agent-loop-ext
Crash-resilient external agent loop with state persistence and CI/CD integration
Evaluate a pinned normalized decision ruleset through an explicitly configured Jev or LLM-subagent binding
$ npx -y skills add jmagly/aiwg --skill decision-evaluate --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/decision-evaluateContext preview
The summary Claude sees to decide when to auto-load this skill.
Evaluate a pinned normalized decision ruleset through an explicitly configured Jev or LLM-subagent binding
namespace: aiwg name: decision-evaluate platforms: [all] description: Evaluate a pinned normalized decision ruleset through an explicitly configured Jev or LLM-subagent binding requires: - feature-enabled: AIWG_DECISION_ENABLED=1 - request: dispatcher request JSON with authored artifact paths and runtime adapter configuration ensures: - normalized-result: returns a decision.aiwg.io/v1alpha1 RulesetResult - backend-boundary: definitions and rulesets contain no vendor request payloads or credentials script: entrypoint: scripts/decision-evaluate.mjs runtime: node cwd: project-root argsHint: "--request <dispatcher-request.json>"
Evaluate one pinned `DecisionRuleset` with one `DecisionBinding`. The dispatcher validates pins, input, capabilities, typed outputs, retry/fallback budgets, and composition before returning an outcome as data. It never authorizes or executes the outcome.
The request document is runtime configuration, not a portable decision artifact. It names `rulesetPath`, `bindingPath`, `definitionPaths`, `inputPath`, `runId`, `invocationId`, optional `receiptDirectory` (which requires `receiptIntegrityKeyRef`; see `docs/operations.md`), `credentials` mappings from logical reference to environment-variable name, and optional `adapterModules` for configured worker transports. Credential values are read only at adapter call time and never written to results.
Network-capable adapters (including the packaged Jev adapter) require `projectionPolicyPath`: a trusted projection policy file, or an array of policies selected by exact adapter and model. Without it the dispatcher refuses before any credential or transport use and exits `2`. Only adapters that declare `egress: { mode: 'none' }` (for example the offline fixture worker) run without a policy. `adapterOptions.jev` sets the Jev `endpoint`, `allowedOrigins` and the operator-declared deployment `region`; the policy origin and region must match them. There is no dispatcher setting that sends unprojected state to a network adapter. See `agentic/code/addons/decision-engine/examples/dispatcher-request-jev.json`.
Set `AIWG_DECISION_ENABLED=1` explicitly. Existing workflows remain unchanged when the flag is absent.
The deployed script loads the compiled runtime from the installed `aiwg` package through `scripts/runtime-root.mjs`: `AIWG_ROOT` when it names a built package, then a project `node_modules/aiwg`, then the `aiwg` executable on `PATH`.
Reusable project context and specialist workflows for the AI tools you already use. Plan software, coordinate specialist reviews, prepare campaigns, investigate incidents, organize research, curate media, and maintain operational knowledge.
Repo: jmagly/aiwg
Crash-resilient external agent loop with state persistence and CI/CD integration
Detect requests for iterative autonomous agent loops and route to the appropriate loop executor
Automatically execute tests when code-generating agents modify source files, enforcing the execute-before-return pattern
Enable agent loops to learn from similar past tasks and share patterns across loops
Query and manage the executable feedback debug memory
Execute tests on generated code and iterate until passing