v0.0.10

Role Model v0.0.10 — Updated UI, Message graph, Knowledge store and other Extensions

Recommended release. v0.0.10 supersedes v0.0.9.

What is fixed

• Adaptive endpoint recovery replaces the blanket ten-minute cooldown. Connection failures enter probation before short 5s/15s/60s backoff, provider 5xx failures recover on a 2s/10s/30s ladder, rate limits honor Retry-After, and successful or quiet endpoints reset automatically. A single half-open probe prevents retry storms.
• Configuration failures are no longer presented as timed outages. Authentication and quota failures stay blocked until configuration changes, while invalid requests do not trip the circuit breaker. Cooldown-only routing now reports endpoint_temporarily_unavailable with retry timing.
• Endpoint operations are visible and durable. The Providers UI shows circuit state separately from endpoint health, and bounded versioned state survives runtime restarts without carrying forward unsafe legacy cooldown records.
• The packaged launcher no longer kills an operational runtime. A valid HTTP 200 /healthz response with sessionBootstrap.status: degraded is now accepted as usable startup, so the UI remains available when only an optional provider endpoint is missing credentials.
• Unsafe startup states still fail closed. blocked, pending, unreachable, malformed, and non-200 health responses are still rejected.
• Message Graph update safety from v0.0.9 is included. Managed Artifact Store digest and encryption keys remain in the persistent Track B state directory and are reused across binary upgrades and restarts.

What the extensions enable

• Message Graph: Artifact Store and Event Log persist connected message/content nodes, edges, references, lifecycle state, and a causal audit/replay timeline.
• Memory and governed knowledge: Memory Store, Knowledge Store, Knowledge Worker, and Repository Context retain scoped memory and reviewable, versioned knowledge projections.
• Trace projection and RL-style improvement: Trajectory Signals projects routed traces into bounded success, latency, and route-choice signals; Profile Learner turns approved projections into reviewable routing proposals without silently training models online.
• Replay and evaluation: Replay Core, Evaluation Core, and Evaluation Runner Local provide deterministic replay and offline/local outcome comparison.
• Asynchronous, privacy-bounded learning: Background Evidence Scheduler and Crowdsourced Learning support lifecycle-tracked background work and authorized aggregate contributions with revocation.

Included production extensions

Extension	User capability
Artifact Store	Message Graph nodes, edges, references, persistence, and lifecycle.
Background Evidence Scheduler	Asynchronous evidence, evaluation, and projection jobs.
Crowdsourced Learning	Authorized, privacy-bounded aggregate contributions and revocation.
Evaluation Core	Evaluation jobs, scorers, scores, and comparisons.
Evaluation Runner Local	Local and offline evaluation execution.
Event Log	Causal audit and replay timeline for the Message Graph.
Knowledge Store	Versioned governed knowledge records.
Knowledge Worker	Reviewed, redacted, rollback-capable knowledge derivation and retrieval.
Memory Store	Scoped long-term memory and ingestion receipts.
Profile Learner	Context-aware estimates and reviewable routing-profile proposals.
Replay Core	Deterministic request, session, and graph replay.
Repository Context	Bounded repository context attached to requests when available.
Trajectory Signals	Bounded signals projected from routed traces.

Updating safely

Back up and preserve the runtime state directory when replacing application files. On Windows production installs it is normally %LOCALAPPDATA%\role-model-runtime; managed Message Graph keys live beneath the runtime scope's track-b\managed-keys directory. Do not regenerate or delete those keys during an update.

Release archives are available for Windows x64, Linux x64, macOS x64, and macOS arm64. Verify downloads against SHA256SUMS.txt.

Source range: v0.0.9..v0.0.10

---

Archived notes from removed superseded releases

The complete notes from v0.0.8 and v0.0.9 are retained below for release-history and audit context. Those releases contained startup defects and their GitHub release pages, downloadable assets, and tags were intentionally removed after v0.0.10 superseded them. Statements inside the archived text about the former release pages or asset immutability are historical quotations.

Original v0.0.8 release notes

> **Superseded release:** v0.0.8 has a production startup defect that can leave the Track B runtime in `runtime_initializing`. Install [v0.0.9](https://github.com/try-works/role-model/releases/tag/v0.0.9) instead. The v0.0.8 tag and assets remain immutable for auditability.

Role Model v0.0.8 — Persistent context, evidence, and learning extensions

Role Model v0.0.8 adds the complete 13-extension production runtime. Routed work can now become persistent, inspectable evidence that feeds memory, knowledge retrieval, replay, evaluation, and privacy-bounded learning workflows.

What this enables

Message Graph — inspect and replay connected work

Message Graph is the user-facing capability. It is implemented primarily by Artifact Store, which persists scoped message/content artifacts and owns the canonical graph: nodes, edges, references, leases, lifecycle, and garbage collection. Event Log supplies the compact append-only causal timeline for routing and extension activity. Together they retain the relationships behind a request—not just its final answer—so users can inspect, materialize, audit, or replay connected work later.

Keep useful context and turn it into governed knowledge

Memory Store retains scoped long-term memory and ingestion receipts, while Knowledge Store manages versioned knowledge records that can be promoted, listed, or tombstoned. Knowledge Worker derives and retrieves knowledge through review, redaction, rollback, and evaluation-aware projection boundaries. Repository Context can attach bounded, request-scoped repository context without making routing depend on that context being available.

This lets users build reusable knowledge from prior work while preserving provenance, scope, and lifecycle controls.

Project traces into replay, evaluation, and RL-style learning signals

Trajectory Signals converts routed traces into bounded success, latency, and route-choice signals. Replay Core builds deterministic request, session, and graph replay plans. Evaluation Core scores and compares outcomes, while Evaluation Runner Local runs those evaluations locally or offline. Profile Learner consumes approved projections to estimate contextual behavior and produce reviewable routing-profile proposals.

Combined, these extensions provide a trace-to-learning pipeline: project trajectories, replay comparable decisions, score outcomes, and propose better routing behavior. This supports offline and RL-style policy improvement; it does not silently train or modify models online.

Improve the system asynchronously and with privacy controls

Background Evidence Scheduler runs evidence, evaluation, and projection jobs asynchronously with retry and lifecycle receipts. Crowdsourced Learning supports previewed, authorized aggregate contributions and revocation, enabling users to contribute bounded learning evidence without treating raw private prompts as shared training data.

Operate extensions safely

The extension host exposes health and lifecycle controls so extensions can be inspected, enabled, disabled, restarted, or allowed to degrade independently. Optional context and learning features can fail closed or step out of the route without taking down the core router.

Included production extensions

Extension	What users can do
Artifact Store	Power the Message Graph by persisting message/content nodes, edges, references, and lifecycle.
Background Evidence Scheduler	Schedule and run asynchronous evidence, evaluation, and projection work.
Crowdsourced Learning	Preview, authorize, send, disclose, and revoke privacy-bounded aggregate contributions.
Evaluation Core	Create evaluation jobs, run scorers, and record scores and comparisons.
Evaluation Runner Local	Run and control evaluation workloads locally and offline.
Event Log	Supply the Message Graph causal timeline for audit and replay.
Knowledge Store	Read, write, promote, list, and tombstone governed knowledge records.
Knowledge Worker	Derive and retrieve reviewed, redacted, rollback-capable knowledge projections.
Memory Store	Read, write, clear, and track scoped long-term memory.
Profile Learner	Estimate context-aware behavior and propose reviewable routing profiles.
Replay Core	Plan, execute, route-only test, and cancel deterministic replays.
Repository Context	Add bounded repository context to a request when available.
Trajectory Signals	Derive and analyze bounded signals from routed trajectories.

Router and release improvements

• Codex subscription and operator-configured OpenAI endpoints are selected using endpoint compatibility metadata instead of brittle hard-coded model lists, improving alias routing and support for newer model IDs.
• Packaged-runtime validation now covers deterministic routed client flows, including alias-based text, image, and degraded-primary routing behavior.
• Release archives are published for Windows, Linux, and macOS with installer scripts and SHA256SUMS.txt.

Source range: v0.0.7..v0.0.8

Original v0.0.9 release notes

# Role Model v0.0.9 — Production startup and update-safety hotfix

Superseded. Install v0.0.10. v0.0.9 fixed managed Message Graph keys, but its packaged launcher could stop an otherwise operational runtime when only optional services were degraded.

What is fixed

• Adaptive endpoint recovery replaces the blanket ten-minute cooldown. Connection failures enter probation before short 5s/15s/60s backoff, provider 5xx failures recover on a 2s/10s/30s ladder, rate limits honor Retry-After, and successful or quiet endpoints reset automatically. A single half-open probe prevents retry storms.
• Configuration failures are no longer presented as timed outages. Authentication and quota failures stay blocked until configuration changes, while invalid requests do not trip the circuit breaker. Cooldown-only routing now reports ^[ndpoint_temporarily_unavailable with retry timing.
• Endpoint operations are visible and durable. The Providers UI shows circuit state separately from endpoint health, and bounded versioned state survives runtime restarts without carrying forward unsafe legacy cooldown records.
• Message Graph starts correctly in production. The runtime now provisions its managed Artifact Store digest and encryption keys on first launch.
• Updates preserve graph readability. Managed keys live under the persistent Track B state directory and are reused across runtime upgrades and restarts. Replacing the application files does not replace these keys.
• Unsafe key states fail closed. Missing halves of a key pair, invalid key lengths, and symlinked key files are rejected instead of silently rotating keys and making existing graph data unreadable.
• Explicit operator-managed keys remain supported. Deployments that supply their own key paths retain that behavior.
• The launcher waits for real readiness. It now checks /healthz and opens the UI only after sessionBootstrap.status is ready, rather than treating a shallow listening socket as successful startup.
• Manual update guidance is documented. The README and installation guide now explain how to upgrade while preserving the state directory and managed keys.

Extensions and user capabilities included

v0.0.9 contains the full 13-extension production runtime introduced in v0.0.8, with the startup and persistence fixes above.

Message Graph — persistent, connected work

Artifact Store powers the Message Graph by retaining message/content nodes, edges, references, leases, lifecycle state, and garbage-collection metadata. Event Log supplies the compact append-only causal timeline. Together they let users inspect, materialize, audit, and replay the connected work behind a request—not just its final response.

Memory and governed knowledge

Memory Store retains scoped long-term memory and ingestion receipts. Knowledge Store manages versioned knowledge records that can be promoted, listed, or tombstoned. Knowledge Worker derives and retrieves reviewed, redacted, rollback-capable knowledge projections. Repository Context can attach bounded request-scoped repository context when available.

Trace projection, replay, evaluation, and RL-style improvement

Trajectory Signals projects routed traces into bounded success, latency, and route-choice signals. Replay Core builds deterministic request, session, and graph replay plans. Evaluation Core scores comparable outcomes, while Evaluation Runner Local runs evaluations locally or offline. Profile Learner consumes approved projections and creates reviewable routing-profile proposals.

This provides a trace-to-learning workflow for offline and RL-style policy improvement; it does not silently train or modify models online.

Asynchronous and privacy-bounded learning

Background Evidence Scheduler runs evidence, evaluation, and projection jobs asynchronously with retry and lifecycle receipts. Crowdsourced Learning supports previewed, authorized aggregate contributions and revocation without treating raw private prompts as shared training data.

Extension operations

The extension host exposes health and lifecycle controls so extensions can be inspected, enabled, disabled, restarted, or allowed to degrade independently without taking down the core router.

Included production extensions

Extension	What it enables
Artifact Store	Message Graph nodes, edges, references, persistence, and lifecycle.
Background Evidence Scheduler	Asynchronous evidence, evaluation, and projection jobs.
Crowdsourced Learning	Authorized, privacy-bounded aggregate contributions and revocation.
Evaluation Core	Evaluation jobs, scorers, scores, and comparisons.
Evaluation Runner Local	Local and offline evaluation execution.
Event Log	Causal audit and replay timeline for the Message Graph.
Knowledge Store	Versioned governed knowledge records.
Knowledge Worker	Reviewed, redacted, rollback-capable knowledge derivation and retrieval.
Memory Store	Scoped long-term memory and ingestion receipts.
Profile Learner	Context-aware estimates and reviewable routing-profile proposals.
Replay Core	Deterministic request, session, and graph replay.
Repository Context	Bounded repository context attached to requests when available.
Trajectory Signals	Bounded signals projected from routed traces.

Updating from v0.0.8

Install v0.0.9 over the application version, but preserve the runtime state directory. On Windows production installs this is normally %LOCALAPPDATA%\role-model-runtime; managed Message Graph keys are retained beneath the runtime scope's track-b\managed-keys directory. Back up the state directory before a manual update. Do not regenerate or delete the managed keys when moving to a new binary.

Release archives are available for Windows x64, Linux x64, macOS x64, and macOS arm64. Verify downloads against SHA256SUMS.txt.

Source range: v0.0.8..v0.0.9