Relay

Pass the baton to the next Sutando session. Where session-handoff.sh writes structured facts (recent commits, open PRs, pending Qs), /relay writes the narrative continuity — what you were just working on, what should be checked first, what might break and how to detect it.

Usage: /relay (writes a new relay file) or /relay --append (appends to the most recent unprocessed relay file rather than creating a new one).

Why this exists

src/session-handoff.sh pulls together the structured state for the next session — system status, recent commits, open PRs, pending questions, tasks, conversation tail, quota, repo stats — and writes it to session-state.md. But "I was about to land PR #X and Mini's review said Y matters most" isn't captured by git log, gh pr list, or pending-questions.md tail. The next session reads structured facts but has to RE-INFER the continuity, which costs context and frequently misses the load-bearing decision.

The relay note encodes intent + judgment — the thing only the LLM that lived through the session can write.

Folder layout

Mirrors Sutando's existing tasks/ and results/ convention:

workspace/relay/
├── relay-{epoch_seconds}.md     # pending relay notes (drained into session-state.md, then archived)
└── processed/
    └── relay-{epoch_seconds}.md # already drained; kept for audit

• File naming: relay-{epoch}.md — sortable + greppable, matches the task-{epoch}.txt shape.
• Multiple files allowed: /relay always creates a NEW file by default. --append appends to the LATEST unprocessed relay-*.md instead of creating a new one.
• Consumption: src/session-handoff.sh reads ALL unprocessed relay-*.md files in sorted order, appends them to session-state.md under ## Relay Notes (from prior sessions), then mvs each one to processed/ — but only after confirming its ### <basename> header landed in the written file, so an interrupt cannot retire an uncaptured note.
• Cleanup: kept indefinitely on local disk. Tiny files (~200-500 bytes each); a year of relay notes is < 1 MB. Sync via the workspace-sync engine (scripts/sync-workspace.sh) for fleet visibility — the legacy sync-memory.sh flow is deprecated in v0.3.0 and removed in v0.4.0.

What to write

Write a narrative note (~150–300 words typical, no fixed schema) covering:

1. What I was just working on — the active PR / thread / decision in flight. Be specific. ("PR #1429 just merged; was about to start the relay skill PR.")
2. What to check FIRST in the next session — concrete first action. ("Pull origin; verify the relay skill SKILL.md lints clean; ping owner on quality-gate decision.")
3. What might go wrong + recovery — known failure modes + how to detect them. ("If session-state.md has no ## Relay Notes section, the relay/ dir is probably orphaned; check ls workspace/relay/.")
4. Implicit context — the why-behind-the-what, decisions that haven't been committed yet, things you'd want a colleague to know if they walked in cold.

Don't write things that are already in the structured snapshot (recent commits, open PRs, pending-questions tail). session-state.md already carries those. Relay's value is the things the structured snapshot can't reach.

Steps

1. Resolve WORKSPACE="$(bash scripts/sutando-config.sh workspace)".
2. mkdir -p "$WORKSPACE/relay" "$WORKSPACE/relay/processed" (idempotent).
3. If --append: find the latest unprocessed file via ls -t "$WORKSPACE/relay/"relay-*.md 2>/dev/null | head -1. If none, fall through to new-file mode. If found, append to it (with a --- separator + timestamp header).
4. Otherwise (new file): generate filename relay-$(date +%s).md under $WORKSPACE/relay/.
5. Write a narrative note (markdown formatting) as described above.
6. Save to the resolved path.
7. Print "Relay note written to <path>" with the absolute path. Mention briefly what's in it.

Quality bar

A good relay note is the difference between the next session starting at full speed vs. spending its first 5–10 minutes re-inferring state. Write something useful, not "I did stuff."

Bad: "Worked on some PRs. Ended session."

Good: "PR #1429 (import-UX) merged at 06:21Z; owner asked to start relay-skill PR next. Lucy's nit on stderr-parity landed pre-merge (commit 0ca0c89). Open thread: catchup PID-stamp variant — local edits applied to repo/skills/, NOT committed; owner is doing E2E test. If they greenlight, PR off staging-workspace-revamp with ~25-line diff in 2 SKILL.md files."

If the session was genuinely uneventful (read-only, no decisions, no in-flight work), say so explicitly: "No new work this session; previous relay note still valid." (Still write the file so the next session has a current heartbeat to read.)

Phase 1 scope

• Manual-invocation only. Auto-refresh (writing/updating from /proactive-loop) deferred to Phase 2 once we see how owners actually use the manual path.
• No quality-gate (refuse-on-thin-note). Always write whatever the LLM produces. Quality-gate deferred to Phase 2 pending observation.
• Read-side lives in src/session-handoff.sh — see its relay-drain block for the read-then-archive flow.

Phase 2 ideas (NOT in scope)

• Auto-refresh from /proactive-loop: if workspace/relay/relay-*.md is mtime-stale (> 30 min) AND substantive work has happened since, write a fresh relay note as part of the loop.
• Quality-gate: refuse to write if the LLM can't list any of active work / first action / known risk]. Could add --skip flag for genuine no-op sessions.
• Hook integration: PreCompact hook could prompt the LLM to write a relay note before compaction discards context.

Where it lives

workspace/relay/relay-{epoch}.md (pending) + workspace/relay/processed/relay-{epoch}.md (drained by session-handoff). Workspace-root parallel to build_log.md + pending-questions.md + tasks/. Owner-readable; both LLM and human can cat it cleanly.