Repository navigation
PocketContext Crew 0.12.0
Crew 0.12.0
One capability set for the reader, MCP and the CLI. Work clients (MCP crew_write and
the CLI write command) can now do everything the owner can do in the reader, except
enter or read webhook URLs and secrets. Who may do what is decided by owner-editable
rules in the skill and MCP instructions, not by server refusals. There is no data format
or protocol version change. The capabilities document changes, so capabilities_digest
changes, and one managed instruction revision pauses webhook delivery until the
dispatcher re-syncs.
Behavior changes
These operations no longer return HTTP 403 on POST /api/crew/agent-write:
review.decideexecution.recoverarchivedonagent.create/agent.update(deactivate or reactivate an agent) and on
bucket.create/bucket.update(archive or restore a bucket)max_bot_handoffsonwebhook.routing.updatereadiness: "ready"on a task stopped by the bot handoff limit
webhook.credentials.update stays reader-only, is still refused before idempotency
replay, and reads still never return the URL or secret. MCP still cannot reach the
reader route. In capabilities every operation except webhook.credentials.update now
has authority: "shared_mcp", and owner_only_fields is gone from
webhook.routing.update.
Execution proof is narrower. A work-client task write (task.update, task.claim,
task.comment, review.request, review.withdraw, review.decide) needs proof only
while a live execution holds the task: dispatching or running with an unexpired lease,
or handed_off before its acknowledgment deadline. Without a live execution the write
applies, even when execution_required_generation is still set. Proof that is supplied
is still checked, and obsolete or mismatched proof is refused with execution_conflict.
A stale worker still cannot write during another worker's live run. The special case that
let agents archive a done governed task is gone, because agents can now archive and
restore governed tasks whenever nothing is live. The handoff_limited refusal code is
retired.
Setting Waiting clears execution_required_generation for every caller, not just the
reader.
A work client that sets a hop-limited task Ready does not reset bot_handoffs or
handoff_limited, so the next work-client reassignment stops the task again. Only a
reader write resets the count. This is intended.
These stay the same, as loop prevention and audit: work-client comments never wake
anyone, work-client reassignments count toward max_bot_handoffs, and events record an
owner: or mcp: actor plus the unverified reported_agent. Attribution is never used to
block.
The review.request bug
Before, a worker that requested review lost the fence it needed afterwards: the request
invalidated its execution, and every unproved write was then refused. Now nothing is
live after the request, so that worker can comment on, edit, archive or restore the task
without proof. execution.complete on the invalidated execution is still refused, and so
are writes that carry its old proof. The worker should drop the proof and write normally.
Default authorization rules
SKILL.md, SKILL.remote.md and the MCP instructions now include an editable
Authorization section. Unless the owner changes it or overrides it in a conversation:
- Decide reviews only when the owner explicitly delegates the decision in the
conversation. - Change
max_bot_handoffsonly when the owner asks. - Archive or restore buckets, and deactivate or reactivate agents, only when the owner
asks. - Recover executions, set Waiting, or set a hop-limited task Ready when the owner asks
or when work is stuck and the owner would otherwise do it in the reader. If unsure,
report first. - Anything that would post, send, publish, invite, pay or ship needs the owner's
authorization in the conversation. An approval never executes it. - Set
reported_agentto your own agent name.
The OAuth consent page, CLI help, MCP tool text and docs were updated to match.
"Owner path" and "agent path" are now called reader writes and work-client writes.
Managed dispatcher instructions
A migration installs a new instruction revision. "bypass human review" becomes "Never
decide reviews", because work clients can now call review.decide and a wake run never
carries the owner's delegation. The protocol stays at 3. Like earlier instruction
upgrades, the migration keeps the dispatcher, mode, mappings, credentials and history,
moves Ready tasks to a new generation and pauses delivery. Hosted dispatchers must
re-sync the managed block before delivery is re-enabled.
Accepted residual risks
A prompt-injected or non-compliant agent could now approve its own draft, archive a
bucket, deactivate an agent, recover or cancel an execution, raise max_bot_handoffs,
or set a hop-limited task Ready. History is append-only, so all of these can be audited
afterwards through actor and reported_agent. External effects stay behind other
systems' logins and the outbound rule. The server still enforces two things: webhook
credentials, and loop prevention (silent work-client comments and hop counting). On
installs without a domain, a CLI agent holds the owner's session and could still call
the reader route with raw HTTP. The skill forbids this, and MCP OAuth clients cannot
reach that route.
Upgrade
Take a complete stopped backup, including the private webhook key. Then replace the
0.11.1 binary or image with 0.12.0 following the upgrade procedure,
upgrade paired clients with the server, and run skill to install the new guidance.
Clients read capabilities again, because the digest changed. Re-sync the dispatcher
routine to the new managed block, run the connection test and re-enable delivery.
Rolling back needs the pre-upgrade backup, because an older binary must not open the
migrated database.
PocketContext Crew 0.12.0: 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/v0.12.0/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/v0.12.0/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.