Upgrade

• Every recommendation shows the technique itself (quoted or code-blocked from its source), names the exact LifeOS file or component it improves, and carries a Prior Status backed by file:line evidence gathered in this run.
• Nothing already implemented, deferred by decision, or rejected is recommended again: implemented items appear under Skipped Content with their evidence, and an idea rejected in MEMORY/KNOWLEDGE/REJECTED/ or the Upgrades store returns only with a named reason the context changed.
• No absence is reported that was not positively probed this run, and no adoption is recommended without a backward-compatibility check against current skills, hooks and workflows.
• A sweep report follows References/OutputFormat.md and arrives on its deadline, listing every source processed, skipped (with the reason) or timed out, so a missing lane lowers coverage without delaying or hiding it.
• A "what's new in AI" request returns one deduplicated, sourced digest saved under MEMORY/RESEARCH/, with any failed source named as unverified (Intents/NewsDigest.md).
• After a model run, no lane sits behind its vendor's latest same-class release unless the report lists it as a decision with its exact one-line fix.

Default intent: a bare "upgrade" or "check for upgrades" runs Upgrade (which includes reflection mining). "Check Anthropic" is the Upgrade sweep's tiered recommendations; "Anthropic updates" is NewsDigest's ANTHROPIC UPDATES digest. Both read the same Tools/Anthropic.ts diff.

Source labels in output: GitHub: claude-code vX.Y.Z · YouTube: Creator @ MM:SS · Docs: Section · Blog: Title.

Model moves: same class auto, cross class propose (principal directive, 2026-09-03). Shared by the Upgrade and Models intents. UpdateModels.ts --cross-vendor --sync writes a lane when the newer roster id is the same family at equal-or-higher stability (GA→GA, preview→GA) and a live probe with our key succeeds; ModelSync --daily runs the same command every day, so an Upgrade run is not the only path. What remains after --sync — a cross-class 🆕 (flash→pro), a GA→preview, a ❌ MISSING, or a ⛔ probe failure — is a decision with the exact one-line fix (edit CROSS_VENDOR in models.ts, or UpdateModels.ts --apply <tier> <id> for a Claude tier).

First-time setup (accounts, keys, files to fill in): Setup.md.

User customizations live at <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Upgrade/ (configRoot is $CLAUDE_CONFIG_DIR when set, else ~/.claude). When that directory exists, its PREFERENCES.md (§ News digest sets the story-explanation format), configurations, resources, youtube-channels.json and user-sources.json override the defaults here; when it is absent, the defaults stand. Monitoring state is install data and lives there too, under State/ (last-check.json, written by Tools/Anthropic.ts; youtube-videos.json and github-trending.json, the seen-lists, created on the first run that needs them). An optional History.md there holds the install's provenance behind generic gotchas and comments, read on demand.

• A reported absence is probed, never inferred (2026-08-06). Internal synthesis is held to the same standard as external findings: before an absence-claim earns a 🆕/CRITICAL tag, the thing claimed missing is grepped or read for in this run. A reported absence you did not check is a fabricated finding — the 2026-08-06 scan shipped a false CRITICAL across seven skills by grepping context: fork while never grepping background:.
• Hard deadline, fail-open — never block on a straggler. The deadline (~4 min per stage, enforced by workflows/UpgradeFanout.js; see the Upgrade intent's Tool contract) is set at dispatch; report with whatever is back when it hits. A missing source is listed as ⏳ timed out in Sources Processed; it degrades coverage, never delays the report. (2026-07-18: one hung GitHub-trending agent stalled a run ~1 hour.)
• Right-size the fan-out (~8 agents). Small-file reads collapse into one agent; network sources get short budgets and are first to drop. Over-fan-out is the failure the reflection corpus flags most.
• Budget the claude-code-guide freshness spawn per-spawn — it dies on a broad ask. One spawn asked to cover the whole Claude Code surface (hook events, settings.json, slash commands, SKILL/subagent frontmatter, MCP, SDK, API) fans out a batch of doc lookups on its first turn, the results blow its context window, and it surfaces as Prompt is too long ~10s after spawn. Prompt length is not the cause; how much the ask makes it fetch is. Cap at ~3 areas per spawn, tell it to look things up rather than pull whole pages, and split the surface across parallel spawns so losing one costs a slice. Tell: sibling agents in the same batch succeed while this one dies. Fallback with no agent: read Claude Code CHANGELOG from sources.json directly and diff the version range. (From a community pull request on the public LifeOS repo; reference in the USER extension's History.md.)
• GitHub search 422s on bare OR between qualifiers with no free-text term. Give every query a free-text term; don't retry a 422 inside the budget.
• Absence from settings.json ≠ a dead hook. It's generated by MergeSettings; hooks also fire via dispatchers (PreToolGuard) and Pulse HTTP. Flag "verify before concluding dark," never assert dead from absence.

User: "upgrade models — new Gemini model just came out; update Gemini, add a new model + agent, add to router"
→ Models intent
→ The Gemini lanes are on the new release after a live probe; the new model has a models.ts lane, an agent named for it and a Router lane; Router.ts status passes

User: "community scan: what have the best practitioners shipped recently? File the new techniques into the Upgrades queue"
→ Upgrade intent
→ A report in the canonical format, every 🔴/🟠/🟡 recommendation written to the Upgrades store, already-done items under Skipped Content with evidence

User: "check Anthropic only"
→ Upgrade intent, single source
→ An abbreviated report from that one source, every recommendation still carrying Prior Status with evidence