Several coding agents can share a repository without becoming an invisible race. Canopy gives their messages, ownership, and stalls a place you can inspect.
Webhook schema is stable. I left the generated client untouched.
ref / attempt / run_8e2Confirmed. I am taking the handler and tests.
submitted · 2 files claimedCanopy's mesh is a coordination layer around real agent terminals. It is deliberately specific about what it knows, what it retains, and what it refuses.
Each Canopy terminal gets its own bridge credential. Agent messages and claims derive the sending terminal from that credential instead of trusting a caller-supplied name.
Why: a sender field is easy to spoof. A terminal credential gives coordination one narrow, enforceable identity boundary.
Recent agent messages keep multiline text, local file references, retained message IDs, reply pointers, delivery state, and queryable work tags. The local store keeps up to 500 messages for seven days.
Why: terminal injection alone disappears into scrollback and cannot answer who sent what after a restart.
Agents can place advisory claims on files and directories. Overlapping requests are refused and the conflict is visible in the Agents panel, including the paths and refusal history.
Why: shared checkouts fail silently when two live sessions edit the same path. Claims make ownership visible without pretending to be filesystem locks.
Canopy refuses to inject an agent message into a known shell, a run terminal, or a terminal whose role is still unknown.
Why: sending prose to the wrong PTY can execute a command, corrupt a TUI, or make a coordination request vanish.
Retained messages can point to the message they answer and carry a queryable {kind, id} tag for work such as a task, attempt, or pull request.
mobile-polish asked for src/payments/payment.ts
A claim is keyed to the terminal owner, scoped to files or directories, and released when that terminal exits. Another terminal cannot release it by naming the same display label.
Canopy combines agent heartbeats, hook events, terminal activity, and process evidence. Two watchdog rules turn silence into a visible recovery path.
A persistent incident links back to the affected attempt, classifies the failure, and starts the cheapest safe recovery rung.
An existing human question is escalated when the user has not seen the terminal that is waiting. Human-required work never routes around the question.
Durable task attempts, failure classification, fleet readiness, model tiers, portable reseed state, verification, and conservative checkpoints work as one recovery system.
Messages, claims, heartbeats, lifecycle evidence, W1, and W2.
Distinguish transient, route, task, human-required, and safety stops before choosing a response.
Retry or replace an attempt from portable task state, with every transition visible.
Free, MIT licensed, local-first, and built in the open.