OpenDDE Harness 0.0.5

OpenDDE Harness 0.0.5 adds fixed-length minibinder design, improves final candidate selection, and separates internal runtime storage from the execution workspace. It also standardizes user-facing residue numbering and improves multi-GPU scheduling and design observability.

Upgrading

This release includes configuration and result-schema changes. Review existing workflows before restarting production tasks.

• Residue positions are now one-based and inclusive. Update legacy zero-based YAML positions, including CDR regions, designable/fixed residues, and target hotspots; for example, "25:34" becomes "26:35" when referring to the same residues. These are sequence positions, not PDB/mmCIF author residue IDs. Internal array indices and compute contracts remain zero-based; do not blindly convert internal payloads. Legacy inputs that do not contain zero can still be accepted with a different meaning, so validation alone is not a migration check.
• Internal storage no longer lives in the execution workspace. Assistant instructions, custom skills, sessions, exports, memory runtime state, and shadow checkpoints are stored under the active configuration file's parent directory, and workspace-specific state is scoped by the resolved workspace path. User project files and project AGENTS.md stay in place. Startup warns about legacy data but does not migrate it automatically.
• Multi-GPU folding defaults to candidate parallelism. A fold GPU list now identifies the candidate-worker pool. Set compute.placement.cp_degree explicitly above 1 to use context parallelism; the fold GPU list must then match that degree.
• Post-filter consumers must handle unranked candidates. Eligible candidates outside the final selection can have rank: null and selection_status not_selected or not_shortlisted. Neither status means a hard eligibility failure.

Preview legacy workspace migration before starting the updated client or initializing assistant templates:

ddeharness workspace migrate --workspace /path/to/workspace --config /path/to/config.json

Stop clients, Python workers, and memory services before applying it:

ddeharness workspace migrate --workspace /path/to/workspace --config /path/to/config.json --apply

Migration preserves backups and supports guarded rollback. Process verification currently requires Linux; other platforms fail closed. Historical memory notes are archived for review, not automatically imported into the external memory service. See workspace storage and migration for recovery and rollback details.

Added

• Minibinder design and optimization: select design.type: minibinder for a fixed-length, single-chain binder with explicit mutable positions. Supports complete sequences and mutable X placeholders, full redesign, point mutation, and inverse folding. Fixed residues and point-mutation budgets are enforced. The mode keeps the sequence length fixed, generates no backbone coordinates, requires a general-protein checkpoint with local or docker folding, and needs a successful parent structure for inverse folding. See protein-design YAML for the full field reference and mode constraints.
• Minibinder-specific scoring and validation: dedicated confidence/contact loss, interface and hotspot gates, prompts, memory separation, and dashboard labels. Antibody-only CDR/framework penalties and developability checks are not applied; unsupported QC properties remain unknown. Minibinder loss values are on their own scale and are not comparable with antibody-mode losses.
• Design-mode-aware checkpoint preparation: prepare antibody, minibinder, or both checkpoints with ddeharness compute prepare --mode local --design-mode antibody|minibinder|both. Published download sizes and SHA256 values are verified, and task readiness checks validate checkpoint availability, readability, and nonzero size on the compute host; these checks do not establish scientific suitability. Explicit checkpoint overrides are preserved for folding and refolding.
• Task status from the CLI: use ddeharness protein-design status --task-id <task-id>, add --json for the saved snapshot, or omit the task ID to list local tasks.
• Per-agent LLM usage in the design dashboard: inspect input/output tokens, cache reads/writes, and cache-hit rate. Missing usage remains unknown, and partial totals are explicitly marked.

Changed

• Bounded PostFilter selection builds a deterministic multi-metric, sequence-diversity-aware shortlist of up to max(80, top_k) eligible candidates. The agent returns only the requested Top-K decisions with bounded explanations, while full candidate evidence remains in the artifact. The shortlist reduces model input rather than scoring biological quality, and it does not guarantee a globally optimal Top-K; large explicit top_k values still raise input and output costs.
• Final results are auditable: shortlist membership is retained, and selected, unselected, non-shortlisted, and hard-rejected candidates are distinguished. Final refolded candidates share the candidate exploration interface, including available metrics and structures.
• Candidate-parallel GPU execution distributes fold batches across resident single-GPU workers in one compute container, uses available subsets of the configured pool, and avoids leasing excess devices when fewer candidates need processing. Conflicting fold/ESM/MPNN leases remain coordinated.
• Workflow context and skill loading use stable role-specific system prompts, shared context-budget checks, and request-scoped skill catalogs. Built-in and advisory skills are loaded on demand and validated against the allowed catalog.
• Memory isolation keeps external long-term memory as the sole recall source when enabled, avoids duplicate local injection, and skips recall and extraction when memory is disabled. Memory-library naming is confined to the adapter boundary.

Fixed

• Invalid PostFilter rankings are retried, and truncated selections are rejected instead of accepting incomplete JSON as a finished result.
• When the PostFilter agent fails, selection falls back to deterministic objective ordering across all eligible candidates while preserving the error and selection mode for inspection.
• Failed-refold candidate decisions are preserved, fallback ranking is corrected, and cancellation is honored through final selection and export.
• Residue-position conversion at user and LLM boundaries is corrected, and sequence hotspots are mapped to structure residue IDs for contact analysis.
• SolubleMPNN clients are reused independently of sampling seeds, with seeds passed to individual sampling requests.
• Redundant code-fence markers are removed from terminal Markdown rendering.

Included changes

• #7 — fix: unify post-filter results and retry invalid candidate rankings
• #9 — feat(plugin): add minibinder workflows and optimize design context
• #10 — feat(protein_design): align residue numbering and separate workspace storage
• #12 — fix(protein_design): bound post-filter selection

OpenDDE Harness 0.0.4

OpenDDE Harness 0.0.4 introduces a pi-tui terminal interface, a shared pi-ai model service, and simpler provider and memory setup.

Upgrading

Run ddeharness onboard after upgrading from 0.0.3, before starting the TUI. This release changes the configuration format, provider credentials, and memory settings. Incompatible configuration is rejected with instructions to reconfigure; it is not migrated automatically.

Onboarding moves incompatible configuration and related compute and memory service settings into backup-<timestamp>/, then creates a new setup. Weights, sessions, and memory notes are preserved. Sign in to Codex again through the new model service; MiniMax now uses an API key instead of device sign-in.

Subsequent onboarding runs preserve configuration with a matching schemaVersion and prefill each step with existing values. Use ddeharness onboard --fresh to back up the current configuration and start over.

Highlights

• Replaced the Ink/React terminal UI with pi-tui for more responsive streaming, long transcripts, and window resizing.
• Unified model access through a Node-based pi-ai service shared by the Python agent, provider login, and long-term memory.
• Added model discovery for declared OpenAI-compatible endpoints and pi-style model selection with saved scopes.
• Simplified onboarding with a shared login flow, compact choices, and clearer download and service-start progress.

Changed

Providers and models

• Provider configuration now follows pi-ai's models.json structure. Built-in providers use their canonical pi IDs; custom providers must declare both baseUrl and api. The protocol is explicit and is never inferred from the URL.
• Renamed provider and model settings: wire to api, extraHeaders to headers, apiBase to baseUrl, and modelOverlay to entries in models. Model fields label, contextWindowTokens, and maxOutputTokens are now name, contextWindow, and maxTokens. Corresponding CLI flags are --api, --headers, and --base-url.
• Credentials use apiKey or "login": "oauth". OAuth credentials live in the model service's credential store. Providers can also use their standard environment variables; endpoints that require no authentication need no key. Codex sign-in is available through ddeharness provider login openai-codex. MiniMax uses providers.minimax.apiKey or providers.minimax-cn.apiKey.
• Model IDs must include a provider, such as openai-codex/<model> or my-vllm/qwen3-32b. Removed agents.defaults.provider; routing and model limits now use the same qualified ID. Bare IDs are rejected with instructions to correct them.
• /model lists models from connected providers, with scoped/all views and current/default indicators. /scoped-models saves the selection in agents.scopedModels. In the TUI, connect providers through /login; the model picker's add-model, remove-model, and disconnect actions are removed.
• Declared OpenAI-compatible endpoints discover models through /models. Results are cached across restarts and refreshed when the picker opens if the endpoint's catalog is at least 24 hours old. Newly declared endpoints are queried immediately.
• Model limits, input modalities, and pricing come from the model service's catalogs and endpoint metadata. Configure model entries through providers.<id>.models or ddeharness provider model set. Unknown limits remain unknown rather than receiving an arbitrary fallback.
• ddeharness provider list shows configured providers. ddeharness provider reset <id> removes the entry and signs out its OAuth credential when applicable.
• Removed the legacy Python provider routes, bundled catalog tables, and custom OAuth flows. Azure OpenAI uses azure-openai-responses with api-version=v1; Chat Completions-only deployments can use openai-completions with the deployment URL as baseUrl.
• Web search now uses the shared web_search tool for every provider: Brave when tools.web.braveApiKey is configured, otherwise DuckDuckGo. Provider-hosted search and its citations are removed. Reasoning embedded in response content as <think> tags is no longer stripped.
• Temperature is sent only when configured on the model entry, including in protein-design tasks. Requests carry the session's cache key for prompt-cache reuse across the conversation.
• Transient assistant failures retry with backoff and a visible turn.retry event, up to agents.defaults.llmRetries (default 3). Partial text from a failed attempt is replaced on retry; deterministic failures stop immediately. Removed fallbackModels. Pre-response failures use the turn's retry budget; Retry-After is no longer honored before the first byte.
• Sessions whose saved model is unavailable use the default model and display a one-time notice on the next turn, with instructions to select a replacement through /model <provider>/<model>.

Memory and context

• Upgraded the memory library to 1.3.1. Long-term memory uses the conversation's default model through the shared model service, including Codex sign-in. Separate memory model, key, and endpoint settings are no longer needed.
• Removed embedding and reranking roles; memory recall uses keywords. Memory configuration and service management now live in the memory plugin, and onboarding only asks whether to enable memory.
• The memory home is fixed at ~/.opendde_harness/memory; previously configured roots are ignored. The host's memory writer again records episodes and the user profile.
• History selection is deterministic. Removed the Curator's model-based planning loop and its context.curatorModel and context.curatorProvider settings.
• Skill selection no longer makes query-rewriting or filtering LLM calls. Catalogs with at most 40 eligible skills are listed in a stable order; larger catalogs select up to 5 skills using BM25. Skills with unmet requires are excluded, and bodies are loaded only through use_skill.
• Removed skillForge.injectionMode and the rewrite* and llmGate* settings, including rewriteEnabled, llmGateEnabled, and llmGateModel. Setting skillForge.enabled=false or skillForge.router.enabled=false now removes the skills block from the prompt.

Onboarding and protein design

• Onboarding and /login share the same provider authentication flow. The wizard offers featured providers and an OpenAI-compatible option; additional providers are available through /login. Advanced protocol configuration, including Azure and --api, remains available through ddeharness provider set.
• Onboarding uses the TUI's branding, palette, translated step labels, and compact layout. Short choices appear in a single row with Tab navigation. Protein-design setup groups compute settings, folding, and weights into sections, and the completion screen uses aligned summaries.
• Service startup displays the current operation and elapsed time, followed by one result line. Compute downloads share a progress display; per-file details and asset bookkeeping go to logs. Memory startup and recovery choices use the same compact layout.
• Protein-design setup defaults to local folding with weights in the compute container. Existing folding-mode choices are preserved.
• Protein-design tasks inherit the initiating conversation's model unless the task YAML overrides it. Workers no longer load code from the launcher's working directory.

Terminal UI and updates

• Tools use pi's names and schemas: bash, read, write, edit, ls, grep, and find.
• /quiet-tools collapses tool rows to one line; Ctrl+O expands up to 1 KiB of a result. /tasks hide|show controls the design task bar, Ctrl+S saves the default model in the picker, and /tracing opens the tracing dashboard. The busy indicator now shimmers.
• Slash commands report timeouts immediately while the underlying operation finishes in the background. The command gateway rejects further commands with "another command is still running" until it completes, and late output is kept out of the terminal.
• PyPI update notices appear in the TUI status bar, once in the launch transcript, and at the top of ddeharness doctor. Notices show the appropriate upgrade command (uv tool upgrade or pip install --upgrade); source checkouts skip update checks.

Fixed

• Models discovered with only an ID no longer fail unconditionally with no_max_tokens. Discovery reads endpoint limit fields such as context_length, max_model_len, and max_completion_tokens, and fills missing limits from matching pi catalog entries. Supported OpenAI-compatible APIs omit an unknown output limit so the server can use its default; APIs that require a positive limit still reject unsized models.
• Compute startup errors distinguish an exited container from a service that is still unresponsive. Set OPENDDE_HARNESS_COMPUTE_KEEP=1 to retain the next container for inspection with docker logs.
• Memory-service startup errors identify exhausted inotify instances and show the command to raise the limit alongside the original error.
• Fixed a CancelledError after saving a provider key during onboarding: leaving the login step no longer cancels the model service's reader.
• Corrected footer colors and prevented the estimated-token marker from resetting the rest of the line's color. The first footer line now shows only the working directory and Git branch.
• Kept the status area at two rows in both idle and active states, preventing layout jumps and separating replies from the editor.
• Added consistent spacing and dim colors to transcript notices, and unified keyboard-hint capitalization in /model and /scoped-models.

OpenDDE Harness 0.0.3

OpenDDE Harness 0.0.3 adds per-model thinking controls, searchable model lists, and clearer progress reporting in the terminal.

Added

• Set per-model thinking levels with /thinking <level> (alias /reasoning) or ddeharness provider model set <provider> <model> --reasoning-effort <level>. Settings are saved in modelOverlay and adapted to the model's supported controls.
• Set the current model's context window with /context 128k, or clear the override with /context default.
• The /model picker discovers models from compatible relay and local endpoints, puts the current model first, and supports typing to filter providers, models, and endpoints.
• Ctrl+T toggles thinking blocks; Ctrl+O expands or collapses tool output. Both apply to live output and conversation history, and save the display preference.

Changed

• The default thinking level is now medium unless overridden per model. Set agents.defaults.reasoningEffort to null to leave it to the provider; Codex subscriptions still default to medium and request reasoning summaries.
• Relay and self-hosted models can inherit catalog limits and reasoning support when their model name has a unique match. Use modelOverlay if the endpoint has lower limits.
• The activity line shows elapsed time, output tokens, retry progress, and how to interrupt.
• Model picker actions use Ctrl+A to add, Ctrl+X to delete, Ctrl+E for endpoints, and Ctrl+D to disconnect. q now filters the list; Esc clears the filter before leaving.

Fixed

• Improved reasoning parameter handling for OpenAI-compatible endpoints, DeepSeek, Z.ai, DashScope, and OpenRouter.
• Output token counts now combine usage reported for completed calls with an estimate for the call still streaming, replacing that estimate when usage arrives.
• Reduced accidental removal of ordinary reasoning text when filtering echoed status messages.

OpenDDE Harness 0.0.2

OpenDDE Harness 0.0.2 improves model configuration, context limits, streaming recovery,
and MCP connection handling. After upgrading, run ddeharness onboard again to reconfigure.

Added

• Explicit protocol selection for OpenAI-compatible endpoints through wire: chat by default, or responses for the OpenAI provider. Configure it per provider or per model through modelOverlay. CLI and TUI configuration screens show the selected protocol.
• Per-model context window and output token limits through modelOverlay, with offline lookups from built-in data, provider-specific models.dev entries, LiteLLM, and canonical models.dev entries. Unknown context windows remain unknown; unknown output limits use an estimated 16384-token ceiling.
• llm_first_token_timeout and llm_idle_timeout to limit the initial wait and gaps between streaming events.
• curator_timeout_seconds to bound curator planning, with a deterministic fallback when planning times out.

Fixed

• Corrected reasoning parameters for supported models served through relays, including DeepSeek, Z.ai, DashScope, OpenRouter, and OpenAI-compatible endpoints.
• Retryable streaming failures now support bounded retries after partial output. The TUI clears partial text before retrying, and multi-endpoint configurations can avoid failed endpoints when another endpoint is available.
• Codex subscription models now use dedicated built-in limits rather than API model entries: 272k context and 128k output, with a 128k context window for the spark tier.
• Removed the 65536-token context fallback. History is no longer trimmed against an assumed window when the model's context limit is unknown.
• History trimming preserves tool results that belong to protected tool calls.
• Isolated MCP transport failures from the agent task. Tool calls remain subject to their configured timeout.
• Preserved reasoning identifiers when replaying Responses API history with tool calls.
• doctor and onboard recognize managed runtime code that can be prepared at first start from cached sources, avoiding unnecessary compute prepare prompts for that case.
• Moved synchronous history trimming and curator processing off the event loop to reduce terminal stalls.
• Improved protocol-mismatch errors and diagnostics for ambiguous 404 responses.

Changed

• Updated the configuration format.
• llm_call_timeout now applies only to non-streamed calls.
• The curator uses the agent's model unless a separate model is configured.
• Raised the minimum LiteLLM version to 1.100.0.

OpenDDE Harness 0.0.1

Introducing OpenDDE Harness - our first public preview!

We're excited to share OpenDDE Harness, an open-source tool for agentic antibody design. It brings LLM-guided sequence design and OpenDDE structure prediction into one workflow, from preparing a target to reviewing candidate sequences and structures.

Describe your design task in natural language, review the plan, and follow the results as the agent proposes, evaluates, and refines candidates. This first release supports VHH, scFv, and paired VH/VL binders while preserving configured fixed residues. A terminal UI guides setup and design, and a tracing dashboard lets you inspect candidate structures, scores, and progress across iterations.

Alongside this release, we're sharing a technical report and [design examples(https://github.com/aurekaresearch/OpenDDE-Harness/tree/v0.0.1/docs/examples) to introduce the approach and help you get started. Install from PyPI or source using the installation guide, then follow the onboarding guide to connect your LLM provider and compute service.

This is an early preview, and you may encounter bugs. Please open an issue with your feedback, reproduction steps, and ddeharness doctor --json output when relevant. Thank you for trying OpenDDE Harness and helping us make it better!

Full Changelog: https://github.com/aurekaresearch/OpenDDE-Harness/commits/v0.0.1