Skip to content

ROADMAP.md

Latest commit

 

History

History
94 lines (77 loc) · 4.27 KB

ROADMAP.md

File metadata and controls

94 lines (77 loc) · 4.27 KB
Raw

ForgeClaw Roadmap

ForgeClaw is a deployment-focused fork of PicoClaw. The roadmap tracks work that matters for the active ForgeClaw deployment while preserving the ability to merge upstream PicoClaw changes.

1. Runtime Reliability

  • Keep async task delivery deterministic across restarts.
  • Continue consolidating tool delivery through the shared delivery coordinator.
  • Reduce duplicate final replies after media/file/message tools.
  • Keep task-registry records bounded, queryable, and useful for debugging.
  • Preserve topic/session routing across Telegram, cron, spawn, and delegate paths.

2. Context Management

  • Keep Seahorse compaction bounded and observable.
  • Prefer asynchronous compaction where it is safe.
  • Fail closed when context still exceeds provider limits after compaction.
  • Keep prompt assembly predictable by reserving budget for tools, non-history prompt sections, and required routing context.
  • Add focused regression tests for long-session and post-compaction behavior.

3. Tooling And Workflow State

  • Keep task_status as the primary user-facing progress/status command.
  • Retire duplicated status surfaces when they no longer add value.
  • Improve deterministic test coverage for tool loops, spawned work, and async completions.
  • Extract a shared path-scope validation package only if non-exec tools need the same workspace/symlink/allowed-path rules currently enforced by shellguard.

4. Provider And MCP Behavior

  • Keep OpenAI OAuth/Codex paths reliable and observable.
  • Preserve streamed output and streamed tool-call behavior in provider adapters.
  • Keep MCP transport failures explicit and fail-fast.
  • Maintain deferred MCP/tool discovery behavior so large tool inventories do not pollute ordinary prompts.

5. Channels And Media

  • Re-architect channel lifecycle so startup, retry, reload, inbound exposure, and readiness are supervised through explicit runtime state instead of being tightly coupled inside channels.Manager.
  • Keep shared webhook/HTTP registration tied to active channel runtimes rather than configured channel presence.
  • Make startup retry generation-aware and cancelable across reload/shutdown.
  • Keep gateway readiness aligned with actual channel delivery capability.
  • Keep Telegram forum-topic routing stable.
  • Preserve media-group handling and forwardable media captions.
  • Keep generated images and files deliverable without duplicate completion messages.
  • Keep channel feedback throttling controlled by real edit intervals.
  • Move reply / adjacent-followup / media-only interpretation toward an explicit inbound relation model so prompt assembly does not have to guess message boundaries from raw history.

6. Automation And Agent Workflows

  • Keep core workflow primitives deployment-agnostic.
  • Support durable queued work without assuming a specific domain or workspace.
  • Make spawned/delegated work observable through shared task status surfaces.
  • Keep webhook, cron, and manual trigger paths consistent.
  • Let deployments layer domain-specific agents and policies outside the core runtime.

7. Upstream Compatibility

  • Merge upstream/main regularly.
  • Keep fork-specific behavior documented in the root README fork note.
  • Avoid repo-wide renames of binary/config/module paths unless there is a strong reason.
  • Prefer small compatibility patches over broad rewrites that increase merge conflict cost.

8. Model Selection Architecture

  • Keep /model as the conversation-scoped model selector and /switch as the explicit workspace-wide operator path until /switch is deprecated.
  • Replace ad hoc override-agent cloning with a clearer effective-model resolution path.
  • Resolve model-selection state once per routed turn, then pass that binding through command/runtime and execution code instead of repeatedly re-reading per-session override state.
  • Keep invalid session overrides self-healing and bounded to the routed conversation key.
  • Preserve current provider/model selection semantics while reducing the amount of agent-instance mutation/cloning required to execute a session override.
  • After the binding layer is stable, consider moving from "derived AgentInstance" materialization toward per-turn provider/model resolution so session overrides stop carrying copied tool/router/provider bookkeeping.