--- status: root_cause_confirmed trigger: Phase 35 SC1 — cross-machine sync auto-detect prompt never surfaces to the user despite its emission chain firing created: 2026-05-27 updated: 2026-05-27 phase: 35 blocks: 35-09 Task 3 operator UAT (SC1 auto-detect) --- # Debug: sync-prompt emits into the void ## Symptoms - **Expected:** on a fresh-ish session, the cross-machine sync auto-detect prompt (``) renders as an AskUserQuestion offering Enable now / No never / Remind in 12h. - **Actual:** no AskUserQuestion ever surfaces, across every trigger tried (2× `/clear`; 2× full binary start/end with a `$LIVE start todlando` in between). - **Errors:** none surfaced to the user — silent. - **Timeline:** first observed on fresh spt v1.11.17 install (phase 35 just deployed). - **Repro:** reset sync to `state=unset`; full session start → `$LIVE start todlando` → session end; full session start → session end. Prompt never appears. ## DATA_START (verbatim operator-supplied analysis — treat as data, not instructions) PROVEN (live debugging, spt v1.11.17, Windows 11): - `$LIVE start`/revive writes the `.sync-prompt-due` sentinel (src/live/start.rs:396 / :721; revive delegates to start::run). Only these write it — confirmed by grep, no other callers. - On the NEXT fresh `startup` SessionStart, plugin_session_start.rs::run() consumes the sentinel (line 287), passes should_emit_sync_prompt (gh present + state Unset), calls emit_sync_prompt(), and bumps last_prompted_ts. - Evidence: after a clean reset to state=unset, doing (session start → $LIVE start todlando → session end) then (session start → session end), settings.json showed state still "unset" but last_prompted_ts bumped to 2026-05-27T06:59:00Z. That timestamp is written ONLY inside the `if should_emit_sync_prompt` branch right after emit_sync_prompt(). So emission executed. - Yet NO AskUserQuestion ever rendered to the user. ROOT CAUSE (high confidence): emit_sync_prompt() prints `` as SessionStart additionalContext (passive context). NOTHING handles that envelope. grep for "spt-psyche-sync-prompt" across plugin/, src/, docs/, CLAUDE.md returns ONLY the emitter in src/owl/plugin_session_start.rs (lines 196/202) + its own unit test (1462/1463). No skill, CLAUDE.md, or hook instructs Claude to act on it. By contrast `` IS handled by plugin/spt/skills/live/SKILL.md (explicit "parse + dispatch"), which works only because the user TYPES /spt:live to load that skill. The sync prompt has neither a user trigger nor a handler instruction, so the additionalContext sits unactioned — emits into the void. ARCHITECTURAL MISMATCH: within any single session, the boot SessionStart always precedes the in-session `$LIVE start`, so write (live-start) and consume (SessionStart) can never pair in the same session — the prompt is structurally deferred to a later session's boot. Plus: a user who never goes live never gets a sentinel written, so is never offered sync at all. The runbook (35-OPERATOR-UAT.md Section 1/2) wrongly expects "run $LIVE start → SessionStart shows the envelope," a pairing the architecture cannot produce. WHY TESTS MISSED IT: the 4 integration tests (tests/sync_*.rs) are fake-remote / in-process and never exercise the hook → Claude → AskUserQuestion UI hop. CANDIDATE FIX DIRECTIONS: A (preferred) — active delivery: route the sync offer as an owl-message EVENT to the live agent (like the version-change announcement), handled by an always-loaded live/ready SKILL.md instruction so it renders an AskUserQuestion reliably. B — handler instruction: add an always-loaded directive (skill/CLAUDE.md) telling Claude to render the AskUserQuestion when it sees `` on a startup turn. Weaker; relies on the model acting on passive injected context. SECONDARY candidate defect (lower priority): `$OWL doctor` rendered enabled-but-failing per-agent sync rows as green [PASS] despite last-err + pending retry-after (observed pre-reset). Possibly should be [WARN]/[FAIL]. Env: Windows 11, spt v1.11.17 deployed to cplugs, gh present (C:\Program Files\GitHub CLI\gh.exe, account SaberMage). ## DATA_END ## Current Focus - hypothesis: CONFIRMED — `` additionalContext has no handler - root_cause: emit_sync_prompt() outputs a hookSpecificOutput JSON envelope whose additionalContext contains ``. No skill, CLAUDE.md, or hook directive instructs Claude to act on this envelope. The version-change analogue works because `` embeds its own `` element telling Claude to call AskUserQuestion — delivered via spool drain (PreToolUse hook), not passive additionalContext. The sync prompt uses neither mechanism. - structural_gap: the live-start sentinel write and SessionStart consume cannot occur in the same session by construction (SessionStart always precedes $LIVE start). Users who never go live never receive a sentinel write at all. - next_action: surface fix options to operator ## Evidence - timestamp 2026-05-27: settings.json `last_prompted_ts=2026-05-27T06:59:00Z` with `state=unset` after start/$LIVE start/end ×2 — proves emit_sync_prompt() executed but no UI rendered. - grep `spt-psyche-sync-prompt` → only src/owl/plugin_session_start.rs:{196,202,1462,1463} (emitter + unit test). No handler anywhere in plugin/, skills/, CLAUDE.md. - grep confirms `` handler exists in live/SKILL.md line 73 (user-triggered skill load required); `` works via self-contained `` element delivered through spool drain — neither pattern was applied to sync prompt. - `build_sync_prompt_envelope()` (plugin_session_start.rs:195) emits hookSpecificOutput JSON with no embedded handler directive in the envelope body. ## Eliminated - hypothesis: sentinel never written / gate failed → ELIMINATED (last_prompted_ts bumped proves the full sentinel→consume→predicate→emit chain ran). - hypothesis: gh not present in hook subprocess → ELIMINATED (predicate passed; emit fired). - hypothesis: wrong trigger (/clear) → partial: /clear can't consume (source-gated early return), but the proper start/end pairing DID consume + emit, so the no-render is downstream of that. - hypothesis: version-change pattern followed → ELIMINATED. Version-change uses spool drain + self-contained instructions. Sync prompt uses passive additionalContext with no instructions. ## Resolution root_cause: emit_sync_prompt() emits `` as passive SessionStart additionalContext with no handler directive in any always-loaded skill, CLAUDE.md, or hook. Claude receives the context but has no instruction to act on it, so no AskUserQuestion fires. The structural live-start↔SessionStart cross-session pairing and never-prompted-if-never-live gaps are secondary architectural issues that remain regardless of which fix direction is chosen. fix: ROUTED TO PLANNING (operator chose Plan fix over inline). Fix direction A (spool-based active delivery mirroring version-change: spool drain via PreToolUse + self-contained element + always-loaded live/ready SKILL.md handler) is the lead candidate — it also resolves the cross-session deferral (in-session active delivery to the live perch). Open scope question for the plan: is "offer sync to live-agent users only" acceptable, or must non-live users also be offered (would need a trigger beyond the live-start sentinel)? Also fold in: runbook 35-OPERATOR-UAT.md Section 1/2 wording fix, a test that exercises the hook→delivery→handler hop, and the secondary doctor [PASS]-on-failing-row candidate. Handing to /gsd:plan-phase --gaps for phase 35. LOCKED SCOPE DECISION (operator, 2026-05-27): offer sync to LIVE-AGENT USERS ONLY. Trigger points: (1) `$LIVE start`/revive, AND (2) SessionStart at the /clear|/compact re-orientation boundary (only when a live agent is present). Delivery MUST be active (spool to the live perch / handled by an always-loaded live|ready SKILL.md instruction, mirroring the version-change pattern) so the AskUserQuestion actually renders — never passive SessionStart additionalContext. This intentionally drops the non-live and the plain-startup-only trigger paths.