8000 agy/agent-e inbox writes silently don't reach recipient on un-symlinked pairs (agent-e↔agent-d, agent-e↔agent-f divergent) · Issue #40 · simon-inkie/inkie-the-brain · GitHub
[go: up one dir, main page]

Skip to content

agy/agent-e inbox writes silently don't reach recipient on un-symlinked pairs (agent-e↔agent-d, agent-e↔agent-f divergent) #40

Open

Description

@simon-inkie

Original Linear description

Filed 2026-06-29 by maintainer. Surfaced by agent-d: agent-e's !6 (INK-581) review completed, wrote its report to an artifact, updated its OWN thread copy, and fired the tmux knock (which reached agent-d) - BUT the verdict never landed in agent-d's inbox. agent-d had to read it off agent-e's pane + the artifact.

Root cause (maintainer diagnosis)

Inbox threads are meant to be ONE canonical file per pair, with one side a symlink to the other, so either agent's write hits the same file. Most of agent-e's pairs are correct:

  • agent-e↔agent-g, agent-e↔agent-h, agent-e↔agent-b, agent-e↔agent-a: agent-e side symlinks to the peer's canonical file. OK.
  • agent-e↔maintainer: maintainer side symlinks to agent-e's canonical file. OK.

But TWO pairs are DIVERGENT - both sides are independent real files (different inodes), never reconciled to the symlink convention:

  • agent-e↔agent-d: aggy/inbox/doctor2.md = real file 8372B (has the review); doctor2/inbox/aggy.md = real file 978B (just agent-d's outbound). agent-e's write landed only in its own copy.
  • agent-e↔agent-f: aggy/inbox/hopkins.md = real file 19865B; hopkins/inbox/aggy.md = real file 2709B. Same divergence.

So agent-e's writes to agent-d and agent-f go to agent-e's local copy and never propagate. Not data loss (agent-e's side has the full thread), but the recipient never sees it in their inbox = silent delivery failure for agy reviews to those two agents.

Fix (two parts)

  1. Reconcile the 2 divergent pairs (agent-d's lane - his tool, touches his + agent-f's canonical files): merge each pair losslessly into one canonical file + replace the other with a symlink, same as the maintainer-agent-c reconcile (doctor2/drafts/reconcile-inboxes.py). This fixes the current divergence AND future propagation (once agent-e's side is a symlink to the canonical, agent-e's writes resolve there).
  2. Systemic prevention (maintainer's lane, INK-567 agy-wake): ensure the agy write/wake path establishes the canonical-file symlink on first contact for a NEW pair, so a pair can never start life as two divergent real files. Audit aggy-gated.sh / the agy delivery path. Also add a cheap divergence check (any agent pair where both sides are real files = alert) so this is caught proactively, not by a teammate noticing a missing verdict.

Severity

agy is a merge-gate reviewer (agent-e approved !6). A review verdict silently not reaching the requester is a real comms-reliability bug, though the knock + artifact provide a backstop today. High.

Relates

  • INK-567 (agy-aware live wake - this is the delivery half of the same lane).
  • The earlier maintainer-agent-c / agent-k-agent-a thread-divergence reconcile (same class).

Origin

Migrated from Linear: INK-583. The original wording is retained except the disclosed privacy and rendering edits below.

Current scope and qualifications

The description above is the original Linear account, including its historical findings and scope. It is not a fresh claim that every described defect still exists.
Current Linear state is Backlog. Prior source review recorded partial, with this remaining scope:

  • Bind historic lossless recovery for both pairs and prove agy first-contact canonicalisation plus divergence alerts.

Publication transfers the task; it does not authorise live recovery, deployment, spending or configuration changes.

Transfer edits

  • Historical fleet identities are aliased; code identifiers and permitted source paths remain references.
  • Linear rich issue embeds are rendered as clickable Markdown links.
  • Additional transfer edits: fleet prose alias; fleet prose aliases; preserve literal evidence path; fleet prose aliases.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions

      0