Known issues & limitations

Release blocker blocks stable cut gate   Active confirmed bug or regression   Investigating repro unclear or fix in progress   Limitation by design or environment

Report something new: GitHub Issues — for /collaborate, say which phase stuck: planning, review, workspace gate, execution, or files. Each item below links to a tracked GitHub issue.

Release blockers

Gates that must pass before stable cut. See Q3 roadmap and stable release checklist.

• Release blocker Platform smoke pending sign-off #18

  Gate 5 platform smoke on a current beta installer (beta.22+) — minimum macOS arm64 + one of Windows x64 or Linux x64. Deferred 2026-07-20 (Mac-only soak day); still pending operator sign-off. Checklist: stable-platform-smoke.md.

Cleared for cut (tracker hygiene)

Former release blockers closed or downgraded to match climb / release-prep evidence (last refresh 2026-07-31).

• Cleared Collab-full layer gate #16

  LAYER=collab-full / collab-scenarios-all PASS on 2026-07-16, the 2026-07-19 climb (8/8), and overnight release-prep 2026-07-31 (Ollama roster). Residual edge flakes live under #20 / #21, not as a cut blocker. Evidence: layer-climb-2026-07-19-1809, test-everything-2026-07-31-0244.

• Cleared Parity soak confirmed #17

  test-parity-stable-restart PASS on climb 2026-07-19, parity-stable-restart-2026-07-20-0255, and again in release-prep-2026-07-31-0244 (3× restart, min 14/sweep).

• Deferred (post-cut) D5 specialist simplification #19

  Phase D backlog — not a v1.2.0 stable cut gate. See PHASE_D_BACKLOG.md.

Active bugs & investigations

Promoted from recent layer-gate collab-core/full runs (Jul 2026). These are beyond normal model variance — see model variance limitation below.

• Fixed layer-gate macOS --verbose make flag #22

  Harness now passes VERBOSE=1 as env (not --verbose to BSD make). Latest collab-core climb ran ~23m. Keep closed unless regression returns.

• Investigating Collab agents silent during planning discussions #20

  Agents blocked or silent — discussion timeout, shouldRespond blocked, msgs=0/3. Scenarios: document-findings-execution, collab-participation-three-agent, planning-two-agent, and others.

• Investigating Cloud provider generation_error loops in collab planning #21

  Claude, Gemini, and Assistant loop generation_error during planning — high error counts, partial agent participation, planning phase stuck.

Collaboration

Multi-agent collab is the hardest surface to make reliable. Beta.5 hardened planning parsers, discussion watchdogs, and approve-plan timing — but local model variance remains. See Collaboration guide and COLLABORATION.md.

• Limitation Chat does not write files #4

  Execution success means approved file-change proposals in Pending changes — not polished markdown in the thread. Chat needs [FILE_CHANGE] proposals you approve. TASK_STATUS: completed alone does not create files on disk.

• Limitation Model variance #5

  Local Ollama models differ in silence, speed, and plan quality. The hub enforces phase budgets and fallbacks; it cannot force a perfect task list from every model.

  Mitigation: Settings → Collaboration planning provider for optional cloud/larger-model planning turns. For implementation sessions: reliable tool model and optional reliable provider (repair round 2+) in Settings → Implementation sessions — the outcome card shows failure reason when repair fails.

• Limitation Smart routing scope #6

  Collaboration smart routing applies to execution tasks only — not normal channel chat. Planning can use a separate optional planning_provider_id in Settings.

Collaboration workarounds

• Upgrade the hub after each beta tag — collab fixes land frequently.
• Run make collab-preflight before live scenarios; use make debug-collab for discussion diagnosis (see COLLABORATION.md).
• Zero-task plans: revise the plan or re-run with fewer agents.
• Inspect per-turn routing trace on collab messages when execution picks an unexpected model.

Chat & context

• Limitation Bounded message history #7

  Per-channel history is capped (5000 messages) and age-pruned after 24h unless marked durable in channel info. SQLite sidecar + Export history — not a full in-app search archive. Use the in-channel find bar for recent history.

Integrations

• Limitation Slack needs local hub #8

  Socket Mode bridge runs in-process — the hub must be running on your machine. No public URL required. Use Settings → Integrations → Slack diagnostics when OAuth, Socket Mode, or channel routing fails.

• Limitation Confluence setup cost #9

  Doc agents need Cloud credentials, indexing time, and reasonable space size for good answers.

• Limitation Room chat needs same LAN #10

  The Room chat pack requires guests on the same LAN as the host hub. Corporate Wi‑Fi client isolation may block joins. Host must enable listen on LAN (listen_all) and configure a hub token. See ROOM_CHAT.md.

Desktop, IDE & packs

• Limitation Web UI is thin #11

  Browser chat at / is a lightweight client — no full workspace, command palette, routing trace drawer, or file-approval flows. Use the Tauri desktop for production work.

• Limitation Git requires dev pack #12

  In-app Git needs the Software development pack, git on PATH, and a git workspace.

• Limitation macOS ad-hoc signing #13

  GitHub Release macOS builds are ad-hoc signed — part of the $0 experiment (no Apple notarization fee). Free paths: Homebrew, Right-click → Open, or Install trust guide. Optional paid fix: notarization when budget allows (v1.2.1+).

Resolved in beta.2+: IDE v4 remote LSP relay and remote collab worktrees — see IDE v4 guide for current remote workspace limits.

Hub, agents & providers

• Limitation Single hub instance #14

  No distributed or multi-region deployment model yet.

• Limitation ReAct tools allowlist #15

  ReAct MCP on models without native tool calling is allowlist-based (gemma3:12b first in beta.5). Other tags use native tools when available, or Qwen swap fallback. See ReAct tools article.

How we test (and close items)

• Go tests — orchestrator, plan parsing: make test-collab-plan, make collab-smoke, go test ./...
• Live collab scenarios — make collab-scenario SCENARIO=…; matrix: collab-matrix.tsv (21/21 PASS as of 2026-06-09; recent layer-gate runs show regressions — see Active & investigating)
• Chat regression — make chat-scenarios-regression (workspace visibility, closure, echo)
• Implement parity — make implement-scenarios, make test-parity-stable (3× with hub restart)
• Release engineering loops — make layer-gate, make layer-fix-loop, make test-growth-loop with artifacts in docs/testing/

When a scenario flips to PASS or a bug is fixed: close the GitHub issue, delete the matching <li> here and the row in KNOWN_ISSUES.md, then note the fix in release notes.