Skip to main content

Known Issues

This page is intentionally blunt. It is for engineering continuity, not marketing. Every limitation listed here is a boundary, not a promise.

Unsupported in v0.1.0

These surfaces exist in code or registry but are not live for users.

SurfaceStatusDetail
Native Windows installerUnsupportedNo Windows-native install path. WSL is best-effort.
Binary artifact-only update via ESTACODA_UPDATE_ARTIFACTReachable, not recommendedThe artifact path exists in code but is not the public v0.1.0 update mechanism. Use estacoda update or package-manager routing.
Cloud browser providers except BrowserbaseRegistered, not livebrowser-use, Firecrawl, and Camofox are registered but cannot create live sessions. Browserbase is implemented behind explicit spend approval.
Web research providers beyond Brave, DDGS, and fetchRegistered, not liveFirecrawl, Parallel, Tavily, Exa, and SearXNG are registered stubs. Brave and DDGS provide live search; guarded built-in fetch extraction is live.
Anthropic Messages API adapterCatalog-known, not runnableCode exists but is not runnable in the current build.
MiniMax LLM adapterCatalog-known, not runnableAppears in metadata but is not a live inference route.
Nous LLM adapterCatalog-known, not runnableAppears in metadata but is not a live inference route.
Arbitrary external memory providersUnsupportedOnly the built-in file provider constructs a live provider. Named providers without code implementation are rejected.

Experimental in v0.1.0

These features are code-gated or maturity-marked. Enable them only if you understand the risk.

FeatureGateRisk
Session compressioncompression.enabled and compression.experimental both trueExperimental-only. Disabled by default.
Agent Evolution autonomy above suggestskills.autonomy modes above suggestRecords reviewable evidence/proposals and shadow-only decisions. Real auto-promotion, auto-rollback, and automatic local skill creation are not active.
Skill evolution/proposal/promotion workflowsskill.propose_patch, skill.rollbackGoverned but not fully autonomous. Promotion runs eval gates.
Workflow runtime integrationSQLite session DB requiredBest-effort. Wires only when the session database is SQLite.
Local TTS providers neutts and kittenttsDeferredThese local/offline TTS providers are not implemented. Edge TTS is implemented separately as a networked provider.
Mistral TTS/STTDeferredNot implemented.
Gateway auto-TTSvoice.autoTts: truePer-reply and per-hour caps apply.
Deferred browser cloud providersbrowser.backend accepts legacy namesFirecrawl and Camofox browser backends report unavailable status. browser-use remains a deferred cloud provider.
Deferred web research providersweb.backend accepts stub namesFirecrawl, Parallel, Tavily, Exa, and SearXNG report unavailable even when configured. DDGS requires explicit managed Python setup and does not auto-install during runtime search.

Present but not live-proven

These channels and providers exist in code but lack live validation evidence for v0.1.0.

SurfaceEvidenceNote
Discord channelExperimentalAdapter is present; live validation is incomplete.
Google LLM providerConfigurable/catalog-knownConfig path exists but not live-proven in this build.
Anthropic LLM providerCatalog-knownMetadata and catalog entries exist, but it is not exposed as a setup/model-picker route and is not runnable in this build.

Known runtime limitations

  • doctor --live can succeed with [empty] response text for some providers.
  • OpenRouter works at runtime but can miss exact-content fidelity checks.
  • Local / Custom OpenAI-compatible endpoint support is implemented but not live-proven in this environment.
  • MCP stdio is live-proven; HTTP and broader third-party coverage need operator validation.
  • MCP workspace-trust ergonomics are coarse-grained.
  • Memory rendering is selective but not ranked. No freshness/staleness handling.
  • On non-vision providers, image analysis degrades to metadata-only.
  • Gateway status reports readiness, not real background-process liveness.
  • Full runtime CLI localization is incomplete. Arabic terminal rendering supports shaped, bidirectional Arabic, but not every CLI string is localized.
  • Evaluation substrate exists but is not a scored automated benchmark.
  • Internal alpha harness is manual and not yet a strict release gate.
  • Provider message content support was widened for vision, but many places still assume string content conceptually.
  • Product logic mixed with formatting/delivery concerns in some channel paths.
  • Live provider capability detection deserves an explicit operator signal.
  • Session recall works, but richer lineage/history management is missing.
  • Profile effects beyond prompt guidance remain incomplete.
  • npm global install requires the package to be published to npm. Package metadata is publish-ready, but publication happens through the release process.
  • Homebrew install depends on an external tap (KemetResearch/homebrew-tap) that is not part of this repository.
  • Docker install depends on GHCR image availability and tag publishing.
  • Auto-stash on dirty worktree is not implemented.

OS support

  • macOS 11 Big Sur and later is the stated floor.
  • Linux: any modern distribution with systemd and glibc. Validated on Ubuntu 22.04+ and Debian 12+.
  • WSL2 is best-effort. Voice/microphone paths and systemd user services have known edge cases.
  • Termux is best-effort. The installer resolves a Termux layout but it is not a primary validation target.
  • Native Windows is unsupported.

What this means

A registered provider is not a working provider. A present channel is not a validated channel. An experimental feature is not a launch guarantee. Check this page before reporting a bug; the bug may be a documented limitation.

  • Testing — validation layers and evidence levels
  • FAQ — short answers on provider and channel maturity