SUCCESSOR Ω: useful local work

This additive workbench preserves the existing personal lab, aggregation workshop, lifecycle, native commands, journals and deployed contracts. Open the workbench, choose a mission, run the example, inspect its files, and save your working HTML copy. No account, provider key, wallet or token is required for these five adapters.

What actually runs

Family	User-controlled inputs	Actual computation and deliverable	Deliberate boundary
Reliability	contains source, interval, test cases	Execute a finite range-function grammar; test four repairs; export source, patch and runnable Node checks	One closed-interval utility; arbitrary repositories are not executed
Performance	Integer values, repeats, repetitions	Measure Set/sort against sort/deduplication; alternate order; check output equality; export timings and selected source	Local timings, no guaranteed advantage or monetary claim
Documentation	Markdown with successor-example JSON blocks	Execute URI encoding/decoding and UTF-8 operations; retain failures; propose corrected examples	Three explicitly supported operations, not an arbitrary SDK
Public data	Two-column group,amount CSV, units	Parse rows, group and reconcile integer totals; export editable CSV, JSON and HTML	Public/synthetic/licensed nonpersonal inputs; no inferred data rights
Numerical	Polynomial coefficients, interval, even step count, tolerance	Calculate exact rational quadrature and analytic integral; compare trapezoid and Simpson errors	Polynomial quadrature, not general scientific validation

Each family runs different domain logic. Changing permitted inputs changes execution and its commitment. Changing a display name or interface language does not create stronger evidence. Failed candidates and negative comparisons remain in the result. A locally completed task can still require repair or retain its incumbent.

Input policies and limits are checked before execution. Unknown fields, duplicate JSON keys, unsupported source, missing rights, unsafe integers and unbounded inputs are refused. CSV exports neutralize formula-like group labels; HTML reports escape supplied text. Imports are inert data, never plugin installers or trust registries.

Native operator and machine interface

Use the existing installed alpha-agent with Python 3.11–3.13 and Node 22 or later. Existing Python-only missions keep their dependencies. From a fresh directory outside the source checkout:

alpha-agent successor-work example reliability --output mission.json
alpha-agent successor-work plan mission.json
alpha-agent --home operator successor-work run mission.json --output result
node result/check.mjs

The JSON response includes the run ID, state, result digest and output directory. Keep that run ID for show, review, control and export. In the following commands, substitute the returned values:

alpha-agent --home operator successor-work show RUN_ID
alpha-agent --home operator successor-work review RUN_ID --result-digest RESULT_DIGEST --decision accepted --note "Reviewed the source and all failures"
alpha-agent --home operator successor-work export RUN_ID --output another-copy
alpha-agent successor-work restore result/evidence.json --output restored-mission.json
alpha-agent --home another-operator successor-work run restored-mission.json --output fresh-result

The machine interface is the same typed JSON CLI, protected by the existing private operator directory. There is no new HTTP listener, unauthenticated localhost API, account service or browser-to-host bridge. For a desktop browser-only start, use the saved HTML. Native Computer Use is a separate explicit setup; see Computer Use.

review records the local operator's acceptance or rejection of delivered work. It never fabricates an independent verifier, institutional admission, production authority or settlement.

Evidence, Jobs and lifecycle integration

The same domain source is bundled into the installed Node worker and the standalone browser page. workbench_sources.json commits every transitive source, including the provenance-checked AGI Jobs compiler. The browser/native pack contains inputs, source identity, artifacts and their hashes, all observations, limits, draft Jobs compilation and elapsed wall time. Provider and settlement spend are zero for fixed local adapters. Host compute cost and actual human effort remain unmeasured, never silently zero.

The pinned AGI Jobs compiler groups its 21 obligations into seven typed lifecycle work packages. These plans have zero funding and require actual reviewer/acceptance commissioning before live dispatch. No token transfer, chain approval, market connection or work settlement happens in this workbench.

Native runs register an Institution, MissionConstitution and CandidateReleaseManifest using the existing SuccessorStore, strict protocol and local-role custody. These records nominate the frozen maintained work method; they do not install it as a serving release. WORLD/POLICY claims, proof, admission and authority retain the existing lifecycle's separate gates. The existing successor-demo and Jobs/ Ascension acceptance suites exercise full bounded delegation, revocation, negative settlement and succession.

Native exports add lifecycle.json and an Ed25519 signed-receipt.json. Verify attribution with a source key retained independently of the downloaded package:

alpha-agent successor-work verify-receipt result/signed-receipt.json --source-key TRUSTED_PUBLIC_KEY_HEX

The adjacent source-key.txt identifies the producing key; importing that file is not independent trust. A signature establishes local attribution and integrity, not factual correctness or current authorization. Retain the existing journal checkpoint separately when freshness and rollback detection are required.

Recovery, ownership and migration

Pause with successor-work control RUN_ID --state paused; use stopped or retired to end the run. Interrupted pure work may restart only with the same task, run ID, --resume, and a new output directory. All attempts share the three-attempt ceiling. Stopped and retired runs cannot resume. The existing global operator pause is also checked during the worker. Concurrent execution uses signed revisions and leases.

Each artifact directory is new; existing files are never overwritten. Public exports exclude private keys and grants. Restoring evidence produces a task and historical, unauthenticated observations with empty active proof and authority. Rerun and review locally. Copying a name or a release number transfers no trust.

Saved HTML works without browser storage or a service worker in modern desktop browsers. Save a new copy after edits; retain full evidence separately. Browser execution is cooperative and isolated from host files; the native fixed worker additionally has a process deadline. Mobile browsers can use the hosted page; mobile file-preview applications are not promised to execute downloaded HTML.

No migration of existing state is needed. New records use @successor-work: and work- institution identities in the existing append-only journal. Optional local role keys occupy workbench-principals. Use the existing installation, backup, update and uninstall procedures; no historical wallet is required.

Extension SDK

examples/successor-workbench/engine.mjs exports createWorkbench(adapters). An operator-installed adapter supplies a versioned manifest, an input validator and a bounded asynchronous execute method. All six output fields are checked by the shared transport, source commitments and draft Jobs compiler. The additional word-inventory example executes without modifying the core registry:

node examples/successor-workbench/extension.mjs

Source adapters are trusted software installed through normal code review. Never load an adapter from a mission/evidence JSON file, a model reply or an untrusted URL. Changing an adapter or supplier requires a new source commitment and fresh evaluation. The SDK is for local public work; institutional effectful tools use the existing ActionGateway and separately authorized resource reservations.

Reproducible checks

python scripts/generate_successor_workbench.py
python -m pytest tests/runtime/test_successor_workbench.py tests/runtime/test_successor_computer.py --confcutdir=tests/runtime -q
node --test examples/successor-workbench/test.mjs
python -m scripts.validate_successor_workbench --output evidence/workbench-browser --axe-script tests/browser/node_modules/axe-core/axe.min.js

The Computer Use container test requires SUCCESSOR_COMPUTER_IMAGE; its absence is a reported skip, not a pass. The release job requires this test with a real immutable image. Live providers, independent evaluation, actual customer acceptance and mainnet commissioning have separate evidence requirements. See the implementation and acceptance record.