feat(home-feed): replace reflection producer with activity-log roll-up [JARVIS-512]

[alex-nork]

Summary

Fourth step of JARVIS-512, stacked on #25593 (instrumentation, open). Pivots the LLM-driven producer from "reflect on relationship state from nothing" → "consolidate recent action items into digests/threads." Same skeleton, same cadence mechanism, new input and new intent.

What changes

Rename: reflection-producer.ts → rollup-producer.ts (git rename tracked so blame history is preserved). The new producer:

• Reads recent action items from readHomeFeed() via a new defaultLoadRecentActions helper — sorted newest-first, capped at MAX_ACTIONS_IN_PROMPT = 30 for token budget. Non-action items are excluded (the input is the raw activity log, not existing consolidations).
• Still loads relationship state for context, but the prompt is explicit that it's secondary framing only — the model is instructed NOT to invent roll-ups from facts alone.
• Narrows the tool schema to type: digest | thread only. No nudges, no actions. Actions come from background jobs; nudges aren't this producer's concern.
• Adds a no_actions skip reason for when the activity log is empty. The producer short-circuits before the provider call.

Scheduler update (feed-scheduler.ts):

• Symbol rename: reflectionRunner/reflectionRan/REFLECTION_INTERVAL_MS/lastReflectionAt → rollup*.
• no_actions treated like no_provider: does NOT advance the cooldown gate, so the next tick retries as soon as new actions land instead of waiting the full 30-minute window. Real LLM attempts (success/empty_items/malformed_output/provider_error) still advance the gate to preserve backoff.

Cadence stays at 30 min in this PR. The drop to 120 min + on-visit refresh in home-feed-routes.ts is the next (and final) step in JARVIS-512.

Loop / cost audit

• The roll-up reads the feed but never writes actions — output is digest/thread only. No feedback path.
• no_actions short-circuit means an empty feed produces zero LLM calls even on every tick (the producer bails before provider.sendMessage).
• readHomeFeed() is the only new data dependency; it's a synchronous file read that the writer's merge loop already uses, so no new I/O cost.

Testing

• bun run typecheck — clean
• bun run lint — clean
• bun test src/home/ — 147/147 pass
• Migrated coverage: all previous reflection-producer tests carry over with digest/thread fixture values
• New tests: no_actions short-circuit, action-item serialization into the prompt, nudge/action types rejected at coercion, no_actions cooldown behavior in the scheduler

Remaining JARVIS-512 scope

• Step 5 — on-visit refresh in home-feed-routes.ts (debounced fire-and-forget from handleGetHomeFeed)
• Step 6 — drop rollup cadence from 30 min → 120 min (bundled with on-visit refresh so there's no staleness window)
• Deferred from earlier PRs: Gmail watcher instrumentation, task-runner direct, skill runner completion

🤖 Generated with Claude Code

---

[devin-ai-integration]

Devin Review found 1 potential issue.

View 4 additional findings in Devin Review.

[devin-ai-integration]

🟡 AGENTS.md violation: new comment references removed code

Lines 7–10 of the new module comment explicitly reference the removed reflection producer: "This is the replacement for the old 'reflect from nothing' producer". This violates the mandatory rule in assistant/AGENTS.md:19-21: "When writing or updating comments, do not reference code that has been removed. Comments should describe the current state of the codebase, not narrate its history." The comment should describe what the rollup producer does without referencing the old implementation.

Suggested change

	* small set of digests or threads." This is the replacement for the
	* old "reflect from nothing" producer: the roll-up starts from real
	* side effects instead of prompting the model to hallucinate signal
	* from relationship state alone.
	* small set of digests or threads." The roll-up starts from real
	* side effects (action items deposited by background jobs) rather
	* than asking the model to invent signal from relationship state
	* alone.

---

Was this helpful? React with 👍 or 👎 to provide feedback.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5f65fbea98

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Require source in roll-up items

The roll-up path depends on source-scoped replacement for digests, but source is optional here (required omits it), so model output without source is accepted and written. In mergeIncoming (feed writer), digest replacement only happens when incoming.source is present, so source-less digests/threads will accumulate as new rows on repeated runs instead of collapsing, which degrades feed quality and can grow the file unexpectedly.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Restrict roll-up input to fresh action events

This loader includes every persisted action item and only sorts by createdAt; it does not filter by freshness/status. Because newly-added schedule/sequence emits create non-expiring actions by default, actions.length will remain non-zero after initial activity, so the new no_actions fast-path almost never triggers and the model keeps reprocessing stale history on each cooldown cycle. That increases token spend and tends to regenerate repetitive summaries rather than reacting to truly new activity.

Useful? React with 👍 / 👎.