roundtable run cannot load configurationConfirm that .roundtable/config.yaml exists (roundtable init creates it), that keys are supported by Configuration Reference, and that nested values use exactly two spaces per level. The current loader is not a full YAML parser. Inline comments and YAML-looking values can become literal string content rather than being rejected; typed integer/boolean fields reject values they cannot convert. Inspect the effective text rather than assuming valid YAML is valid Roundtable configuration.
Ensure the configured database parent directory is writable and that no incompatible process is holding a lock. SQLite runs with WAL, foreign keys, and a 5-second busy timeout. Preserve -wal and -shm files while a database is open; deleting them can lose or corrupt live state. Back up through the supported maintenance process.
Check that the parent directory for mcp.socket_path exists or can be created, the path is not already owned by another Roundtable process, and the user can bind/remove the socket. Unix sockets are local IPC and must not be exposed by copying or forwarding them to untrusted users.
An active run may be waiting for initial work, a pending agent turn, unresolved task status, a proposal review, or human action. Inspect roundtable table, roundtable watch --run RUN_ID, task/proposal state, and agent.turn_* events through MCP. A no-task run intentionally waits for incoming MCP work rather than completing immediately.
The execution supervisor launches configured external CLI workers inside containers and captures supported structured session IDs. It does not retain provider-home state after execution cleanup, so a captured ID alone cannot resume a later execution. Confirm container runtime/image availability and inspect persisted execution metadata. See Agent Adapters and Sessions and Resume.
Set the required NEXT_PUBLIC_* variables before building, verify they target the API origin and workspace you intend, and check browser network errors, API CORS policy, and API health. NEXT_PUBLIC_* settings are compiled into the client bundle. The UI surfaces missing configuration and normalized API problem details; it does not supply bearer credentials itself.
Follow the full generation/preparation/build/check sequence in Deployment, not just mkdocs build. MkDocs reads staged .docs-build, so stale content can survive in a build if preparation is omitted. Strict mode promotes emitted warnings to failures; it is not a complete coverage, semantic-link, or browser validator. Add pages to mkdocs.yml and run scripts/verify-docs-navigation.py for the separate rendered-sidebar checks. Local non-Markdown assets are not automatically copied by preparation. Task packs and prompts are intentionally excluded.
refusing to overwrite existing file is expected when a starter target already exists. Initialization does not skip that file and continue. Inspect the current scaffold and configuration before retrying: earlier directories/files can remain from a partially failed attempt. Do not use --force merely to silence the error; it replaces customized protocol, project, policy, configuration, bootstrap task, and integration files. An already-scaffolded project normally starts with run, not another init. See Quickstart.
Check whether the command uses direct mode or --socket. Direct mode opens local state itself; socket mode requires a listening server and has no explicit response deadline. --args-file is a process-local file containing the tool argument object, not a socket request envelope. A hanging client or lost response does not prove a server-side command stopped or a mutation was rolled back. Inspect records and artifacts before retrying. In particular, avoid blindly repeating proposal creation, patch application, or test execution. See CLI Reference.
An exact tool not implemented: repo.dependency_context error is not caused by a bad manifest or missing package manager. The name is advertised but has no dispatch handler. Use the documented file-read/search alternatives; there is no resolved dependency graph behind that tool.
Elapsed wall-clock TTL does not itself change stored status. Conflict checks still include exact active claims until an expiration operation transitions them. Inspect claims list and, when authorized, run claims expire with the intended root/run attribution. Invalid expiry timestamps are skipped by expiration. Do not delete claim rows to force access or infer that session shutdown releases leases. Reconciliation checks content hashes, not heartbeat age or the stored resume policy. See Resource Claims.
Standalone mcp serve accepts requests but does not tick the Chair scheduler. Confirm an orchestrator is running against the same database and run ID. Drain MCP table.watch cursor pages and correlate the request event ID with agent.turn_scheduled, agent.turn_started, and agent.turn_completed. An outstanding scheduled turn blocks later scheduling even when its start acknowledgment never arrives; there is no scheduler timeout or automatic stale-owner release. Do not send repeated requests or fabricate lifecycle events for another agent. Inspect ownership and coordinate recovery with the human/Chair. Turn ownership does not launch provider processes.
Convergence reads database-wide tasks/proposals and uses exact status rules, not test success. Unknown task statuses count as active; unknown proposal statuses do not. Failed tasks can be terminal for convergence. Historical work can affect a new run. A bounded headless invocation can exit successfully on its iteration limit while its run remains active; this is not completion. Inspect the persisted run, individual work records, tests, and approvals separately. See Orchestration.
The TUI and CLI watch use polling; CLI follow suppresses previously printed IDs but fetches a limited recent window without a cursor. More arrivals than the window can fit between polls can be missed. Use MCP cursor polling for reliable database replay. API WebSocket live delivery uses an in-process buffered bus that can drop messages; a separate runtime writing the same database does not notify that bus automatically. Reconcile persisted sequences and replay/snapshot state rather than treating an open connection as complete history. See Architecture and TUI and Watch.
Briefings are assembled from current local rows, not frozen session-start snapshots. Task/claim/proposal selection is agent-based, decisions are database-wide, and the displayed run ID does not scope all those reads. Completed assigned tasks are not excluded, and only exact accepted/rejected proposals are omitted. Prefer an explicit session ID and verify each selected entity before acting. A heartbeat with status active does not clear a prior end timestamp. See Sessions and Resume.
The tool envelope reports successful command execution/persistence handling, not a passing test verdict. Read test_run.status: a command failure is stored as failed while the tool can still return ok: true. Missing or unreadable saved log files can yield empty log text without a tool error. Results are not bound to a patch hash or repository revision, and patch application does not automatically run expected tests. See Tests and Evidence.
A 409/5xx, timeout, or disconnected client does not establish rollback. Runtime patch application mutates files before completing artifact/database persistence. Branch switching can change Git before recording workspace revision and attempts only best-effort restoration on failure. Notification acknowledgment can save read state before a failed reload; diagnostics can persist a test row before failing event insertion. Preserve the response/request IDs and inspect the relevant files, entity rows, audit/outbox, and artifacts before retrying. Do not change keys merely to bypass a cached conflict. See Transactions, Repository Control, and Retry Semantics.
This is a known HTTP authorization limitation, not evidence that the agent acquired human consent. Security middleware assigns agent identity in context, but the shared human helper reads the original role header. An accepted agent token with an allowlisted human-role header and actor ID can satisfy that helper. Do not distribute agent credentials assuming narrow read-only authority. Restrict both token types to trusted participants and inspect raw actor/decision provenance; changing an advertised role label does not repair the boundary. The current documentation records the defect rather than claiming it is fixed. See API Authorization. Local MCP socket trust is a separate boundary.
Effective settings overlay only the latest workspace revision’s one section onto defaults; they do not replay earlier section updates. Saving another section can make the prior section appear reverted, and a partial update does not inherit omitted keys from older revisions. Inspect the current version and raw revision history through trusted database access before resubmitting values. HTTP settings also do not rewrite runtime YAML, enforce loaded policy, or restart components. Governance confirmation tokens bind workspace/section/version, not proposed values. See Settings; do not infer applied runtime configuration from a successful save.
HTTP deliberation start changes stored status without launching a provider, advancing rounds, or enforcing its budget/time metadata. HTTP session creation records starting without spawning a CLI. The separate coordinator execution supervisor launches configured external CLIs inside containers and captures supported structured session IDs, but provider-home persistence and automatic resume across disposable containers are not implemented. Check the actual runtime process and execution evidence separately; a static readiness label or stored status is not a process probe. Do not repeatedly create sessions/deliberations to force execution. See Deliberation Administration, Sessions, and Adapters.
HTTP compensation copies the transaction’s rollback path into a high-risk pending proposal. Runtime rollback artifacts are JSON recovery entries, not unified reverse diffs, and that endpoint does not convert them or create affected-resource rows. Preserve the artifact, inspect current files for later work, and prepare a reviewed, valid patch through the governed workflow. Do not feed rollback JSON to Git or treat compensation creation as restoration. See Transaction Control.
The actionable count includes only unread, unresolved rows marked exactly 1; the actionable list can include read-but-unresolved rows. Recipient filters use exact equality and do not automatically include broadcast *. Broadcast acknowledgment changes one shared row, not a per-user receipt. Marking read does not resolve the underlying approval or proposal. Verify recipient, read/resolution timestamps, and authoritative linked entity separately. See Notifications.
The current exporter emits HELP/TYPE declarations followed by bare numeric values rather than named metric samples. HTTP 200 therefore does not establish valid Prometheus ingestion. The counters are process-local, unlabeled, and updated after handler completion; they are not persisted governance statistics. Inspect the actual response and collector parsing evidence, and avoid building operational guarantees on the declared names alone. See HTTP Observability.
Checkpoint does not stop writers, and its own subsequent audit insert can create new WAL content. A completed response with nonzero busy also does not prove truncation. Do not delete sidecars. A database backup excludes its own later creation audit and does not include external patch/log artifacts. Coordinate all writers and verify restoration separately. See Maintenance.