Adaptive inventory application

This executable example applies ALGAL's application lifecycle to inventory observations and replenishment decisions. It uses actual bounded file probes, the native query engine and replay verification. Proposals are authored deterministic programs. No model call, purchase or external inventory write is performed.

Build the native CLI, then give the scenario a new directory:

cargo build --locked
ALGAL_MEMORY_NATIVE="$PWD/target/debug/algal" \
ALGAL_BIN="$PWD/target/debug/algal" \
bun examples/adaptive-inventory/run.ts /tmp/algal-inventory-example

ALGAL_MEMORY_NATIVE is required and its executable is pinned by SHA-256 for the run. ALGAL_BIN optionally renders the initial unresolved and final captured views using application report. Existing output directories are rejected. All fixture mutations and retained application state live beneath the supplied new directory.

Three inhabitants have separate manifest budgets and authority declarations:

The host enforces the investigation/proposal route declarations. Every manifest has a zero model-call budget. These are independent program identities and durable deliveries within one application, not separate operating-system sandboxes or a claim of measured multi-agent performance.

The active revision retains a goal to establish current stock and discover the current tool before selecting a recommendation. Its typed query joins both observations. Unknown and stale goal statuses drive investigation; a supported goal selects the planner entrypoint. The evidence records the sequence unknown → supported → stale → supported → supported, and the final view retains the exact goal definition, state, memory, and derivation reference. Support establishes current evidence, not that inventory was replenished.

Eight separate Bun host processes execute the scenario. The parent waits for each host to exit successfully before starting its successor; it does not own an application service. The native query executable is pinned across every invocation. This demonstrates actual exit/reopen continuity, not abrupt termination at every filesystem publication boundary.

The first host commits an investigation, captures a fresh query at that exact head, and exits. Its view retains the unresolved query, declared host probes and their dependencies, and pending investigation request. Its successor observes inventory.json and toolchain.json, joins the two admitted observations in a native query, and executes the incumbent planner. Its selected fixture tool at tools/inventory-a is run and checked. The fixture then moves that executable to tools/inventory-b, updates its configuration, and changes stock. A new host confirms the old evidence is stale and schedules fresh investigation.

After rediscovery, an authored proposal produces a planner that reads the declared current tool location. Evaluation freezes two train, three validation, and two holdout cases. Both candidates execute train and validation; only the winner executes holdout, for twelve retained case executions. The candidate must improve validation without regressing an incumbent pass, then pass the holdout. Activation occurs in another host invocation against the exact parent.

Two competing writers then attempt to publish an episode against one captured state. Exactly one succeeds and a stale retry is rejected. Another host executes the pinned planner and verifies its actual retained receipt; the fixture checks that its selected tool at B runs. A final host confirms no settled episode is redelivered. There are four observation probes and two planner episodes.

Inspect evidence.json for counts, process identities, and content identities; phase-*.json for each exited host's evidence; tool-*.json for the selected fixture tool checks; view-unknown.json for the initial unresolved capture; view.json for the final supported capture; and optional report-unknown.html and report.html for their passive renderings. evidence.json retains both view identities and the corresponding report paths.

The final drilldown links current observations to their scopes, raw records, receipts, decoders, and admission identities; it retains query facts/result/source references, declared probes, accepted revision evaluation references, and episode bindings and settlement. Organism probes display their VM budgets; these host-backed inventory probes display a null VM budget. Each evidence category is capped at 32 rows with its own truncation flag. Immutable state references are captured together, while work statuses are separately observed dispatch records. The rendered report does not query, replay, authenticate observations, or dispatch. Its unknown and supported labels describe the selected evidence, not external inventory completion.

The application store retains observations, receipts, evaluation evidence and revision history. The fixture proves these deterministic mechanisms; it does not establish commercial inventory accuracy, model learning or general coding improvement.