cjm-substrate

A dependency-isolated capability-composition runtime: heterogeneous tools run in their own environments behind a uniform HTTP/JSON boundary; the host composes them into workflows with resource-aware scheduling and threads provenance through a context graph.

Modules

• cjm_substrate
• cjm_substrate.bootstrap — One-call factory that assembles a CapabilityManager + JobQueue + capability
• cjm_substrate.cli — CLI tool for declarative capability management.
• cjm_substrate.core
• cjm_substrate.core._telemetry — Shared GPU/CPU attribution helpers used by both JobQueue._sample_resource_snapshot (CR-6 Stage 3) and CapabilityManager._record_sample_safe (CR-7).
• cjm_substrate.core.adapter — The typed-task half of the capability-unit fracture (pass-2 Thread 3) —
• cjm_substrate.core.adapter_manifest — The ADAPTER unit's registration manifest + the surface-based compatibility matcher (CR-17 pt 2, stage 4). Pass-2 Thread 3: registration/discovery = per-unit manifests generated in-env and found by discover_manifests(); compatibility is DERIVED, not declared — the capability records only its structural surface, the adapter declares its protocol (recorded here as member names + parameter lists), and the substrate matches manifest-vs-manifest. Works against UNLOADED capabilities with zero protocol imports host-side.
• cjm_substrate.core.capability — The tool-capability interface — the manage-the-tool half of the capability-unit fracture (pass-2 Thread 3): identity, lifecycle, config, cancellation, and observability for a tool running in a worker process, serving both concrete capabilities (in workers) and remote proxies (in the host). The task channel is deliberately NOT here: typed task contracts live on task adapters (core.adapter + the per-task cjm--adapter-interface libraries) and results cross the worker boundary through the typed wire layer (core.wire); execute_stream's transitional default remains for fused-era capabilities only. Also home to the CR-12 worker-env contract (EnvVarSpec + the Q1-A template vocabulary), the SG-44/T28 action-dispatcher convention, and the pass-2 Thread-3 structural-surface derivation.
• cjm_substrate.core.config — Project-level configuration for paths, runtime settings, and environment management.
• cjm_substrate.core.config_store — Persistent storage for per-instance capability configuration (config + enabled flag + worker-env override), keyed by instance_id.
• cjm_substrate.core.diagnostics_store — CR-14 (stage 7): the disposable diagnostic-narrative class. Worker-written
• cjm_substrate.core.empirical_store — Persistent store for empirically-observed resource usage per (instance_id, config_hash) pair. CR-7's data foundation — record_sample is called from CapabilityManager.execute_capability* finally blocks; aggregates feed eviction-candidate selection + future UI hints + cost-aware retry decisions.
• cjm_substrate.core.errors — Typed exception hierarchy + JobError dataclass + default classification of bare Python exceptions. The substrate's CR-5 implementation per the 2026-05-19 substrate audit.
• cjm_substrate.core.journal_store — CR-14 (stage 7): the durable account-of-action. One substrate-derived,
• cjm_substrate.core.manager — Capability discovery, loading, and lifecycle management via process isolation.
• cjm_substrate.core.manifest_format — Typed parser + writer for the nested v2.0 manifest layout (2026-05-19 substrate audit, CR-8).
• cjm_substrate.core.metadata — Data structures for capability metadata.
• cjm_substrate.core.platform — Cross-platform utilities for process management, path handling, and system detection (Linux, macOS, Windows).
• cjm_substrate.core.ports — Capability compositions as DAGs of invocation nodes with typed input/output
• cjm_substrate.core.proxy — Host-side bridge to isolated capability workers: RemoteCapabilityProxy implements ToolCapability but forwards every call over HTTP to a Universal Worker subprocess running in the capability's own environment. Owns worker process management (spawn with the manifest's python_path, SG-4 parent-bound listening socket / FD inheritance, suicide-pact --ppid, CR-14 stream pump + lifecycle journaling), zero-copy input transfer (FileBackedDTO -> temp files), the typed error contracts crossing the wire (409 -> CapabilityCancelledError; _job_error 500 bodies and stream terminal chunks -> typed exceptions; CR-7 Track A worker-death classification), and a dual sync/async interface for scripts and FastHTML hosts.
• cjm_substrate.core.queue — Resource-aware multi-lane job queue for capability execution (stage 3 / CR-16 rework).
• cjm_substrate.core.scheduling — Resource scheduling policies for capability execution.
• cjm_substrate.core.secret_store — CR-12: project-local secret storage for API-based capabilities (file-backed, 0600).
• cjm_substrate.core.wire — Typed data transfer at the worker boundary — the zero-copy FileBackedDTO
• cjm_substrate.core.worker — FastAPI server that runs inside isolated capability environments (the Universal Worker): dynamically loads the capability class named on the CLI, exposes the HTTP lifecycle / execute / task / monitor surface for the proxy, monitors the parent process (the suicide-pact watchdog prevents zombie workers), and reports process-subtree telemetry for resource-scheduling decisions. This module is a process ENTRYPOINT (SG-39): host code never imports it.
• cjm_substrate.core.workspace — Workspace resolution: the marker-rooted directory that owns a pipeline's local artifacts (runs, graph data, substrate stores, TUI sidecars).
• cjm_substrate.utils
• cjm_substrate.utils.cache_paths — Per-(input-content, config) deterministic cache directories for capability outputs. Same (input content, action, config) always resolves to the same directory; any change to input content OR config produces a different one — no silent overwrites, no stale-artifact accumulation, and chained invalidation for capability sequences (see the cache-paths-design-provenance note for the ffmpeg-bug origin story).
• cjm_substrate.utils.hashing — Shared cryptographic hashing primitives for content integrity verification.
• cjm_substrate.utils.sidecar — JSON sidecar for shell view-state — settings and bookmarks that persist
• cjm_substrate.utils.validation — Validation helpers for capability configuration dataclasses.

API

cjm_substrate.bootstrap

• Pipeline class — Assembled substrate stack: manager + queue + capability bindings.
• create_pipeline function — Assemble a CapabilityManager + JobQueue + capability bindings in one call.

cjm_substrate.cli

• generate_adapter_manifest function — CR-17 pt 2 (stage 4): introspect a task-adapter impl in-env and write its adapter manifest.
• install_all function — Install and register all capabilities defined in capabilities.yaml.
• list_capabilities function — List installed capabilities from manifest directory.
• list_secrets function — List the secret KEY NAMES stored for a capability — never the values (CR-12).
• logs_command function — Tail / follow the observability stores (CR-14).
• main function — cjm-substrate CLI for managing isolated capability environments.
• regenerate_manifest function — Re-run introspection for an installed capability and rewrite its manifest.
• remove_capability function — Remove a capability's manifest and conda environment.
• retention_command function — Apply the diagnostics retention policy now (CR-14).
• run_cmd function — Run a shell command and stream output.
• set_secret function — Store a capability secret in the project-local SecretStore (CR-12).
• setup_host function — Install interface libraries in the current Python environment.
• setup_runtime function — Download and setup micromamba runtime for project-local mode.
• validate_file function — SG-6 + T23: validate a manifest / capabilities.yaml / capability source.
• workspace_doctor_cmd function — Run workspace integrity checks; exits non-zero when any check warns.
• workspace_init function — Declare a workspace: write the marker + create the conventional layout (idempotent; an existing marker is left untouched).

cjm_substrate.core._telemetry

• attribute_gpu_to_worker_subtree function — Attribute GPU memory across the worker's process subtree.

cjm_substrate.core.adapter

• TaskAdapter class — Base for task adapters — the typed-task half of the capability-unit

cjm_substrate.core.adapter_manifest

• AdapterManifest class — A discovered ADAPTER unit (CR-17 pt 2) — the registration record for one
• adapter_manifest_from_dict function — Reconstruct an AdapterManifest from its on-disk JSON shape.
• is_adapter_manifest function — Route a manifest file by the unit discriminator (capability manifests
• match_protocol_against_surface function — Surface-based compatibility (pass-2 Thread 3) — host-side, manifest-vs-

cjm_substrate.core.capability

• ConfigOption class — CR-11: one live option for a dynamic config field, with optional metadata.
• EnvVarSpec class — CR-12: one entry of a capability's spawn-time worker-environment contract.
• FieldOptions class — CR-11: the live option domain for one dynamic config field.
• ToolCapability class — Tool-capability interface: manage the tool/worker — identity, lifecycle,
• capability_action function — Marker decorator tagging a capability method as the handler for action_name.
• collect_capability_actions function — Collect action names from @capability_action-decorated methods on cls.
• derive_structural_surface function — Record a capability class's structural surface by pure self-introspection.
• expand_worker_env_template function — Substitute ${VAR} placeholders in template using placeholders.
• template_check_placeholders function — Return the set of placeholder names referenced by a worker-env template.

cjm_substrate.core.config

• CJMConfig class — Main configuration for cjm-substrate.
• CondaType class — Type of conda implementation to use.
• RuntimeConfig class — Runtime environment configuration.
• RuntimeMode class — Runtime mode for the capability system.
• SubstrateConfig class — Substrate behavior toggles.
• get_config function — Get current config (loads defaults if not set).
• load_config function — Load config with layered resolution (CLI > env vars > yaml > workspace > defaults).
• reset_config function — Reset to unloaded state (for testing).
• set_config function — Set current config (called by CLI callback).

cjm_substrate.core.config_store

• CapabilityConfigRecord class — Persisted state for a capability INSTANCE: config + enabled flag + worker-env override.
• CapabilityConfigStore class — Protocol for persisting per-instance CapabilityConfigRecord across sessions.
• LocalCapabilityConfigStore class — SQLite-backed default implementation of CapabilityConfigStore.
• delete function — Remove the record for an instance.
• get function — Fetch the record for an instance.
• list_all function — Return all stored records keyed by instance_id.
• set function — Persist a record. Stamps updated_at to the current time.

cjm_substrate.core.diagnostics_store

• DiagnosticRecord class — One structured worker log record (CR-14 diagnostics class).
• DiagnosticsLogHandler class — Worker-side logging handler writing DiagnosticRecords (CR-14).
• DiagnosticsStore class — Protocol for the disposable diagnostic-narrative store (CR-14).
• LocalDiagnosticsStore class — SQLite-backed default DiagnosticsStore (CR-14).
• StreamChunk class — One raw stdout/stderr line the host pump captured (death-rattle floor).
• append_chunk function — Persist one raw stream line.
• append_record function — Persist one structured record.
• apply_retention function — Retention as a QUERY (the CR-14 reframe's mechanical payoff).
• install_worker_diagnostics function — Configure worker-process logging (replaces the old basicConfig).
• normalize_stream_line function — Collapse CR progress frames to the final frame; drop empty results.
• query_chunks function — Raw stream read, session-scoped.
• query_records function — Filtered structured-record read.

cjm_substrate.core.empirical_store

• EmpiricalResourceRecord class — Aggregated empirical resource profile for a (instance_id, config_hash) pair.
• EmpiricalResourceStore class — Protocol for persisting empirically-observed resource usage.
• LocalEmpiricalResourceStore class — SQLite-backed default implementation of EmpiricalResourceStore.
• ResourceSample class — Single observation captured after an execute call completes.
• compute_config_hash function — CR-7: hash a capability instance's effective config for empirical-record keying.

cjm_substrate.core.errors

• CapabilityCancelledError class — Cooperative cancellation signal raised from ToolCapability.check_cancel().
• CapabilityConfigError class — Unknown / invalid keys in a config dict against a capability's config schema.
• CapabilityDisabledError class — JobQueue / execute_capability rejected: the capability is currently disabled.
• CapabilityError class — Base for substrate-recognized capability exceptions.
• CapabilityFatalError class — Bug / irrecoverable state. The capability cannot complete this job; retrying won't help.
• CapabilityInputError class — User-fixable error: bad config, invalid argument, missing file.
• CapabilityNotLoadedError class — Caller submitted to a capability that was never loaded.
• CapabilityResourceError class — Resource exhaustion: GPU VRAM, system RAM, disk full.
• CapabilityTimeoutError class — A per-job timeout fired before the capability finished.
• CapabilityTransientError class — Temporary failure: timeout, network blip, brief resource contention.
• JobError class — Structured failure summary recorded on a completed Job.
• ResourceShortfall class — Quantitative gap between what a capability needed and what was available.
• TracebackPolicy class — How much exception detail the substrate records on a JobError.
• WorkerOOMError class — The worker subprocess died with a kill-signal during an active execute call.
• classify_exception function — Return the substrate category for any exception.
• map_bare_exception_to_job_error function — Convert any exception into a structured JobError.

cjm_substrate.core.journal_store

• JournalEvent class — One durable observability record (CR-14).
• JournalStore class — Protocol for the durable account-of-action (CR-14).
• LocalJournalStore class — SQLite-backed default JournalStore (CR-14).
• SubstrateEventType class — Journal vocabulary beyond the job-scoped JobEventType set (CR-14).
• append function — Persist one event; sets and returns event.seq.
• count function — Total journal rows (volume regression checks).
• query function — Filtered read; all filters AND-combined.
• terminal_state_events function — The durable job history (_history migration rider): terminal

cjm_substrate.core.manager

• CapabilityBinding class — Pre-bound view of a single capability through a shared CapabilityManager.
• CapabilityManager class — Manages capability discovery, loading, and lifecycle via process isolation.

cjm_substrate.core.manifest_format

• CodeSection class — Code-derived facts refreshed by cjm-ctl regenerate-manifest.
• DriftTracking class — Witness hashes for drift detection.
• InstallSection class — Deployment-specific facts populated at install time.
• ManifestV2 class — Top-level v2.0 manifest with four named sections plus format_version.
• compute_config_schema_hash function — Hash a JSON Schema with stable canonicalization.
• compute_structural_surface_hash function — Hash a structural surface with stable canonicalization.
• load_manifest function — Load a manifest file and return a typed ManifestV2.
• manifest_to_dict function — Serialize a ManifestV2 to a v2.0 dict.
• write_manifest function — Serialize a ManifestV2 to disk in v2.0 nested layout (indent=2).

cjm_substrate.core.metadata

• CapabilityInstance class — Per-instance runtime state for a loaded capability (CR-10 multi-instance).
• CapabilityLoadSpec class — One entry in CapabilityManager.load_capabilities_concurrent's batch input (CR-10).
• CapabilityMeta class — Metadata about a capability.
• ResourceRequirements class — Binary hard-facts about what a capability needs to run (Phase 5a).

cjm_substrate.core.platform

• build_conda_command function — Build a complete conda/mamba/micromamba command.
• conda_env_exists function — Check if a conda environment exists (cross-platform).
• download_micromamba function — Download and extract micromamba binary to the specified path.
• ensure_runtime_available function — Check if the configured conda/micromamba runtime is available.
• get_conda_command function — Get the conda/mamba/micromamba base command with prefix args for local mode.
• get_current_platform function — Get current platform string for manifest filtering.
• get_micromamba_binary_path function — Get the configured micromamba binary path for the current platform.
• get_micromamba_download_url function — Get the micromamba download URL for the specified or current platform.
• get_popen_isolation_kwargs function — Return kwargs for process isolation in subprocess.Popen.
• get_python_in_env function — Get the Python executable path for a conda environment.
• is_apple_silicon function — Check if running on Apple Silicon Mac (for MPS detection).
• is_linux function — Check if running on Linux.
• is_macos function — Check if running on macOS.
• is_windows function — Check if running on Windows.
• run_shell_command function — Run a shell command cross-platform.
• terminate_process function — Terminate a subprocess + its entire process subtree (grandchildren, etc).
• terminate_self function — Terminate the current process (for worker suicide pact).

cjm_substrate.core.ports

• Composition class — A static DAG of capability-invocation nodes, submitted as one unit.
• CompositionBindingError class — An OutputRef could not be resolved against the producer's recorded
• CompositionNode class — One capability invocation in a composition.
• CompositionNodeRun class — Live state of one node within a composition run.
• CompositionRun class — Tracks a submitted composition through execution (lives in
• CompositionValidationError class — A composition failed submit-time validation (duplicate ids, unresolved
• NodeState class — State of one composition node (and, for the terminal subset, of a
• OutputRef class — Binding marker: this kwarg's value comes from an upstream node's result.
• extract_output_field function — Extract a field from an upstream result for binding into a kwarg.
• new_composition_run function — Validate a composition and build its run record.
• resolve_node_kwargs function — Materialize a node's kwargs by resolving its OutputRef markers.
• validate_composition function — Validate a composition and return its derived dependency map.

cjm_substrate.core.proxy

• RemoteCapabilityProxy class — Proxy that forwards capability calls to an isolated Worker subprocess.

cjm_substrate.core.queue

• CancelPhase class — Phase of a cancellation in progress (CR-6 + CR-4 pairing).
• Job class — A queued capability execution request (CR-6 reshape; stage-3 composition
• JobEvent class — A push-based job event (CR-6; stage-3 composition tags).
• JobEventType class — Push-based job event types (CR-6; stage-3 composition rework; CR-14
• JobQueue class — Resource-aware multi-lane job queue with journal-primary observability
• JobQueueDependencies class — Substrate dependencies the JobQueue requires (CR-6 + stage 3).
• JobStatus class — Status of a job in the queue.
• QueueStats class — Aggregate counts returned by JobQueue.get_stats() (CR-6).
• ResourceSnapshot class — Point-in-time resource usage for one job (CR-6 Stage 3).

cjm_substrate.core.scheduling

• PermissiveScheduler class — Scheduler that allows all executions (Default / Dev Mode).
• ResourceScheduler class — Abstract base class for resource allocation policies.

cjm_substrate.core.secret_store

• LocalSecretStore class — File-backed default SecretStore (0600 JSON under secrets_dir).
• SecretStore class — Protocol for resolving per-capability secrets (API keys, tokens).
• delete_secret function — Remove a secret, pruning now-empty capability/scope containers.
• get_secret function — Resolve a secret value.
• list_keys function — Return the names of secrets stored for a capability (never the values).
• set_secret function — Persist a secret value.

cjm_substrate.core.wire

• CallEnvelope class — Substrate-owned per-call identity + control block (CR-14 / CR-15).
• FileBackedDTO class — Protocol for Data Transfer Objects that serialize to disk for zero-copy transfer.
• begin_account_capture function — Start a fresh account list for the current call span (worker endpoint
• drain_accounts function — Return + clear the current span's recorded accounts ([] outside a span
• flat_from_dict function — Default reconstruction for FLAT wire DTOs (no nested-DTO fields).
• get_call_envelope function — The current call envelope, or None outside any call span.
• record_account function — Record one substrate-family account for the current call span.
• reset_call_envelope function — Restore the prior envelope (always pair with set_call_envelope in finally).
• set_call_envelope function — Set the current call envelope; returns the token for reset_call_envelope.
• wire_decode function — Reconstruct a typed result from its tagged envelope (host side).
• wire_encode function — Wrap a registered wire DTO in its tagged envelope (worker side).
• wire_type function — Register a dataclass as a typed wire DTO under kind.

cjm_substrate.core.worker

• EnhancedJSONEncoder class — JSON encoder that handles dataclasses and other common types.
• create_app function — Create FastAPI app that hosts the specified capability.
• parent_monitor function — Monitor parent process and terminate self if parent dies.
• run_worker function — CLI entry point for running the worker.

cjm_substrate.core.workspace

• Workspace class — A resolved workspace: the marker-rooted directory owning pipeline artifacts.
• WorkspaceError class — A workspace was named (flag or env) but could not be resolved.
• find_workspace_root function — Upward walk — the git-style discovery that makes workspace identity launch-cwd-independent.
• init_workspace function — Declare a workspace: write the marker + create the conventional layout.
• relativize_recorded function — Writer half of the 5daadfc4 recording contract (rung f).
• resolve_recorded_tree function — Reader half of the recording contract (rung f; anchor rule ratified 2026-07-19).
• resolve_workspace function — Resolve the active workspace: explicit flag > CJM_WORKSPACE env > upward walk > None.
• workspace_doctor function — Integrity-check skeleton for the workspace doctor verb.

cjm_substrate.utils.cache_paths

• cache_dir_for_config function — Return (and optionally create) a per-(input-content, config) cache directory.
• list_cache_entries function — Enumerate all per-config cache directories for a given (input, action).
• prune_cache_for_input function — Delete per-config cache directories for (input, action), optionally

cjm_substrate.utils.hashing

• hash_bytes function — Compute a hash of byte content.
• hash_dict_canonical function — Hash a dict via canonical JSON encoding.
• hash_file function — Stream-hash a file without loading it entirely into memory.
• verify_hash function — Verify content against an expected hash string.

cjm_substrate.utils.sidecar

• SidecarState class — One JSON sidecar file: never-raise load, merge-on-save, best-effort write.

cjm_substrate.utils.validation

• config_to_dict function — Convert a configuration dataclass instance to a dictionary.
• dataclass_to_jsonschema function — Convert a dataclass to a JSON schema for form generation.
• dict_to_config function — Create a configuration dataclass instance from a dictionary.
• extract_defaults function — Extract default values from a configuration dataclass type.
• validate_config function — Validate all fields in a configuration dataclass against their metadata constraints.
• validate_field_value function — Validate a value against field metadata constraints.

Dependencies

Depends on: fastapi, fastcore, httpx, psutil, pyyaml, typer, uvicorn Used by: cjm-capability-pyannote, cjm-capability-pysbd, cjm-context-graph-projection, cjm-graph-storage-adapter-interface, cjm-markdown-decompose-core, cjm-sentence-segmentation-adapter-interface, cjm-speaker-diarization-adapter-interface, cjm-transcript-correction-core, cjm-transcript-correction-qt, cjm-transcript-correction-tui, cjm-transcript-decomp-core, cjm-transcript-decomp-qt, cjm-transcript-decomp-tui, cjm-transcription-core, cjm-transcription-qt, cjm-transcription-tui, cjm-vad-adapter-interface, cjm-workflow-hub-qt, cjm-workflow-hub-tui