Skip to content

progress.md

Latest commit

 

History

History
292 lines (238 loc) · 15.2 KB

progress.md

File metadata and controls

292 lines (238 loc) · 15.2 KB
Raw

Progress Log

Session: 2026-03-02 - Re-scoped planning for PR #50, #49, #48

Status

  • Exact PR review complete.
  • Planning files updated to correct scope.
  • No code implementation performed yet.

Actions Completed

  1. Replaced initial branch-based assumption after user clarified target PRs (50, 49, 48).
  2. Fetched and reviewed:
    • PR metadata (draft/open/base/head/additions/files)
    • Changed file lists
    • Patch content for all three PRs
  3. Compared PR scope against current local main to confirm changes are not yet merged.
  4. Identified cross-PR dependency and merge-order risks (especially PR #50 depending on PR #49 behavior).
  5. Rewrote task_plan.md and findings.md for exact PR scope.

Key Outcomes

  • PR #50 is not self-contained against main in its current form.
  • PR #49 and PR #50 currently define overlapping tunnel strategies.
  • PR #48 is suitable as a docs-first strategic artifact, pending issue/owner extraction.

Next Actions (Execution)

  1. Decide canonical tunnel approach (profile vs override) and enforce in PR #49/#50.
  2. Resolve PR #50 base/dependency issue (self-contained or stacked rebase).
  3. Add Windows-compatible sharing path for tunnel helper tooling.
  4. Finalize merge sequencing and acceptance checks.

Files Updated This Session

  • task_plan.md
  • findings.md
  • progress.md

Environment Notes

  • gh CLI unauthenticated in this environment; GitHub API and patch endpoints used instead.

Session: 2026-03-06 - Cursor parity debugging

Status

  • Cursor code path inspected.
  • Local Cursor CLI validated on this Windows machine.
  • Implementation bugs identified; patching started.

Actions Completed

  1. Compared Cursor registration/invocation code against Claude/Codex/OpenCode.
  2. Ran existing server/tests/cursor.test.ts, server/tests/mcp.test.ts, and bun run typecheck to establish baseline.
  3. Validated local Cursor install:
    • cursor.cmd present
    • standalone agent.cmd present
    • agent.ps1 is blocked by PowerShell execution policy
  4. Confirmed via agent.cmd --help that TalkTo should be using a resumable chat ID for --resume.
  5. Confirmed current prompt/docs guidance around CURSOR_TRACE_ID is inconsistent with actual CLI semantics.

Next Actions

  1. Patch server/src/sdk/cursor.ts discovery/auth handling.
  2. Update Cursor registration guidance in prompts and MCP tool descriptions.
  3. Add targeted tests for Cursor registration semantics.

Completed

  1. Patched server/src/sdk/cursor.ts for robust CLI discovery, Windows fallbacks, and auth/error handling.
  2. Updated Cursor registration guidance in README, MCP tool descriptions, and prompt templates.
  3. Added MCP test coverage for Cursor registration guidance.
  4. Posted the integration finding to TalkTo channel #project-talkto.

Validation

  • bun test server/tests/cursor.test.ts ✅
  • bun test server/tests/mcp.test.ts ✅
  • bun test server/tests/prompt-engine.test.ts ✅
  • bun run typecheck ✅

Environment Notes

  • Running bun test server/tests/cursor.test.ts server/tests/mcp.test.ts in one combined command hit EADDRINUSE because the server test boot path expects isolation. Running the suites separately passed.

Session: 2026-03-07 - Workspace sharing / lifecycle stability

Status

  • New multi-feature pass started.
  • Planning refreshed from prior work and current dirty tree.
  • Architecture/lifecycle audit underway.

Requested Scope

  1. Simplify workspace sharing and critique the current cross-machine collaboration architecture.
  2. Add delete operations for channels and agents.
  3. Allow editing agent profiles and stop relying solely on agents to self-report provider type.
  4. Ship adjacent stability improvements where justified.

Next Actions

  1. Inspect workspace creation/share/join/auth code paths and current UI/API shape.
  2. Inspect channel/agent lifecycle routes, services, and frontend affordances.
  3. Inspect current provider detection logic across MCP registration and setup flows.

Completed

  1. Added TALKTO_PUBLIC_BASE_URL support in server config so share links and advertised MCP URLs can target a real public origin.
  2. Added admin PATCH /api/agents/:agentName, DELETE /api/agents/:agentName, and DELETE /api/channels/:channelId.
  3. Implemented agent/channel admin management inside the workspace settings sheet.
  4. Hardened provider detection so OpenCode evidence overrides a bad claude_code self-report, while explicit Codex/Cursor still short-circuit correctly.
  5. Added live invalidation events for agent/channel lifecycle changes and fixed Cursor invocability labeling in the sidebar.
  6. Added targeted API + MCP coverage and revalidated frontend build/tests.

Validation

  • bun run typecheck ✅
  • bun test server/tests/mcp.test.ts ✅
  • bun test server/tests/api.test.ts ✅
  • cd frontend && bun run test ✅
  • cd frontend && bun run build ✅

Environment Notes

  • Frontend vite/vitest commands hit spawn EPERM inside the sandbox because esbuild could not start worker processes there. Re-running those commands outside the sandbox passed without code changes.

Session: 2026-03-07 - Context menus and test DB isolation

Status

  • Right-click operator menus shipped for project channels and agents.
  • Integration tests now use temp databases and skip live server binding.
  • Validation complete.

Requested Scope

  1. Add shadcn context menus for projects and agents so delete and related actions are available on right click.
  2. Ensure tests use a separate database so they do not populate the real product.
  3. Investigate adjacent product improvements such as image paste/forward support.

Completed

  1. Added a shared Radix/shadcn-style context-menu primitive at frontend/src/components/ui/context-menu.tsx.
  2. Added project-channel context menus with open/copy/delete actions in frontend/src/components/workspace/channel-list.tsx.
  3. Added agent context menus with DM/copy/delete actions in frontend/src/components/workspace/agent-list.tsx.
  4. Added frontend API coverage for channel deletion and agent update/delete requests.
  5. Added server/tests/test-env.ts so integration suites always use a temp TALKTO_DATA_DIR.
  6. Added TALKTO_DISABLE_SERVER=1 test bootstrap handling in server/src/index.ts so test imports do not bind a real port.
  7. Fixed server/tests/api.test.ts to create its own onboarded admin user inside the isolated test DB.

Validation

  • bun run typecheck ✅
  • bun test server/tests/api.test.ts server/tests/auth.test.ts server/tests/mcp.test.ts server/tests/messages-write.test.ts server/tests/ownership.test.ts server/tests/where-chaining.test.ts ✅
  • cd frontend && bun run test ✅
  • cd frontend && bun run build ✅

Environment Notes

  • Frontend bun run test and bun run build still require execution outside the sandbox because Vite/esbuild worker spawning fails with spawn EPERM inside the sandbox.
  • The prior EADDRINUSE issue on combined server integration runs was eliminated by disabling live server bootstrap under tests.

Session: 2026-03-07 - Claude integration assessment

Status

  • Investigation started.
  • Failure evidence captured from the provided log artifact.
  • Local Claude executable presence confirmed.

Actions Completed

  1. Loaded current project state and prior investigation notes.
  2. Extracted Log talkto.docx by unpacking the .docx archive and reading word/document.xml.
  3. Confirmed the key failure in the supplied log is Claude Code process exited with code 1 from @anthropic-ai/claude-agent-sdk/sdk.mjs.
  4. Confirmed local Claude CLI presence via where.exe claude and PowerShell command resolution.
  5. Compared TalkTo's wrapper with the installed SDK surface and found a likely contract mismatch: permissionMode: \"bypassPermissions\" is used without the required allowDangerouslySkipPermissions: true.
  6. Confirmed that claude auth status cannot be trusted inside the sandbox because the CLI hits uv_spawn 'reg' EPERM before reporting auth state.
  7. Re-ran claude auth status outside the sandbox and confirmed the local install reports as logged in.
  8. Ran a minimal unrestricted claude -p command and found a separate environment issue: the stored OAuth token is expired and headless calls currently fail with 401 authentication_error.

Next Actions

  1. Validate the suspected permission-flag mismatch against the current SDK typings and CLI behavior.
  2. Inspect the installed SDK/CLI source to confirm whether TalkTo's missing allowDangerouslySkipPermissions flag is a hard validation requirement.
  3. Determine whether the break is code-level, environment/setup-level, or both.

Completed

  1. Patched server/src/sdk/claude.ts to include the required allowDangerouslySkipPermissions: true flag and to clear Claude liveness on prompt failure.
  2. Updated prompts/claude_global_rules.md, prompts/registration_rules.md, and server/src/mcp/server.ts to stop advertising PID fallback for Claude session registration.
  3. Added MCP-side rejection for numeric Claude session IDs.
  4. Added regression coverage in server/tests/claude.test.ts and server/tests/mcp.test.ts.

Validation

  • bun test server/tests/claude.test.ts ✅
  • bun test server/tests/mcp.test.ts ✅
  • bun run typecheck ✅

Follow-up Completed

  1. Patched server/src/services/agent-invoker.ts so Claude failures surface a clearer operator message and immediately ghost the failed agent.
  2. Patched server/src/services/agent-discovery.ts so stale-credential cleanup also ends active sessions and broadcasts agent_status=offline.
  3. Added server/tests/agent-discovery.test.ts for the offline/credential-clearing path.
  4. Re-validated Claude CLI after re-login; a direct claude -p call now succeeds.

Follow-up Validation

  • bun test server/tests/agent-discovery.test.ts ✅
  • bun test server/tests/claude.test.ts ✅
  • bun test server/tests/mcp.test.ts ✅
  • bun run typecheck ✅

Session: 2026-03-08 - Registration verification / lifecycle simplification

Status

  • New lifecycle refactor started.
  • Existing planning files synchronized.
  • Code-path audit and Claude-specific root-cause work underway.

Actions Completed

  1. Confirmed the current subprocess lifecycle model is based on in-memory liveness Sets, not durable provider-backed verification.
  2. Confirmed OpenCode is the only provider with a direct session health API today.
  3. Verified Claude session recovery from local transcript files under .claude/projects and tightened prompt guidance to match cwd before reading sessionId.
  4. Isolated the recent Claude invocation failure to TalkTo resuming from the wrong working directory.
  5. Patched Claude SDK calls to pass projectPath as cwd and validated that change with targeted tests plus a live external resume probe returning OK.

Next Actions

  1. Finish auditing every registration/discovery/UI path that still depends on subprocess ghost probing.
  2. Move subprocess verification to register().
  3. Remove proactive subprocess liveness checks while keeping failure-time invalidation.
  4. Revalidate backend behavior and update operator-facing status semantics if needed.

Session: 2026-03-08 - Startup session reconciliation / active-agent model

Status

  • Follow-up architecture pass started.
  • Planning files refreshed for a startup reconciliation refactor.
  • Provider health primitive audit underway.

Actions Completed

  1. Confirmed the remaining stale-agent problem is not just frontend rendering: TalkTo still lacks a startup reconciliation pass for persisted subprocess credentials.
  2. Verified that the frontend loads /api/agents immediately and therefore faithfully shows any stale rows left in the database.
  3. Confirmed OpenCode already has a cheap session health API, while Claude/Codex/Cursor wrappers currently do not.
  4. Reframed the product model around "active = persisted credentials that pass cheap provider verification", not "active terminal" or "open MCP connection".

Next Actions

  1. Probe local Claude/Codex/Cursor storage and CLI surfaces for non-interactive session verification options.
  2. Implement a reconciliation pass that runs automatically at startup and/or first app load.
  3. Auto-delete stale non-system agent rows and align the frontend to the new active-agent model.

Completed

  1. Added cheap provider-backed verification helpers for Claude, Codex, and Cursor using local session/transcript metadata instead of real prompts.
  2. Added a backend reconciliation pass and wired it into startup bootstrapping.
  3. Added app-load reconciliation before the workspace renders, so stale agent rows are removed before the UI fetches its steady-state agent list.
  4. Added GET /api/agents?reconcile=1 and upgraded bulk cleanup to use the same provider-backed reconcile path.
  5. Updated the DM header to derive agent reachability from is_invocable instead of transient websocket status noise.
  6. Validated the new behavior with targeted server tests and a clean bun run typecheck.

Progress Log

Session: 2026-03-08 - Channel session history API

Status

  • Backend implementation complete.
  • Focused validation complete across REST, MCP, and write-path test coverage.
  • Final handoff summary pending.

Actions Completed

  1. Loaded the brainstorming and planning-with-files skill instructions.
  2. Ran session catchup for planning-with-files.
  3. Inspected existing planning files and decided to preserve prior task context by prepending a new task section.
  4. Recorded the new goal, phases, open questions, and initial findings.
  5. Audited the schema, message routes, invocation path, and every message producer (messages, message-router, users, webhooks, seed data).
  6. Cross-checked the local provider CLIs (claude, codex.cmd, opencode, cursor-agent) while evaluating how TalkTo uses session identifiers.
  7. Implemented persisted channel sessions in the DB plus backfill/migration logic for existing messages.
  8. Added session-aware message creation/invocation behavior and two new session history APIs.
  9. Updated the isolated server test schema and added API coverage for session grouping and root-message deletion resilience.
  10. Fixed an old-database startup regression by deferring channel_session_id index creation until after additive migrations have added the column.

Next Actions

  1. Deliver the API shape and implementation summary to the user.
  2. Optionally wire the new session-history endpoints into the frontend once the desired UI is defined.

Files Updated This Session

  • task_plan.md
  • findings.md
  • progress.md
  • server/src/db/index.ts
  • server/src/db/schema.ts
  • server/src/db/seed.ts
  • server/src/routes/messages.ts
  • server/src/routes/users.ts
  • server/src/routes/webhooks.ts
  • server/src/services/agent-invoker.ts
  • server/src/services/broadcaster.ts
  • server/src/services/channel-sessions.ts
  • server/src/services/message-router.ts
  • server/src/types/index.ts
  • server/tests/api.test.ts
  • server/tests/setup.ts

Validation

  • bun test server/tests/api.test.ts ✅
  • bun test server/tests/messages-write.test.ts ✅
  • bun test server/tests/webhooks.test.ts ✅
  • bun test server/tests/mcp.test.ts ✅
  • bun run typecheck ✅
  • bun run dev:server against the existing local data/talkto.db ✅
  • bun test server/tests/api.test.ts server/tests/messages-write.test.ts server/tests/webhooks.test.ts server/tests/mcp.test.ts ✅