Add CrewRegistryConfig to RigEntry allowing crew members to be defined
in rigs.json and synced across machines. The new `gt crew sync` command
creates missing crew members from the configuration.
Configuration example:
"rigs": {
"gastown": {
"crew": {
"theme": "mad-max",
"members": ["diesel", "chrome", "nitro"]
}
}
}
Closes: gt-tu4
Replace bd subprocess spawns with direct SQLite queries:
- queryEpicsInDir: direct sqlite3 query vs bd list subprocess
- getLinkedConvoys: direct JOIN query vs bd dep list + getIssueDetails loop
- computeGoalLastMovement: reuse epic.UpdatedAt vs separate bd show call
Also includes mailbox optimization from earlier session:
- Consolidated multiple parallel queries into single bd list --all query
- Filters in Go instead of spawning O(identities × statuses) bd processes
177x improvement (6.2s → 35ms) by eliminating subprocess overhead.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Replace bd subprocess calls in gt commands with daemon RPC when available.
Each subprocess call has ~40ms overhead for Go binary startup, so using
the daemon's Unix socket protocol significantly reduces latency.
Changes:
- Add RPC client to beads package (beads_rpc.go)
- Modify List/Show/Update/Close methods to try RPC first, fall back to subprocess
- Replace runBdPrime() with direct content output (avoids bd subprocess)
- Replace checkPendingEscalations() to use beads.List() with RPC
- Replace hook.go bd subprocess calls with beads package methods
The RPC client:
- Connects to daemon via Unix socket at .beads/bd.sock
- Uses JSON-based request/response protocol (same as bd daemon)
- Falls back gracefully to subprocess if daemon unavailable
- Lazy-initializes connection on first use
Performance improvement targets (from bd-2zd.2):
- gt prime < 100ms (was 5.8s with subprocess chain)
- gt hook < 100ms (was ~323ms)
Closes: bd-2zd.2
The previous approach using KillPaneProcessesExcluding/KillPaneProcesses
killed the pane's main process (Claude/node) before calling RespawnPane.
This caused the pane to close (since tmux's remain-on-exit is off by default),
which then made RespawnPane fail because the target pane no longer exists.
The respawn-pane -k flag handles killing atomically - it kills the old process
and starts the new one in a single operation without closing the pane in between.
If orphan processes remain (e.g., Claude ignoring SIGHUP), they will be cleaned
up when the new session starts or by periodic cleanup processes.
This fixes both self-handoff and remote handoff paths.
Fixes: hq-bv7ef
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The `gt plugin run` command was recording a "success" run even though it
only prints plugin instructions for an agent/user to execute - it doesn't
actually run the plugin.
This poisoned the cooldown gate: CountRunsSince counted these false
successes, preventing actual executions from running because the gate
appeared to have recent successful runs.
Remove the recording from `gt plugin run`. The actual plugin execution
(by whatever follows the printed instructions) should record the result.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Under high concurrency (17+ agents), the bd version check spawns
multiple git subprocesses per invocation, causing timeouts when
85-120+ git processes compete for resources.
This fix:
- Caches successful version checks to ~/.cache/gastown/beads-version.json
- Uses cached results for 24 hours to avoid subprocess spawning
- On timeout, uses stale cache if available or gracefully degrades
- Prints warning when using cached/degraded path
Fixes: https://github.com/steveyegge/gastown/issues/503
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The Stop hook with `gt costs record` was causing 30-second timeouts
on every session stop due to beads socket connection issues. Since
cost tracking is disabled anyway (Claude Code doesn't expose session
costs), this hook provided no value.
Changes:
- Remove Stop hook from settings-autonomous.json and settings-interactive.json
- Remove Stop hook validation from claude_settings_check.go
- Update tests to not expect Stop hook
The cost tracking infrastructure remains in costs.go for future use
when Claude Code exposes session costs via API or environment variable.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add assignee display to both list and single-goal views. In list view,
assignee appears on the second line when present. In single-goal view,
it appears as a dedicated field after priority. JSON output also includes
the assignee field.
Closes: gt-libj
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The getCurrentWork function was returning ANY in_progress bead from the
workspace rather than only beads assigned to the current agent. This caused
crew workers to see wisps assigned to polecats in their status bar.
Changes:
- Add identity parameter to getCurrentWork function
- Add identity guard (return empty if identity is empty)
- Filter by Assignee in the beads query
This complements the earlier getHookedWork fix and ensures both hooked
AND in_progress beads are filtered by the agent's identity.
Fixes gt-zxnr (additional fix).
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Root cause: tmux statusline showed wrong hook for all java crewmembers
because GT_CREW env var wasn't set in tmux session environment.
Changes:
- statusline.go: Add early return in getHookedWork() when identity is empty
to prevent returning ALL hooked beads regardless of assignee
- crew_at.go: Call SetEnvironment in the restart path so sessions created
before GT_CREW was being set get it on restart
Fixes gt-zxnr.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
KillPaneProcesses was killing ALL processes in the pane, including the
gt handoff process itself. This created a race condition where the
process could be killed before RespawnPane executes, causing the pane
to close prematurely and requiring manual reattach.
Added KillPaneProcessesExcluding() function that excludes specified PIDs
from being killed. The handoff command now passes its own PID to avoid
the race condition.
Fixes: gt-85qd
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
When deacon patrol molecules completed, their child step wisps were not being
closed automatically. This caused orphan wisp accumulation - 143+ orphaned
wisps were found in one cleanup session.
The fix ensures that when a molecule completes (via gt done or gt mol step done),
all descendant step issues are recursively closed before the molecule itself.
Changes:
- done.go: Added closeDescendants() call in updateAgentStateOnDone before
closing the attached molecule
- molecule_step.go: Added closeDescendants() call in handleMoleculeComplete
for all roles (not just polecats)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The Deacon patrol formula now clearly documents the continuous loop:
1. Execute patrol steps (inbox-check through context-check)
2. Squash wisp, wait for activity via await-signal (15min max)
3. Create new patrol wisp and hook it
4. Repeat from step 1
Changes:
- Formula description emphasizes CONTINUOUS EXECUTION with flow diagram
- loop-or-exit step renamed to "Continuous patrol loop" with explicit
instructions for creating/hooking new wisps after await-signal
- plugin-run step now clearly shows gt plugin list + gt dog dispatch
- Deacon role template updated to match formula changes
- Formula version bumped to 9
Fixes gt-fm2c: Deacon needs continuous patrol loop for plugin dispatch
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
gt goals was only querying the default beads location (town-level
with hq- prefix), missing epics from rig-level beads (j-, sc-, etc.).
Now iterates over all rig directories with .beads/ subdirectories
and aggregates epics, deduplicating by ID.
Wisp molecules (gt-wisp-* IDs, mol-* titles) are transient operational
beads for witness/refinery/polecat patrol, not strategic goals that
need human attention. These are now filtered by default.
Add --include-wisp flag to show them when debugging.
Fixes gt-ysmj
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add `bd tree <id>` to Key Commands in bd prime template (beads.go)
- Add `bd tree <issue>` to prime_output.go for mayor/polecat/crew roles
- Helps agents understand bead ancestry, siblings, and dependencies
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add three new flags for filtering convoys by epic relationship:
- --orphans: show only convoys without a parent epic
- --epic <id>: show only convoys under a specific epic
- --by-epic: group convoys by parent epic
These support the Goals Layer feature (Phase 3) for hierarchical
focus management.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Implements gt goals command to show epics sorted by staleness × priority.
Features:
- List all open epics with staleness indicators (🟢/🟡/🔴)
- Sort by attention score (priority × staleness hours)
- Show specific goal details with description and linked convoys
- JSON output support
- Priority and status filtering
Staleness thresholds:
- 🟢 active: moved in last hour
- 🟡 stale: no movement for 1+ hours
- 🔴 stuck: no movement for 4+ hours
Closes: gt-vix
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Create goals.go with basic command structure for viewing strategic
goals (epics) with staleness indicators. Includes --json, --status,
and --priority flags. Implementation stubs return not-yet-implemented
errors.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add explicit guidance on the Mayor → Crew → Polecats delegation model:
- Crew are coordinators for epics/goals needing decomposition
- Polecats are executors for well-defined tasks
- Include decision framework table for work type routing
Closes: gt-9jd
Implements the Overseer Experience epic (gt-k0kn):
- gt focus: Shows stalest high-priority goals, sorted by priority × staleness
- gt attention: Shows blocked items, PRs awaiting review, stuck workers
- gt status: Now includes GOALS and ATTENTION summary sections
- gt convoy list: Added --orphans, --epic, --by-epic flags
These commands reduce Mayor bottleneck by giving the overseer direct
visibility into system state without needing to ask Mayor.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add explicit guidance on the Mayor → Crew → Polecats delegation model:
- Crew are coordinators for epics/goals needing decomposition
- Polecats are executors for well-defined tasks
- Include decision framework table for work type routing
Closes: gt-9jd
Adds clear guidance that crew members are coordinators, not implementers:
- Lists 4 key responsibilities: Research, Decompose, Sling, Review
- Clarifies "goal-specific mayor" model - own outcomes via delegation
- Documents when to implement directly vs delegate (trivial fixes, spikes, etc.)
Closes: gt-gig
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add --convoy flag to gt sling that allows adding an issue to an existing
convoy instead of creating a new one. When specified:
- Validates the convoy exists and is open
- Adds tracking relation between convoy and issue
- Skips auto-convoy creation
Changes:
- Add slingConvoy variable and --convoy flag registration
- Add addToExistingConvoy() helper function in sling_convoy.go
- Modify auto-convoy logic to check slingConvoy first
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Add flag variable declarations and Cobra flag registrations for:
- --epic: link auto-created convoy to parent epic
- --convoy: add to existing convoy instead of creating new
Closes: gt-n3o
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Change molecule step completion instructions to use `gt mol step done`
instead of `bd close`. This ensures polecats get fresh context between
each step, which is critical for multi-step review workflows like
shiny-enterprise where each refinement pass should have unbiased attention.
The `gt mol step done` command already:
1. Closes the step
2. Finds the next ready step
3. Respawns the pane for fresh context
But polecats were being instructed to use `bd close` directly, which
skipped the respawn and let them run through entire workflows in a
single session with accumulated context.
Updated:
- prime_molecule.go: step completion instructions
- mol-polecat-work.formula.toml
- mol-polecat-code-review.formula.toml
- mol-polecat-review-pr.formula.toml
Fixes: hq-0kx7ra
Three fixes to make dog dispatch work end-to-end:
1. Add BuildDogStartupCommand in loader.go
- Similar to BuildPolecatStartupCommand/BuildCrewStartupCommand
- Passes AgentName to AgentEnv so BD_ACTOR is exported in startup command
2. Use BuildDogStartupCommand in dog.go
- Removes ineffective SetEnvironment calls (env vars set after shell starts
don't propagate to already-running processes)
3. Add "dog" case in mail_identity.go detectSenderFromRole
- Dogs now use BD_ACTOR for mail identity
- Without this, dogs fell through to "overseer" and couldn't find their mail
Tested: dog alpha now correctly sees inbox as deacon/dogs/alpha
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Recovered from reflog - these commits were lost during a rebase/force-push.
Dogs are directories with state files but no sessions. When `gt dog dispatch`
assigned work and sent mail, nothing executed because no session existed.
Changes:
1. Spawn tmux session after dispatch (gt-<town>-deacon-<dogname>)
2. Set BD_ACTOR=deacon/dogs/<name> so dogs can find their mail
3. Add dog case to AgentEnv for proper identity
Session spawn is non-blocking - if it fails, mail was sent and human can
manually start the session.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Dogs can now reset their own state to idle after completing work:
gt dog done # Auto-detect from BD_ACTOR
gt dog done alpha # Explicit name
This solves the issue where dog sessions would complete work but remain in
"working" state because nothing processed the DOG_DONE mail. Now dogs can
explicitly mark themselves idle before handing off.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Fixes intermittent 'timeout waiting for runtime prompt' errors that occur
when Claude takes longer than 60s to start under load or on slower machines.
Resolves: hq-j2wl
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Phase 1 of agent security model: Set distinct email addresses for each
agent type to improve audit trail clarity.
Email format:
- Town-level: {role}@gastown.local (mayor, deacon, boot)
- Rig-level: {rig}-{role}@gastown.local (witness, refinery)
- Named agents: {rig}-{role}-{name}@gastown.local (polecat, crew)
This makes git log filtering by agent type trivial and provides a
foundation for per-agent key separation in future phases.
Refs: hq-biot
Mayor now checks `gt escalate list` between hook and mail checks at startup.
This ensures pending escalations from other agents are handled promptly.
Other roles (witness, refinery, polecat, crew, deacon) are unaffected -
they create escalations but don't handle them at startup.
The daemon's isRigOperational() was only checking wisp config layer
for docked/parked status. However, 'gt rig dock' sets the status:docked
label on the rig identity bead (global/synced), not wisp config.
This caused the daemon to auto-restart agents on docked rigs because
it couldn't see the bead-level docked status.
Fix:
- Check rig bead labels for status:docked and status:parked
- Also updated mol-deacon-patrol formula to explicitly skip
DOCKED/PARKED rigs in health-scan step
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
TestRoleAgentConfigWithCustomAgent tests custom agent "opencode-mayor"
which uses the opencode binary. Without the stub, the test falls back
to default and fails assertions about prompt_mode and env vars.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Extract platform-specific syscall code into proc_unix.go and proc_windows.go
with appropriate build tags. This fixes TestCrossPlatformBuild which failed
because syscall.SysProcAttr.Setpgid is Unix-only.
Changes:
- setSysProcAttr(): Sets process group on Unix, no-op on Windows
- isProcessAlive(): Uses Signal(0) on Unix, nil signal on Windows
- sendTermSignal(): SIGTERM on Unix, Kill() on Windows
- sendKillSignal(): SIGKILL on Unix, Kill() on Windows
Fixes gt-3078ed
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
gt hook status failed with "cannot determine agent identity (role: unknown)"
when run from rig root directories like ~/gt/beads. The cwd-based detection
only recognizes specific agent directories (witness/, polecats/foo/, etc.)
but not rig roots.
Now falls back to GT_ROLE env var when detectRole returns unknown, matching
the pattern used by GetRoleWithContext and other role detection code.
Fixes gt-6pg
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
All agents now receive their startup beacon + role-specific instructions via
the CLI prompt, making sessions identifiable in /resume picker while removing
unreliable post-startup nudges.
Changes:
- Rename FormatStartupNudge → FormatStartupBeacon, StartupNudgeConfig → BeaconConfig
- Remove StartupNudge() function (no longer needed)
- Remove PropulsionNudge() and PropulsionNudgeForRole() functions
- Update deacon, witness, refinery, polecat managers to include beacon in CLI prompt
- Update boot to inline beacon (can't import session due to import cycle)
- Update daemon/lifecycle.go to include beacon via BuildCommandWithPrompt
- Update cmd/deacon.go to include beacon in startup command
- Remove redundant StartupNudge and PropulsionNudge calls from all startup paths
The beacon is now part of the CLI prompt which is queued before Claude starts,
making it more reliable than post-startup nudges which had timing issues.
SessionStart hook runs gt prime automatically, so PropulsionNudge was redundant.
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
When contributing to upstream repos or wanting human review before merge,
the --no-merge flag keeps polecat work on a feature branch instead of
auto-merging to main.
Changes:
- Add --no-merge flag to gt sling command
- Store no_merge field in bead's AttachmentFields
- Modify gt done to skip merge queue when no_merge is set
- Push branch and mail dispatcher "READY_FOR_REVIEW" instead
- Add tests for field parsing and sling flag storage
Closes: gt-p7b8
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
* fix(costs): read token usage from Claude Code transcripts instead of tmux
Replace tmux screen-scraping with reading token usage directly from Claude
Code transcript files at ~/.claude/projects/*/. This enables accurate cost
tracking by summing input_tokens, cache_creation_input_tokens,
cache_read_input_tokens, and output_tokens from assistant messages.
Adds model-specific pricing for Opus 4.5, Sonnet 4, and Haiku 3.5 to
calculate USD costs from token counts.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* fix(costs): correct Claude project path encoding
The leading slash should become a leading dash, not be stripped.
Claude Code encodes /Users/foo as -Users-foo, not Users-foo.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
* fix(costs): make --by-role and --by-rig fast by defaulting to today's costs
When using --by-role or --by-rig without a time filter, the command was
querying all historical events from the beads database via expensive bd list
and bd show commands, taking 10+ seconds and returning no data.
Now these flags default to today's costs from the log file (same as --today),
making them fast and showing actual data. This aligns with the new log-file-based
cost tracking architecture.
Co-Authored-By: Claude Haiku 4.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
The polecat manager was using rig/name format while sling sets
rig/polecats/name. This caused gt polecat status to show 'done'
for working polecats because GetAssignedIssue couldn't find them.
Fixes gt-oohy
* fix(config): preserve all RuntimeConfig fields in fillRuntimeDefaults
fillRuntimeDefaults was only copying Command, Args, InitialPrompt, and Env
when creating a copy of RuntimeConfig. This caused fields like PromptMode,
Provider, Session, Hooks, Tmux, and Instructions to be silently dropped.
This broke custom agent configurations, particularly prompt_mode: "none"
which is needed for agents like opencode that don't accept a startup beacon.
Changes:
- Copy all RuntimeConfig fields in fillRuntimeDefaults
- Add comprehensive tests for fillRuntimeDefaults
- Add integration tests for custom agent configs with prompt_mode: none
- Add tests mirroring manual verification: claude-opus, amp, codex, gemini
* fix(config): deep copy slices and nested structs in fillRuntimeDefaults
The original fix preserved all fields but used shallow copies for slices
and nested structs. This could cause mutation of the original config.
Changes:
- Deep copy Args slice (was sharing backing array)
- Deep copy Session, Hooks, Tmux, Instructions structs (were pointer copies)
- Deep copy Tmux.ProcessNames slice
- Add comprehensive mutation isolation tests for all fields
- Fix TestMultipleAgentTypes to test actual built-in presets
- Add TestCustomClaudeVariants to clarify that claude-opus/sonnet/haiku
are NOT built-in and must be defined as custom agents
Built-in presets: claude, gemini, codex, cursor, auggie, amp, opencode
Custom variants like claude-opus need explicit definition in Agents map.
* docs(test): add manual test settings to TestRoleAgentConfigWithCustomAgent
Document the settings/config.json used for manual verification:
- default_agent: claude-opus
- Custom agents: amp-yolo, opencode-mayor (with prompt_mode: none)
- role_agents mapping for all 6 roles
- Manual test procedure for all 7 built-in agents
* fix(test): address CodeRabbit review feedback
- Fix isClaudeCommand to handle Windows paths and .exe extension
- Use isClaudeCommand helper instead of brittle equality check
- Add skipIfAgentBinaryMissing to tests that depend on external binaries
(TestMultipleAgentTypes, TestCustomAgentWithAmp)
* fix: always send handoff mail
* fix: remove trailing slash from mayor in detectSenderFromCwd
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
Add functions to wake Claude Code's event loop in detached tmux sessions:
- IsSessionAttached: Check if session has attached clients
- WakePane: Always trigger SIGWINCH via resize dance
- WakePaneIfDetached: Smart wrapper that skips attached sessions
When Claude runs in a detached tmux session, its TUI library may not
process stdin until a terminal event occurs. Attaching triggers SIGWINCH
which wakes the event loop. WakePane simulates that by resizing the pane
down 1 row then back up.
NudgeSession and NudgePane now call WakePaneIfDetached after sending
Enter, covering all 22 nudge call sites in the codebase.
Fixes: gt-6s75ln
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
The procps-ng kill binary (version 4.0.4+) has an argument parsing issue
where negative PIDs like "-12345" are misinterpreted as options because
they start with a dash. This causes the binary to fall back to "-1" which
means "all processes", resulting in kill(-1, signal) being called.
This was discovered through an extensive murder investigation after 26
Claude AI instances were killed when running TestCleanupOrphanedSessions.
Each investigator left notes for their successor, eventually leading to
the discovery that:
- exec.Command("bash", "-c", "kill -KILL -PGID") is SAFE (uses bash builtin)
- exec.Command("kill", "-KILL", "-PGID") is FATAL (uses /usr/bin/kill)
Verified via strace:
/usr/bin/kill -KILL -12345 → kill(-1, SIGKILL) # WRONG!
/usr/bin/kill -KILL -- -12345 → kill(-12345, SIGKILL) # Correct!
The fix adds "--" before negative PGID arguments to explicitly end option
parsing, ensuring the negative number is treated as a PID/PGID argument.
Full investigation: https://github.com/groblegark/gastown/blob/main/MURDER_INVESTIGATION.md
Co-authored-by: Refinery <matthew.baker@pihealth.ai>
Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
