only connection to the private working copy. Each view is an
explicit allowlist of columns crew_read already returns; webhook URLs and
credentials, fencing handles, auth records, OAuth state and PocketBase internals
are excluded. An SQLite authorizer allows only SELECT and a fixed function list.
Results are capped at 2 seconds, 1000 rows and 1 MiB. SQL views include the new
wake and handoff fields. Use SQL for counts, joins, aggregates and history, and
compact reads or search for ordinary lookups.

Measured results

A development instance with real Grok bots in parent_bridge mode, compared with
the 0.6.0 test baseline:

Measure 0.6.0 0.7.0
Parent-bridge handoff protocol 14 calls / ~190 KB 8 calls / 14.5 KB
Full handoff task with a memory lookup 13 calls / 18.4 KB
Self-work protocol 9 calls / ~126 KB 5 calls / ~10.5 KB
Managed block 6,659 B 3,986 B
Capabilities read 48.8 KB full ~5.5 KB execution subset

A synthetic retrieval benchmark (500 notes, 300 tasks) returned identical answers
with 273 times fewer bytes for "ready work for an agent" and 611 times fewer for
"counts by status and assignee". These numbers describe those runs, not general
provider behavior.

Upgrade

Take a complete stopped backup, including the private webhook key. The migrations
add the task fields and set max_bot_handoffs to 4 without changing
subscriptions, then install two instruction revisions in order. Dispatcher,
dispatch mode, mappings, credentials and history are retained; old work
generations are fenced and delivery pauses at the latest instruction revision. No
tests or backlog are sent. Refresh connector tool schemas, apply the current
managed block to the routine once, run an explicit connection test (with a mapped
teammate in parent_bridge mode), then re-enable delivery. Opt in to comment if
wanted; any integration, subscription or credential change pauses delivery and
fences in-flight work, so change it between runs and re-enable. Update installed
skills with the release. See upgrades.

Compatibility

Protocol 3 is unchanged, and so are the write envelope, execution fencing, the
data._crew path for cached connectors and the owner-only review boundary.
Omitting the new read options keeps earlier crew_read behavior, and the full
capabilities document is still returned for any other id. SQL is read-only and
the new tools never return webhook URLs, credentials or redacted fields. A
0.6.0 binary must not open the migrated database; roll back only with the
complete pre-upgrade backup.

Verification

Synthetic tests cover comment wakes and deferral, the handoff limit, migration
from the 0.6.0 layout, protocol overhead per flow, rule retention in the compact
block, read filters and projections, search freshness and archive semantics, SQL
allowlists, rejected secret columns, mutations and limits, and the retrieval
benchmark. On the live development instance the following were verified: owner
comment wake, deferred comment wake, MCP comments never waking, the handoff limit
stop (write applied, task Waiting, audit event, no wake), the owner-only limit,
and SQL rejection of secret columns and mutations. Other accounts still require
live validation. See webhook setup and MCP.


PocketContext Crew 0.7.0: shared agent memory, tasks and human review in one binary.

Linux amd64 and arm64: static musl binaries. macOS arm64: macOS 15 or later, ad hoc signed. Windows amd64: statically linked MinGW build, unsigned.

Install on Linux or macOS into the project folder: curl -fsSL https://github.com/pocketcontext-labs/pocketcontext-crew/releases/download/v0.7.0/install.sh | sh -s -- /absolute/project/folder

Install on Windows (PowerShell): $env:CREW_INSTALL_DIR = 'C:\absolute\project\folder'; irm https://github.com/pocketcontext-labs/pocketcontext-crew/releases/download/v0.7.0/install.ps1 | iex

Both installers require the target folder, usually the project folder beside pocketcontext.db (it may be inside iCloud Drive, OneDrive, Dropbox or Google Drive).

See the README.