Mission Control day summary — 2026-05-28

Plain-English summary of what happened today.

Big picture

Mission Control started the day looking stuck: lots of tickets were in the pending bucket, but most of them were not actually waiting for Elmar. The real problems were a mix of unclear board labels, stale runtime state, review-gate noise, and technical tickets being misclassified as human work.

By the end of the day we had turned that into a cleaner control plane:

• the main Mission Control repair chain is done;
• completed tickets no longer keep scary stale retry/death badges;
• the board now exposes clearer “why is this pending?” reasons;
• the shepherd checks MC every 5 minutes;
• a separate pulse gives Elmar visibility without spamming;
• the root causes of today’s “pending but not moving” noise were fixed and landed.

What we fixed

1. Made MC stop stranding completed workers

We fixed the harvest / signal-write race that could leave tickets in needs_input or pending-looking states after the worker had actually completed. This was one of the core reasons the board looked stuck.

Related ticket: MC-4357.

2. Stopped shadow-review from bouncing valid controller closures

Shadow review was falsely failing some tickets because it inspected the wrong branch/repo context or treated non-code/operator work like code work. That caused already-closed controller tickets to reappear as pending.

Related ticket: MC-4371.

3. Improved the MC Home / mobile experience

A lot of today’s work was around the MC Home and board/workbench UX:

• AionUI-style Home polish.
• Visible active sessions / chat entry points.
• Better mobile session information architecture.
• Better campaign cockpit / ticket creation flow.
• Pending cards now show a clearer reason/action instead of burying the reason in comments.

Related tickets: MC-4370, MC-4324, MC-4326, MC-4376, MC-4348.

4. Made review gates stricter where they need to be

For content/audio/research work, we tightened the rule that “tests passed” is not enough. If the acceptance criterion is a user-facing artifact, the reviewer must verify the actual artifact or explain why it cannot be verified.

Related ticket: MC-4383.

5. Fixed stale red retry/death badges on done tickets

Some tickets that were already Done still showed old n/5-style failure badges. Those were stale counters, not active failures. We normalized terminal tickets so Done rows do not keep showing current-pressure warnings.

6. Added autonomous MC shepherding

We set up the MC board shepherd:

• runs every 5 minutes;
• inspects DB state, runtimes, tmux, scheduler/pickup logs, and service health;
• gates reviews, repairs row drift, clears stale terminal pressure, parks true backlog, and requeues runnable technical work;
• writes a shared pulse-state file.

Then we added a separate 30-minute progress pulse so Elmar gets visibility without noise.

7. Found and fixed why “pending” still looked stuck

After the first cleanup, the board still looked like it had unmoving tickets. The live audit showed:

• MC-3921 was intentionally parked backlog / upstream Telegram topic work.
• MC-4323 was intentionally parked low-value PKA/vector-recall follow-up.
• MC-4288 was waiting on Iris.
• MC-4193 and MC-4194 were Life document review items assigned to Elmar.
• MC-4333 was incorrectly assigned to Elmar even though it was technical operator-tune work.

We reassigned MC-4333 to Luci and got it moving.

8. Fixed the operator tuner crash

MC-4333 found the immediate bug: the operator tuner embedded a huge prompt in a command-line argument, which hit Linux ARG_MAX and failed with Argument list too long.

Fix: pipe the prompt through stdin instead of argv.

Related ticket: MC-4333.

9. Fixed the deeper operator-noise root causes

MC-4397 then fixed the root causes that made MC look messy even after the immediate crash was fixed:

• graphify-out/ generated files are now classified as generated noise by the operator.
• Active-lane stale threshold was raised from 2 hours to 4 hours so legitimate longer-running lanes are not escalated too aggressively.
• repeated scheduler failure logging now dedupes by task_run id so the same failure is not counted repeatedly across overlapping operator runs.

Related ticket: MC-4397.

What is still open intentionally

These are not current MC-core blockers:

• MC-3921 — parked backlog: upstream/postinstall Telegram topic routing hardening.
• MC-4323 — parked backlog: low-value PKA/vector recall follow-up waiting on external Mac/PKA context.
• MC-4288 — waiting on Iris handoff.
• MC-4193 and MC-4194 — Life document review items assigned to Elmar.

Git / landing status

Mission Control fixes were merged into the active master branch and pushed.

Workspace-level operator-tuner fix was already on workspace master and pushed.

This summary is intended to make the day readable without needing to inspect every MC ticket or commit.