| name | state-ops |
| description | Inspect anvil project state — list tasks, show task details, find the next claimable task, summarize active claims and blockers, check file-conflict warnings, and reconcile state with the filesystem and git. Use this skill when you want to see what anvil knows without changing anything. |
State-Ops — State Inspection
Read the canonical SQLite state that anvil maintains: project summary, task inventory, claim activity, dependency graph, conflict warnings, and filesystem reconciliation. This skill makes no mutations.
When to Use
- When orienting before starting a work session — run
status first, always.
- Before claiming a task — confirm the PRD has passed review (
reviewed or approved) and the task is ready.
- When a claim was interrupted and the state of the queue is unclear.
- When multiple agents are active and conflict risk is non-trivial.
- When a task shows as
blocked and the blocker chain needs tracing.
- When suspicious that orphan branches or stale packets exist on disk.
- When reporting project progress to a human or another skill.
State-ops is NOT for authoring or reviewing PRDs — use the prd skill. State-ops is NOT for generating plans or scoring tasks — use the plan skill. State-ops is NOT for claiming work or submitting evidence — use the claim and execute skills. This skill only reads.
Prerequisites
The project must have run anvil init at least once. State lives in the layout anvil chose (the HOME workspace by default, ~/.anvil/workspaces/<key>/.anvil/...), not in the repo, so check init through the CLI rather than a literal path:
anvil status >/dev/null 2>&1 || echo "MISSING: run anvil init first"
If it reports MISSING, refuse to proceed and tell the caller to run:
anvil init --name "<project-name>"
Do not attempt to read state, list tasks, or call any other anvil command until anvil status confirms the project is initialized.
Workflow
Step 1 — Get the project summary
anvil status
Plain anvil status prints a human-readable summary like:
anvil for "myproject" (id: myproject)
Path: /Users/you/.anvil/workspaces/myproject/.anvil
Initialized: 2026-06-22T06:19:23.815479Z
PRD: approved
Tasks: 2 total (2 ready, 0 in_progress, 0 blocked)
Active claims: 0
Sync: off
Schema: 6
The Path: line tells you where state actually lives for this project (the HOME
workspace by default, an in-repo .anvil/ only under ANVIL_STATE_LAYOUT=local).
To inspect a different project's state, pass --cwd <project-dir> — the flag
points at the project directory anvil derives the workspace from. There is no
--workspace flag: never point a flag at the ~/.anvil/workspaces/<key>/
path directly — that option does not exist (passing one prints No such option '--workspace'). Select state by project via --cwd, and read the resolved
location off the Path: line.
PRD: is one of none, draft, reviewed, or approved. The Tasks: line gives
the total plus the ready / in_progress / blocked counts.
Run this first. The output orients every subsequent decision — it answers whether work is even possible before touching any other command.
Key signals to look for:
PRD: draft — the claim gate is closed. No task can be claimed until the PRD reaches reviewed or approved. Proceed to the prd skill instead.Tasks: N total (0 ready, ...) — all tasks are either upstream of ready or already active. Check blockers or the plan skill.(... 0 blocked) becoming non-zero — identify which tasks are blocked before picking new work.Active claims: N — determine whether adding another claim creates conflict risk.
For a single-line machine-readable form (with the prd-status: / ready-tasks: /
blockers: / active-claims: tokens), use anvil status --hook-format — see
Hook-Friendly Output below.
Step 2 — List tasks by filter
anvil list [--open] [--summary] [--status STATUS] [--feature FEATURE_ID] [--type TYPE]
Prints a table with columns TaskID | Title | Status | Priority | Type | Score | Feature (the Score column is complexity/agent_suitability, e.g. 2/2, or unscored until anvil score runs):
TaskID Title Status Priority Type Score Feature
-------------------------------------------------------------------------------------
T001 Implement argument parsing ready high feature 2/2 F001
T002 Implement list output ready medium feature 2/2 F001
2 task(s) listed.
Filters:
--open — only unfinished tasks (hides terminal done/accepted;
rejected awaits rework and stays open). The fastest answer to "is there
anything left to do?".--summary — per-PRD rollup (PRD | Open | Total | Breakdown) instead of
one row per task; combine with --open to see only PRDs with open work.--status ready — tasks available to claim right now.--status in_progress — tasks currently under active claims.--feature F001 — all tasks scoped to a specific feature.--type feature|bugfix|refactor|modify — tasks of a single type.
Use this to audit plan coverage, pick the next work item by priority, or confirm a feature is fully drafted before raising it for review.
Step 3 — Drill into a specific task
anvil show TASK_ID
Example:
anvil show T001
Returns: full task detail in multi-section format — title, feature, status, priority, the six-dimension scores (complexity, parallelizability, context_load, blast_radius, review_risk, agent_suitability) with an explanation, dependencies, conflict groups, acceptance criteria, verification commands, the task Likely Files, active claims (agent, lease, heartbeat — (none) if unclaimed), and recent events. The task's narrative body is its description.
Run show before claiming to confirm:
- The acceptance criteria are concrete and independently verifiable.
complexity is under 4 (if 4+, expand first via the plan skill).- No active claim already covers this task.
- The
Likely Files do not overlap files currently claimed by another agent (cross-check with anvil conflicts).
Step 4 — Find the next claimable task
anvil next
Returns the single highest-priority claimable task without claiming it:
Next recommended task: T001
Title: Implement argument parsing
Priority: high
Complexity: 2
Run `anvil claim T001` to acquire the lease.
This is the standard agent-loop entry point. Run it instead of manually scanning list output when the goal is simply to find work. After next returns a task ID, run show TASK_ID to read the full detail before claiming.
If next finds nothing claimable, it prints that the queue is empty (exit 0 — an empty queue is not an error; with -q/--quiet it prints nothing and exits 3 instead). Check status — the queue is either empty, fully claimed, or PRD-gated.
Step 5 — Check for conflicts
anvil conflicts
Returns the persisted conflict groups — sets of tasks whose likely_files overlap. Each group lists its ID, the member task IDs, and the shared files:
1 conflict group(s):
CG-T001-T002: T001, T002
Tasks T001 and T002 share overlapping files: src/cli.py
These groups are produced by planning inference and persisted to the conflict_groups table; conflicts reads them back (read-only and deterministic). Run it proactively before claiming a task that touches shared files (e.g., a module's __init__.py, a schema file, or a config). A conflict warning at this step is cheaper than a merge conflict later.
anvil claim also warns on overlap at claim time, but conflicts gives the full picture across all groups, not just the one being attempted.
Step 6 — Reconcile with filesystem and git
anvil sync
Bare anvil sync runs reconciliation and returns a report of orphans — branches that exist in git but have no corresponding claim in anvil's state, work packets with no matching task, and claims with expired leases that were never force-released. On a clean project it prints:
Reconciliation scanned at 2026-06-22T06:21:17.805045+00:00
No discrepancies found.
For a purely local report (no external sync provider, always exits 0), anvil drift covers the same ground read-only: orphan branches, orphan worktrees, orphan packets, stale claims, and tasks whose plan-declared files have vanished from disk.
Run reconciliation after a session ends abruptly, after a force-push cleans up stale branches, or periodically during long-running projects to keep state clean.
To apply fixes:
anvil sync --fix
The --fix flag applies each suggested fix; in a non-interactive context it requires --yes to skip the confirmation prompt (confirm this is safe first). Run anvil sync without --fix first to read the report.
Hook-Friendly Output
The SessionStart hook (detect-state.sh) calls status in compact form for machine consumption:
anvil status --hook-format
Emits exactly one line:
active-claims:N ready-tasks:N blockers:N prd-status:STATUS
Where STATUS is one of: none (project initialized, no PRD yet), draft, reviewed, or approved. When the project itself is not initialized for this directory, --hook-format prints the single literal token uninitialized instead of the four-token line (and still exits 0, so a hook never fails the session).
Parse this format when reading status from a hook or another skill. Do not parse the human-readable status output — its layout may change. The --hook-format contract is stable.
Example of a healthy project at the start of an execute session:
active-claims:0 ready-tasks:4 blockers:0 prd-status:approved
Example of a PRD-gated project:
active-claims:0 ready-tasks:8 blockers:0 prd-status:draft
The second example blocks claiming even though ready-tasks is non-zero — the PRD gate is enforced by the claims manager, not by the ready-tasks count. A draft (or rejected) PRD blocks; reviewed or approved clears the gate.
Common Pitfalls
- Claiming before the PRD passes review. The claims manager enforces this gate —
anvil claim errors while the PRD is draft or rejected. A reviewed or approved PRD clears the gate. Run anvil status first and check the PRD: line before attempting any claim.
- Manually editing the state database. Do not open the SQLite
state.db directly to fix state. Every mutation should go through the CLI so the change is recorded in the event log. Manual edits produce state that cannot be replayed or audited.
- Assuming stale claims block the queue. Stale leases are detected and cleared automatically on the next CLI or MCP operation — no manual intervention is needed. Wait one cycle (run any
anvil command) and the task returns to ready.
- Confusing
conflicts (file overlap) with blockers (dependency blockers). status reports both separately. blockers are tasks stuck in blocked status due to unmet task dependencies. conflicts are active claims that overlap on files. Address them differently.
- Running
sync --fix --yes without reading the report first. Run sync (without --fix) first to read the orphan report, then decide whether auto-remediation is appropriate.
Composition with Other Skills
State-ops fits into a repeating inspection-then-action cycle:
| Sequence | Skill |
|---|
| Nothing precedes state-ops | State-ops is read-only and safe to run at any point in any session |
status shows PRD: draft | Proceed to the prd skill to author and review the PRD |
status shows PRD: approved (or reviewed) with 0 ready tasks | Proceed to the plan skill — tasks may need scoring, expand, or review-to-ready promotion |
list or next returns a task ID | Proceed to the claim skill to take ownership |
show TASK_ID reveals complexity of 4 or more | Return to the plan skill — run anvil expand TASK_ID before claiming |
conflicts shows overlap | Resolve the conflict first (wait for the other claim to complete, or coordinate with the other agent) before proceeding to claim |
State-ops is the starting point of every agent work session. It answers "what is true right now" before any skill decides "what to do next."
Sync operations
Beyond local reconciliation, anvil sync github (and anvil sync provider <id>) push/pull against an external provider. See docs/github-sync.md for the full CLI reference, provider variants, and conflict-resolution strategies.
