Dashboard-first orchestration runtime for autoresearch.
Turn drifting agent work into durable runs, locked work items, and reviewable verdicts.
v0.2.6.4 authority/discovery patch · closed discussion authority reaches legacy work items, and dashboard sync refreshes stale scaffolds
THOTH CONTROL PLANE
Claude Code surfaces Codex surfaces
/thoth:* command set $thoth command set
\ /
\ /
+-------------+
|
v
+----------------------------------------------------------------------------+
| Layer 1. Host Surface |
| |
| init discuss run loop review auto status |
| doctor dashboard |
+----------------------------------------------------------------------------+
|
v
+----------------------------------------------------------------------------+
| Layer 2. Planning Authority |
| |
| init -> bootstrap, migrate, or resync .thoth authority |
| discuss -> record discussions, decisions, and work items |
| |
| Discuss -> Decision -> Work Item Object Graph |
| | |
| v |
| Ready Work (--work-id) |
+----------------------------------------------------------------------------+
|
v
+----------------------------------------------------------------------------+
| Layer 3. Execution Runtime |
| |
| run -> one durable execution packet |
| loop -> one durable recoverable loop packet |
| review -> structured findings through the same protocol |
| auto -> priority-driven child loops for actionable work |
| |
| +---------------------------+ |
| | Ready Work (--work-id) | |
| +-------------+-------------+ |
| | |
| +----------+----------+ |
| | | |
| v v |
| Run Loop |
| | | |
| +----------+----------+ |
| | |
| v |
| Run Ledger / Events / Artifacts / Result |
| | |
| v |
| Mechanical Validation / Acceptance |
| |
| attach watch resume stop |
+----------------------------------------------------------------------------+
|
v
+----------------------------------------------------------------------------+
| Layer 4. Read Surfaces |
| |
| dashboard -> human-visible runtime workbench |
| status -> active / stale / attachable run summaries |
| doctor -> strict health, projection, and runtime-shape audit |
| report -> available through status --report |
| |
| +-----------+-----------+-----------+-----------+ |
| | | | | |
| v v v v |
| Dashboard Status Report Doctor |
+----------------------------------------------------------------------------+
Key invariants:
- .thoth is the shared machine/runtime authority
- .agent-os is the human governance layer
- run and loop are strict --work-id surfaces
- auto executes only actionable ready/active/failed work; blocked and draft work require human decisions
- dashboard, status, report, and doctor are read surfaces, not authority writers
- run, loop, and auto progress through the RuntimeDriver until terminal or paused
Thoth is a dashboard-first orchestration runtime for autoresearch. It assumes chat alone is not an operating system: truth must survive the session, progress must stay visible, and completion must be mechanically testable.
| Problem | Why it matters |
|---|---|
| Work is not persistent | Long-running work dies with the session, so the agent cannot keep working while you sleep and there is no durable state to resume or audit. |
| Parallel work is invisible | Multiple threads or delegated runs drift apart, and humans cannot see what is actually active. |
| Agents can claim completion too early | A fluent summary can hide that nothing mechanical passed. |
| Docs and state rot over time | Discussions, decisions, work items, and runtime facts drift until nobody knows which layer is authoritative. |
| Mechanism | What it does | Counters |
|---|---|---|
| Hooks + watchdog + runtime | Keep execution attached to durable ledgers and observable lifecycle events. | Work is not persistent |
| Dashboard-first visibility | Show live, stale, attachable, and host-specific runtime truth in one read surface. | Parallel work is invisible |
| Mechanical yes/no acceptance | Force validators, ledgers, and result payloads to decide whether work really passed. | Agents can claim completion too early |
| Object graph + execution system + locked work items | Freeze what is allowed, compile it into runnable work items, and keep authority layers from drifting. | Docs and state rot over time |
Humans should not spend their attention tracking every grain of sand in the funnel. Thoth lets AI own the middle of the hourglass, while the dashboard shows the gold that survives: decisions, work items, runs, results, and the current verdict.
| Stage | Purpose | Input | Output |
|---|---|---|---|
| Intent | Capture the user request and operating boundary. | Human goals, constraints, repo context | Direction for planning |
| Decision | Lock key choices before execution drifts. | Intent, open questions, policy constraints | Recorded decisions |
| Work Item | Freeze goal, constraints, execution plan, eval, runtime policy, and decisions. | Discussion, decisions, requirements, acceptance rules | Ready or blocked work item |
| Run | Execute one frozen work_id@revision through phase results. |
Work item, controller policy, host surface | .thoth/objects/run plus .thoth/runs/<run_id> ledger |
| Result | Produce a mechanical verdict instead of narration alone. | Validator outputs, artifacts, runtime checks | Structured result and acceptance evidence |
| Dashboard | Let humans read the final state without replaying the chat. | Portable authority plus local ledgers and read models | Inspectable project truth |
Thoth project state is intentionally split into three layers.
Portable authority is the Git state needed to continue work after a fresh clone. Commit AGENTS.md, CLAUDE.md when the Claude surface is enabled, .thoth/objects/project/, .thoth/objects/work_item/, .thoth/objects/discussion/, .thoth/objects/decision/, .thoth/docs/agent-entry.md, .thoth/docs/project.json, and .thoth/docs/source-map.json. These files define the project, discussion history, decisions, and runnable work item graph.
Runtime evidence is local by default. New projects get .thoth/.gitignore rules for .thoth/runs/, .thoth/derived/, .thoth/docs/work-results/, .thoth/objects/run/, .thoth/objects/artifact/, .thoth/objects/controller/, and .thoth/objects/phase_result/. These ledgers remain on disk for local audit, but a fresh machine should start a new run rather than attach to old PIDs, leases, workers, supervisors, or dashboard processes.
Dashboard dependencies and cache are also local. Thoth writes idempotent ignore rules for tools/dashboard/frontend/node_modules/, tools/dashboard/frontend/dist/, Vite cache, backend Python cache, and the dashboard SQLite read model under .thoth/derived/dashboard/. If a team intentionally wants to carry a run to another machine, export a concise report with thoth status --report or archive the specific .thoth/runs/<run_id> evidence bundle explicitly; Thoth does not add every runtime ledger to Git by default.
thoth init --sync refreshes the managed dashboard scaffold when the installed plugin has newer runtime/read-model fixes. The previous scaffold is copied to ignored .thoth/derived/dashboard-sync-backups/ before overwrite, so stale dashboard code can be recovered without making local backup files Git-visible.
Fresh-clone recovery means:
git clone <repo>
cd <repo>
codex plugin marketplace upgrade thoth
thoth doctor --version
thoth status --json
thoth run --work-id <ready-work-id>That flow resumes from committed authority and starts new local runtime evidence. It does not take over old machine-local processes.
- Install Thoth on the host surfaces you use.
claude plugin marketplace add SeeleAI/Thoth --scope user
claude plugin install thoth@thoth --scope user
codex plugin marketplace add SeeleAI/ThothFor Codex, adding the marketplace is the source step. Then install or enable the thoth plugin from the Codex plugin directory.
After the plugin is installed, two different entry layers exist on purpose:
- Public plugin surface:
Claude /thoth:*,Codex $thoth <command>, and the plugin-provided shell wrapperthoth <command> - Source-repo development fallback:
python -m thoth.cli <command>
Use the plugin-installed thoth wrapper in fresh repos or empty directories. Use python -m thoth.cli only when you are intentionally running against a checked-out Thoth source tree and want execution pinned to that exact checkout.
- Initialize the repository you want Thoth to manage.
/thoth:init
$thoth init
- Start the first strict run from a compiled task.
/thoth:run --work-id task-1
$thoth run --work-id task-1
- Open the read surface.
/thoth:dashboard
$thoth dashboard
| Host | First install | Stable upgrade | Important note |
|---|---|---|---|
| Claude Code | claude plugin marketplace add SeeleAI/Thoth --scope user then claude plugin install thoth@thoth --scope user |
claude plugin marketplace update thoth then claude plugin update thoth@thoth --scope user |
Restart Claude Code after plugin update so the new version is applied. |
| Codex | codex plugin marketplace add SeeleAI/Thoth, then install or enable thoth from the Codex plugin directory |
codex plugin marketplace upgrade thoth |
add takes a source such as SeeleAI/Thoth; upgrade takes the configured marketplace name, which is thoth in this repo. |
Default development verification is intentionally targeted-only. Broad or full sweeps are not the normal workflow.
- The public selftest entrypoint is now atomic-only:
python -m thoth.selftest --case plan.discuss.compile --case runtime.run.livepython -m thoth.selftestwithout any--casefails on purpose and prints the available case catalog.- Every case runs in its own workdir and artifact directory, writes a per-case report entry keyed by
case_id, and must not depend on side effects from an earlier case. - Release, regression, and closeout gates must record explicit case IDs instead of broad aliases such as
hardorheavy. - The current catalog is split into repo-local capability probes such as
plan.discuss.compile,runtime.run.live,runtime.loop.sleep,review.exact_match,observe.dashboard,hooks.codex, plus host-surface probes such assurface.codex.run.live_prepareandsurface.claude.loop.stop.
- Allowed developer entrypoints:
python -m pytest -q tests/unit/test_selftest_registry.py
python -m pytest -q tests/unit/test_selftest_helpers.py::test_validate_pytest_invocation
python -m pytest -q --thoth-target selftest-core- Blocked by default: bare
pytest, directory-wide invocations such aspytest tests/unit, and broad tier sweeps such aspytest --thoth-tier heavy. - Broad runs are reserved for explicit release or CI situations and require
--thoth-allow-broadorTHOTH_ALLOW_BROAD_TESTS=1. --thoth-tieris retained only as an explicit override path for those exempted broad runs; it is not the default development interface.- The target manifest lives in
thoth/test_targets.py. - Use the helper below to translate changed paths into recommended pytest targets and atomic selftest cases:
python scripts/recommend_tests.py thoth/observe/selftest/runner.py tests/conftest.py| Command | Host Surface | Purpose | Input | Result |
|---|---|---|---|---|
init |
Claude: /thoth:initCodex: $thoth init |
Audit, initialize, migrate, or resync canonical Thoth authority. | --sync, --migrate preview, --migrate apply, --migrate --preview, --migrate --apply, or optional config payload |
Portable .thoth authority, migration ledger, ignore rules, generated projections, dashboard scaffolding, scripts, and tests |
discuss |
Claude: /thoth:discussCodex: $thoth discuss |
Record planning decisions without entering code execution. | Topic, decision payload, or work payload | Updated discussion, decision, or work_item objects plus generated docs view |
run |
Claude: /thoth:runCodex: $thoth run |
Execute one ready work item through a durable runtime packet. | --work-id, optional host or executor controls, optional attach/watch/stop |
Durable run ledger with state, events, phase results, artifacts, and terminal result |
loop |
Claude: /thoth:loopCodex: $thoth loop |
Iterate on one ready work item through a controller service. | --work-id, optional resume or sleep controls |
Controller object, child run lineage, and bounded iteration history |
review |
Claude: /thoth:reviewCodex: $thoth review |
Produce structured findings without modifying source code. | Review target, optional --work-id, optional executor controls |
Structured review result recorded through the shared protocol |
auto |
Claude: /thoth:autoCodex: $thoth auto |
Run the priority queue while the user is away. | Optional --sleep, --rounds, --scope, or explicit --work-id |
Auto controller, child loop lineage, monitor events, and terminal or paused summary |
status |
Claude: /thoth:statusCodex: $thoth status |
Show project health, active durable runs, doctor, report, or dashboard views. | Optional --json, --doctor, --report, or --dashboard |
Shared status snapshot and read-only derived views |
doctor |
Claude: /thoth:doctorCodex: $thoth doctor |
Alias for status --doctor; strictly audit health and runtime shape. |
Optional --quick or --json |
Health report with validation findings |
dashboard |
Claude: /thoth:dashboardCodex: $thoth dashboard |
Alias for status --dashboard; manage the local dashboard runtime. |
Optional action: start, stop, or rebuild |
Local dashboard process and read endpoints backed by authority plus local .thoth ledgers |
| Signal | What you can inspect |
|---|---|
| Local runtime truth | .thoth/runs/* keeps run, state, events, artifacts, and result payloads on the current machine by default. |
| Locked planning authority | .thoth/objects/discussion/, .thoth/objects/decision/, and .thoth/objects/work_item/ define what execution is allowed to do. |
| Script-backed verification | Validators, doctor checks, and selftests decide pass or fail mechanically. |
| Shared read model | status, report, and dashboard all read from the same authority instead of chat memory. |
| Good fit | Why |
|---|---|
| Research and experimentation repos | They need durable memory, replayable results, and visible long-running work. |
| Engineering teams using AI for real changes | They need code execution, review, and acceptance to stay auditable. |
| Teams that want Claude Code and Codex parity | They need one host-neutral command model rather than two drifting workflows. |
| Current boundary | Implication |
|---|---|
run and loop are strict --work-id surfaces |
Free-form execution is intentionally rejected. |
| Host parity is semantic, not identical UX | Claude and Codex still need their own install and local runtime wiring. |
| Dashboard is a local service, not a hosted control plane | Operators need a machine that can run the backend and frontend assets. |
| The hero logo currently ships as a raster PNG | A clean SVG and icon-family refinement is still useful for smaller surfaces and plugin packaging. |
Built in public by contributors who want AI work to remain inspectable.
Contribution path: open a pull request or start a discussion.
MIT. See LICENSE.