Skip to content

Project backlog

Canonical home for concise future-work requests and unresolved defects. Detailed plans, investigations, and review reports belong under .scratch/. Read this before planning work in any listed area.

Use the inbox for ideas that have not been investigated yet. Promote an item to a detailed entry when its scope and current status are understood. Remove items when they are implemented, obsolete, or no longer valuable.

Inbox

  • Consider making mevedel's data buffers hidden
  • Check if memory and journal processing could be async

Execution

Harness-enforced Bash timeouts

Managed Bash commands run until they exit or are stopped. 145c66a removed mevedel-bash-timeout (120 s default) and the model-facing timeout_seconds, so only a native timeout the model writes itself bounds a command. A forgotten command keeps running and holds the system sleep inhibitor. When Bash completion wake-ups and the Monitor tool were added (2026-10), the user decided against restoring a Bash deadline for now; only Monitor expires (timeout_ms). Revisit if forgotten commands become a problem in practice. Reference point: Claude Code caps background commands at 10 minutes by default and 2 hours at most. Expiry would reuse the existing TERM/KILL process-group teardown and report a timeout outcome.

Whiteboard

Excalidraw parity left out of the element model

ADR 0121 stores and draws Excalidraw elements but cherry-picks interaction. Not yet implemented: segment dragging for elbow arrows, object and grid snapping, a frame tool and moving children with their frame, sticky-note font auto-fit with its lifted corner and date footer, the constant-width laser-pointer stroke, and embedding the scene in PNG/SVG downloads. Add each when a board needs it; Excalidraw's behavior is compiled in excali-mode's docs/excalidraw-spec.md (https://github.com/yibie/excali-mode).

Review

Automatic turn advisor

Consider an opt-in second-model review of successful root turns, with bounded feedback delivered to the next request. Reuse the Stop hook, reviewer runtime, and reminder delivery. Keep findings distinct from Buddy's advisory notes, and specify evidence scope, deduplication, cooldown, cancellation, and cost before implementation. Pursue when recurring review failures justify the extra model call; keep it advisory, without restarting completed turns. Implement this bounded feature before considering generic prompt or agent hook handlers. There is no automatic root-turn reviewer/instruction-injection path today. The original request came from the 2026-08-20 watchdog/advisor discussion.

Hook extensions

Task transition events

Consider a task transition event when notifications or tracker synchronization need to observe every completion. A hook matching TaskWrite misses automatic completion by mevedel-tool-task-finalize-owner in mevedel-tool-task.el. Observe actual state transitions across both paths rather than tool calls alone. Defer until a concrete integration needs that coverage.

Deferred extensions

  • HTTP, prompt, MCP-tool, and agent handlers: revisit when a named workflow cannot be served adequately by command/Elisp handlers. Define permission, cancellation, and recursive-hook behavior before adding a handler type.
  • Handler-level predicates: keep conditions inside handlers unless repeated workflows or measured launch cost justify declarative filtering. Prefer an event-specific matcher where sufficient, such as skill-name matching for UserPromptExpansion, rather than a general predicate language.
  • Configuration, cwd/files, tool-batch, and shell-environment events: add individually when a caller needs a boundary existing events cannot express; define the event's scope and control effects before implementation.