Skip to content

attend: remaining work from ADR-113/114 implementation #2

@aaronsb

Description

@aaronsb

Context

Branch feat/adr-113-114-attend-active-awareness shipped the core attend implementation across 28 commits. The sensor loop, peer messaging, focus groups, context tracking, and ways integration are all working. Two Claude instances + a human terminal successfully collaborated through attend in a live session.

This issue tracks the remaining work that emerged during implementation.

Audit 2026-04-12: re-checked every item against the codebase. The vast majority of what was "remaining" has since shipped across PRs #1, #3, #4, #7, #10, #11, #14, #18. Only a handful of genuinely unstarted items remain — see "Still open" below.

Still open

Signal model / awareness

  • Build event batching — process sensor reports cargo/rustc start+exit as individual events. A higher-level "build started → completed (pass/fail)" observation would be cleaner for peer observers. Candidate: aggregate in the process sensor or as a derived sensor.
  • Consequence model / projection — context sensor does velocity + projection, but the general consequence model (ADR-113 consequence/ module) isn't built. Needed for non-context consequences like "branch is diverging from main at N commits/day."
  • Insistence tracker — unacted observations should re-surface with escalating urgency. Currently observations are fire-and-forget. ADR-113's insistence emitter + acknowledgment tracking is not started. (ADR-119 shipped the action-potential model that replaces part of this — revisit scope before starting.)
  • Emission modes — ADR-114 defines silent/informational/affordance/insistent/critical. Currently attend has magnitude-based priority but no formal mode system. (ADR-119 may subsume this; check.)

Operational

  • Signal TTL and cleanup subcommand — signals older than 5 minutes are cleaned up by the peer sensor during polls, but there's no periodic cleanup for dirs nobody is polling. An attend cleanup subcommand or cron-based cleanup would help.
  • Way vocabulary tuning — attend way scores 0.44 on "attend focus group" but only 0.31 on "send signal to another claude". More test prompts and threshold tuning via /ways-tests would improve matching.

Shipped (formerly "remaining")

ADR-113

ADR-114

  • trigger.type: attend in ways schema
  • Affordance strings in notifications
  • Linter attend-awareness
  • Signal handler ways: context-pressure, build-complete, reflection-overdue — all exist under hooks/ways/meta/attend/

Infrastructure

ADR-118 follow-ups (from PR #14 review)

References

  • ADR-113 Accepted · ADR-114 Accepted · ADR-115 Draft (code shipped) · ADR-116 Accepted · ADR-117 Accepted · ADR-118 Accepted · ADR-119 Accepted · ADR-120 Draft

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions