Take stuck work back out of Symphony's queue

[tptodorov]

S refused to unschedule any ticket Symphony held a session for. That is right while an agent is running, but wrong for the two states where nothing is progressing — blocked (waiting for an operator) and retrying (backing off after failing to start). Those are precisely the tickets you need to pull back, fix, and put in again, and they were the ones being refused.

Why removing the label is enough

The previous reasoning was that unscheduling "would not stop the work, only make the dashboard disagree with what is happening". Checked against Symphony's orchestrator rather than assumed, that does not hold for either stuck state:

• retrying — when the retry timer fires, Symphony re-fetches the issue and re-tests retry_candidate_issue? → Issue.routable?(issue, required_labels). With the label gone that is false, so it falls through to release_issue_claim ("Issue left active states, removing claim").
• blocked — reconcile_blocked_issue_state has an explicit !issue_routable? → release_issue_claim branch ("Blocked issue no longer routed to this worker").

Both re-read the required labels before acting, so the claim really is dropped.

running is still refused: no label change interrupts a turn already in progress. That needs stopping the session in Symphony, whose API is read-only apart from POST /api/v1/refresh, so devdash cannot do it for you. symphonyHasIt becomes symphonyIsWorkingOnIt to name the state it actually means.

Also in here

• The long-form ? help still described the old rule and contradicted the new behaviour — corrected alongside the key table in the help, the in-app view, and the README.
• The README marker table gains the grey scheduled marker it never listed.
• A Stop hook (.claude/settings.json) builds, tests, and installs to ~/.local/bin only when both pass, so the binary on PATH matches the source. Its deploy path is machine-specific; .claude/settings.local.json is the untracked home for it if that becomes awkward.

Verification

• gofmt -l . clean, go vet ./... clean, go test -race -cover ./... passes (53.1% of statements).
• TestCannotUnscheduleWhileAnAgentIsRunning covers the refusal; TestCanUnscheduleStuckWork covers blocked, retrying, scheduled and no-session, asserting each removes the label and leaves the status untouched.
• Confirmed against a live Symphony on port 10000 that /api/v1/state really does return a blocked array — worth checking, since Symphony's SPEC.md does not document one and the Go implementation does not emit it, so on that backend blocked would never appear.
• The hook's failure path was exercised with a deliberate syntax error: it reported the compile error and left the previously installed binary untouched.

Not covered: the live dry-run harness (DEVDASH_DRYRUN_DIR=... go test -run TestDryRunSchedulePlans) needs JIRA credentials that were not available, so the plan for every real ticket has not been eyeballed.

🤖 Generated with Claude Code