Skip to content

opencode-porting-tracker.md

Latest commit

 

History

History
279 lines (256 loc) · 20.3 KB

opencode-porting-tracker.md

File metadata and controls

279 lines (256 loc) · 20.3 KB

OpenCode Feature Porting Tracker

Started: Feb 5, 2026 Source: /tmp/opencode (anomalyco/opencode, dev branch) Comparison: /tmp/claude-code-vs-opencode-comparison.md

Status Legend

  • Not started
  • [~] In progress
  • Complete
  • [-] Blocked/Deferred

Phase 0: Setup

  • Run existing test suite, confirm all pass
  • Understand build system (bun, turbo)
  • Create working branch for changes

Phase 1: Easy wins (build codebase familiarity)

1. Compact with user instructions ✅

  • Research: read session/compaction.ts, server endpoint, TUI command
  • Implement: add optional instructions param to /compact command
  • Implement: pass instructions through to compaction prompt
  • Test: unit tests (schema, storage) + integration test (real LLM via Claude Max OAuth)
  • Verify: 891 pass, 0 fail
  • Files changed:
    • src/session/compaction.ts - instructions in process() and create()
    • src/session/message-v2.ts - instructions in CompactionPart schema
    • src/session/prompt.ts - pass task.instructions to process()
    • src/server/routes/session.ts - instructions in /summarize endpoint
    • src/cli/cmd/tui/component/prompt/index.tsx - handle /compact [text] in prompt
    • packages/sdk/js/src/v2/gen/sdk.gen.ts - instructions in summarize() method
    • test/session/compaction.test.ts - 3 new schema tests
    • test/server/session-summarize.test.ts - 6 new endpoint + storage tests
  • Complexity: Low
  • Notes: Verified end-to-end: instructions stored in compaction part AND influenced LLM summary output

2. PR review status in TUI

  • [-] Deferred: user uses Gitea/GitLab, not GitHub
  • Notes: Would need Gitea/GitLab API support instead of gh CLI

3. Session-linked PRs

  • [-] Deferred: user uses Gitea/GitLab, not GitHub
  • Notes: Would need Gitea/GitLab API support

4. Context fork for skills

  • [-] Skipped: OpenCode already supports this via its existing subagent/task system
  • Notes: Different mechanism than Claude Code's context: fork frontmatter, but same result. Commands support subtask: true for subagent execution.

Phase 2: Medium features (deeper codebase knowledge)

5. Prompt suggestions ✅

  • Research: read editor/prompt components, understand how responses complete
  • Design: background LLM call using small/fast model after agent finishes
  • Implement: SessionSuggestion module with generate() using Provider.getSmallModel()
  • Implement: Bus event session.suggestion → SSE → client store → placeholder text
  • Implement: Tab to accept into input, Enter to accept+submit, grayed-out display
  • Implement: experimental.suggestions config flag (opt-out)
  • Test: unit tests (891 pass), type check clean, integration test (real LLM via Google OAuth)
  • Verify: end-to-end SSE event received by client
  • Files changed:
    • src/session/suggest.ts - NEW - generates suggestions via small model (e.g. gemini-3-flash, claude-haiku-4.5)
    • src/session/prompt.ts - calls SessionSuggestion.generate() at loop exit (fire-and-forget)
    • src/config/config.ts - suggestions: z.boolean().optional() in experimental schema
    • src/cli/cmd/tui/context/sync.tsx - session_suggestion store + SSE event handler + clear on busy
    • src/cli/cmd/tui/component/prompt/index.tsx - suggestion memo, placeholder display, Tab/Enter handlers
    • packages/sdk/js/src/v2/gen/types.gen.ts - EventSessionSuggestion type in Event union
  • Complexity: Medium
  • Notes: Uses Provider.getSmallModel() (same as title generation) for cost/speed. Silently skips if model returns empty or session goes busy. Subtask sessions excluded.

6. Background bash tasks (Claude Code's Ctrl+B) ✅

  • Research: Claude Code backgrounds individual bash commands, not whole sessions
  • Design: Approach A - background individual bash tool calls (matches Claude Code behavior)
  • Study: PR #9681 (4,878 lines, unmerged) - adapted cleaner version without persistence
  • Implement: src/task/events.ts - Bus events (task.created, task.output, task.completed, task.killed)
  • Implement: src/task/index.ts - In-memory TaskManager (adopt, list, get, kill, read, tail, remove)
  • Implement: src/tool/bash.ts - Foreground process registry + migrateToBackground() function
  • Implement: src/tool/process-query.ts - LLM tool for agents to check on background tasks
  • Implement: src/server/routes/task.ts - REST API (list, get, output, tail, kill, migrate, foreground)
  • Implement: TUI Ctrl+B keybind (context-sensitive: only when session is busy + bash running)
  • Implement: /tasks command + DialogTasks (list, kill, view output)
  • Implement: Toast notification on background task completion
  • Test: type check clean, 887 unit tests pass, integration test verified full flow
  • Files changed:
    • src/task/events.ts - NEW - Bus event definitions
    • src/task/index.ts - NEW - TaskManager (in-memory, cleaned up on exit)
    • src/tool/bash.ts - Foreground process registry, migration support, callID in metadata
    • src/tool/process-query.ts - NEW - LLM tool (list, status, read_output, read_last_n, search)
    • src/tool/registry.ts - Register ProcessQueryTool
    • src/server/routes/task.ts - NEW - Task REST API
    • src/server/server.ts - Mount TaskRoutes at /task
    • src/cli/cmd/tui/app.tsx - /tasks command, toast on task.completed
    • src/cli/cmd/tui/component/dialog-tasks.tsx - NEW - Tasks dialog with kill/view
    • src/cli/cmd/tui/routes/session/index.tsx - Ctrl+B handler
  • Complexity: Medium
  • Notes: No persistence (matches Claude Code - tasks die with server). ~500 lines total vs PR #9681's 4,878. Ctrl+B is context-sensitive: only triggers when session busy + bash running, otherwise acts as cursor-left.

7. Persistent agent memory ✅

  • Research: OpenCode already has AGENTS.md/CLAUDE.md loading, global instructions, contextual loading via resolve()
  • Part A: /memory command - dialog to list/create/edit memory files in $EDITOR
  • Part B: .opencode/rules/*.md with path-scoping (like Claude Code's .claude/rules/*.md)
  • Implement: scan .opencode/rules/**/*.md and .claude/rules/**/*.md with gray-matter frontmatter parsing
  • Implement: unconditional rules (no paths frontmatter) loaded in system prompt alongside AGENTS.md
  • Implement: path-scoped rules (paths frontmatter with glob patterns) loaded contextually via resolve() when matching files are read
  • Implement: rules cached per-instance, claimed per-message (no duplicate injection)
  • Test: 11 new unit tests (unconditional loading, frontmatter stripping, path matching, claiming, compatibility)
  • Verify: 922 pass, 0 fail, type check clean
  • Files changed:
    • src/session/instruction.ts - RuleFile type, scanRules(), getRules() cache, rules in system() and resolve()
    • src/cli/cmd/tui/component/dialog-memory.tsx - NEW - Memory file browser with $EDITOR support
    • src/cli/cmd/tui/app.tsx - /memory command registration
    • test/session/instruction-rules.test.ts - NEW - 11 tests for rules loading
  • Complexity: Medium
  • Notes: Supports both .opencode/rules/ (native) and .claude/rules/ (compatibility). YAML frontmatter with paths field for glob-based scoping. Unconditional rules go in system prompt, path-scoped rules injected as <system-reminder> when agent reads matching files.

8. Task list with dependencies ✅

  • Research: OpenCode todo system (4-field flat list), Claude Code Tasks (full deps), 7 related issues
  • Fix #5612: Changed status from z.string() to z.enum(["pending","in_progress","completed","cancelled","blocked"]), priority to z.enum(["high","medium","low"])
  • Implement: depends_on optional field (array of todo IDs) in Todo.Info schema
  • Implement: blocked status — auto-set when deps are incomplete, auto-unblocked when deps complete
  • Implement: dependency validation — circular dep detection (DFS), strip refs to non-existent IDs, auto-resolve on update
  • Implement: TodoItem UI — ⊘ icon for blocked tasks, red color for blocked state
  • Fix #5934: Inject incomplete todos into compaction prompt so they survive context compaction
  • Update: todowrite.txt prompt with dependency instructions + example
  • Test: 14 new unit tests (schema validation, auto-block, auto-unblock, chain deps, multiple deps, invalid refs)
  • Verify: 936 pass, 0 fail, type check clean
  • Files changed:
    • src/session/todo.ts - Status and Priority enums, depends_on field, resolveDependencies(), hasCircularDeps(), hasUnresolvedDeps()
    • src/session/compaction.ts - Inject incomplete todos into compaction prompt context
    • src/tool/todowrite.txt - Dependency documentation and example
    • src/cli/cmd/tui/component/todo-item.tsx - statusIcon() function, blocked/cancelled rendering
    • test/session/todo.test.ts - NEW - 14 tests
  • Complexity: Medium
  • Notes: Also fixes issues #5612 (status validation) and #5934 (todos lost after compaction). Dependency resolution uses AND logic (all deps must complete). Circular deps are warned but not blocked to avoid breaking the LLM flow.

Phase 3: Complex features

9. Auto-memory ✅

  • Research: Studied OpenClaw's memory system (vector search, session-memory hooks, pre-compaction memory flush, memory_search/memory_get tools)
  • Design: Conservative approach — memory_save tool the LLM calls when user explicitly asks to remember something
  • Implement: MemorySaveTool — writes facts to .opencode/rules/memory.md with timestamps
  • Implement: Auto-inject via existing rules loading system (#7) — no new code needed, facts load in future sessions automatically
  • Also fixed: Todo list now injected into system prompt via Todo.systemContext() for reliable persistence across compaction (simpler than compaction prompt hack)
  • Test: 5 new unit tests (create file, append, date inclusion, directory creation, instruction system integration)
  • Verify: 944 pass, 0 fail, type check clean
  • Files created:
    • src/tool/memory-save.ts - NEW - MemorySaveTool tool definition
    • src/tool/memory-save.txt - NEW - Conservative usage instructions
    • test/tool/memory-save.test.ts - NEW - 5 tests
  • Files changed:
    • src/tool/registry.ts - Register MemorySaveTool
    • src/session/todo.ts - Added Todo.systemContext() for system prompt injection
    • src/session/prompt.ts - Inject todos into system prompt every turn
    • src/session/compaction.ts - Removed compaction prompt hack (replaced by system prompt injection)
  • Complexity: Low (conservative approach leverages existing rules infrastructure)
  • Notes: OpenClaw has much more sophisticated memory (embeddings, SQLite vector search, pre-compaction flush). Can layer those on later. Current approach: LLM writes to .opencode/rules/memory.md, rules system loads it automatically.

10. Vim mode

  • Research: read textarea-keybindings.ts, understand input handling
  • Implement: mode state (normal/insert)
  • Implement: normal mode motions (h/j/k/l, w/e/b, 0/$, gg/G, f/F/t/T)
  • Implement: operators (d/c/y) with motions and text objects
  • Implement: visual indicator of current mode
  • Test: write tests
  • Verify: run test suite
  • Files likely touched: cli/cmd/tui/component/textarea-keybindings.ts, prompt component
  • Complexity: Medium-High (mechanical but lots of cases)
  • Notes:

11. Compact with user instructions (enhanced) ✅

  • Implement: partial compaction with boundaryMessageID ("summarize before X")
  • Implement: validation that boundaryMessageID exists in message list, fallback to full compaction
  • Verify: type check clean, integrated with /compact command
  • Files changed:
    • src/session/compaction.ts - boundaryMessageID param in process() and create()
  • Complexity: Low (extension of #1)
  • Notes: Allows targeting specific conversation ranges for compaction

Phase 4: Agent teams (the big one) ✅

12. Agent teams - primitives ✅

  • Design: file-locked JSON state per team under .opencode/teams/<name>/
  • Design: TeamMember schema (name, sessionID, agent, status, model, planApproval)
  • Design: TeamTask schema (id, content, status, priority, assignee, depends_on)
  • Implement: src/team/index.ts - Team namespace: create, get, list, addMember, removeMember, setMemberStatus, setMemberPlanApproval, setDelegate, findBySession, cleanup. TeamTasks: list, update, add, claim, complete, resolveDependencies. File locking via Lock.write().
  • Implement: src/team/events.ts - 10 Bus events (Created, MemberSpawned, MemberStatusChanged, Message, Broadcast, TaskUpdated, TaskClaimed, PlanApproval, ShutdownRequest, Cleaned)
  • Implement: src/team/messaging.ts - TeamMessaging.send() and broadcast() inject synthetic user messages into recipient sessions
  • Test: 12 server route tests pass via bun test
  • Verify: 1030 pass, 0 fail
  • Files created:
    • src/team/index.ts - NEW - Team and TeamTasks namespaces
    • src/team/events.ts - NEW - Schemas and Bus event definitions
    • src/team/messaging.ts - NEW - Cross-session messaging
    • src/server/routes/team.ts - NEW - REST API (list, get, tasks, by-session, delegate toggle)
    • test/server/team-routes.test.ts - NEW - 12 route tests
  • Complexity: High

13. Agent teams - orchestration ✅

  • Implement: 9 team tools (team_create, team_spawn, team_message, team_broadcast, team_tasks, team_claim, team_approve_plan, team_shutdown, team_cleanup) behind OPENCODE_EXPERIMENTAL_AGENT_TEAMS flag
  • Implement: team_create with delegate mode (restricts lead to coordination-only by denying bash/write/edit/apply_patch)
  • Implement: team_spawn — spawns teammate as child session, starts SessionPrompt.loop() in background, auto-claims task
  • Implement: Plan approval (require_plan_approval: true) — teammate starts in read-only mode with write tools denied; lead uses team_approve_plan to grant access
  • Implement: Multi-model support — team_spawn accepts model param in provider/model format (e.g. anthropic/claude-sonnet-4-20250514, google/gemini-2.5-pro), validated via Provider.getModel() with fuzzy suggestions on error
  • Implement: Model resolution priority: explicit param > agent.model > lead's current model > Provider.defaultModel()
  • Implement: team_shutdown with negotiation (teammate can approve/reject)
  • Implement: Child session permission rules (deny team management tools, deny todowrite/todoread)
  • Implement: Idle notification when teammate loop finishes
  • Implement: POST /session/:sessionID/team-message route (follows shell() pattern)
  • Implement: SessionPrompt.teamMessage() function
  • Test: team e2e, scenarios, edge cases, integration tests
  • Verify: 1030 pass, 0 fail
  • Files created:
    • src/tool/team.ts - NEW - 9 team tool definitions (~650 lines)
  • Files changed:
    • src/tool/registry.ts - Register team tools behind flag
    • src/flag/flag.ts - OPENCODE_EXPERIMENTAL_AGENT_TEAMS dynamic getter
    • src/session/prompt.ts - teamMessage() function
    • src/server/routes/session.ts - POST /session/:sessionID/team-message
    • src/config/config.ts - max_turns, max_budget_usd experimental config
  • Complexity: High

14. Agent teams - UI ✅

  • Implement: dialog-team.tsx — team dialog with member list (status icons), shared tasks, Enter to navigate, m for message, l to go to lead
  • Implement: TeamBadge — compact: "Team: name (N active, N idle, N total)" or "memberName @teamName"
  • Implement: TeamStatusBar — per-member: status icon, model label, plan approval state, current task (truncated), task progress, delegate indicator
  • Implement: Shift+Up/Down teammate selection with "Messaging: @name" indicator bar
  • Implement: @teammate autocomplete with plan approval state and model in description
  • Implement: Prompt submit interception — when teammate selected, routes through POST /session/:sessionID/team-message
  • Implement: <leader>j/<leader>k teammate navigation, <leader>w team dialog, <leader>d delegate toggle
  • Implement: /team slash command, /delegate slash command
  • Implement: Team event toasts (member status, task claimed, cleanup)
  • Implement: SSE handlers for all team events + session.suggestion in sync.tsx
  • Test: 11 PTY smoke tests + 10 PTY live tests (all pass)
  • Verify: type check clean, all tests pass
  • Files created:
    • src/cli/cmd/tui/component/dialog-team.tsx - NEW - Team management dialog
    • test/tui/pty-harness.ts - NEW - Reusable PTY test harness (bun-pty)
    • test/tui/tui-smoke.ts - NEW - 11 PTY smoke tests
    • test/tui/tui-live.ts - NEW - 10 PTY live tests (real Anthropic OAuth)
  • Files changed:
    • src/cli/cmd/tui/app.tsx - Team commands, toasts, session.limit handler
    • src/cli/cmd/tui/context/sync.tsx - Team store type (with model field), SSE handlers
    • src/cli/cmd/tui/routes/session/header.tsx - TeamBadge, TeamStatusBar
    • src/cli/cmd/tui/routes/session/index.tsx - Teammate selection, messaging bar, keybinds
    • src/cli/cmd/tui/component/prompt/index.tsx - Tab-to-accept suggestions, team message routing
    • src/cli/cmd/tui/component/prompt/autocomplete.tsx - @teammate autocomplete with model info
    • src/config/config.ts - Keybinds: team_show, team_next, team_previous, team_delegate
  • Complexity: Medium-High

15. Session guards ✅

  • Implement: max_turns and max_budget_usd in experimental config
  • Implement: Toast notifications when limits are reached (via session.limit Bus event)
  • Implement: 80% budget warning (fires once per session via budgetWarned Set)
  • Implement: LimitEvent BusEvent definition in prompt.ts
  • Implement: TUI handler in app.tsx — "Budget Warning" / "Limit Reached" toasts
  • Verify: type check clean, 1030 tests pass
  • Files changed:
    • src/session/prompt.ts - LimitEvent, budgetWarned, guard logic with Bus.publish
    • src/cli/cmd/tui/app.tsx - session.limit SSE event handler
    • src/config/config.ts - max_turns (int), max_budget_usd (number) in experimental schema
  • Complexity: Low
  • Notes: Novel feature — Claude Code doesn't have per-session turn/budget limits. Our implementation: silent loop break + toast notification + server-side log.

Deferred / Out of scope

  • Chrome integration (separate product)
  • Sandboxing (design doc at opencode-sandbox-integration.md, pending implementation)
  • Reverse search (low value)
  • Status line (low value)

Session Log

Date Session Work done Issues
Feb 5, 2026 1 Comparison sheet created, tracker created Wasted time on wrong Go repo
Feb 5, 2026 1 Compact with instructions: implemented, tested (unit + integration with real LLM) Server endpoint tests need Instance cache handling
Feb 5, 2026 2 Prompt suggestions: implemented server-side generation, SSE plumbing, TUI display + accept Initial test showed empty suggestion (trivial prompt); fixed by using small model
Feb 5, 2026 3 Prompt suggestions: debugged integration test, switched to small model, verified e2e Flash model returns short suggestions for trivial convos (acceptable)
Feb 5, 2026 4 Background tasks: studied PR #9681, designed leaner approach, implemented full feature Ctrl+B is context-sensitive (only when session busy). stdin pipe broke cat test (fixed). Route ordering mattered for /foreground vs /:taskId.
Feb 5, 2026 5 Persistent memory: /memory command + .opencode/rules/*.md with path-scoping. Fixed compact test on wrong branch. Investigated 200K token limit (Claude Max OAuth restriction, not OpenCode). context-1m beta header not available on Claude Max subscription.
Feb 5, 2026 5 Task dependencies: depends_on field, blocked status, auto-resolve, circular dep detection. Fixed #5612 (enum validation) + #5934 (todos lost after compaction). Researched Claude Code's new Tasks system (addBlockedBy/addBlocks) and 7 related OpenCode issues.
Feb 5, 2026 6 Auto-memory: memory_save tool + todo system prompt injection. Studied OpenClaw's memory system. Replaced compaction prompt hack with simpler Todo.systemContext() in system prompt.