Keystone — What it can't do (yet)

Answering & chat

• No live web fetch by default. Answers draw on the model's training knowledge; "latest / today / current price" questions can be stale unless web-grounding supplies fresh sources. [observed]
• Media only with known-real URLs. It renders inline images and YouTube embeds, but won't invent a media URL — so for things it isn't sure of, it links the source page instead of embedding. [design]
• Mislabeled/misrouted questions. The intent classifier used substring matching, so pr matched inside "comprehensive" and ui inside "build" — tagging ordinary questions as the "code / GitHub route". [fixed] — now whole-word matching (convergance-os/model-router.js).
• No image/file input in the answer path. The composer's attachment is for journaling, not "analyze this screenshot and answer." [observed — verify]
• No automatic cross-session recall in chat. Within a session, context carries (#772); across sessions it does not auto-recall — that lives in the journal / CSF and is manual. [observed]

Autonomy & the convergence loop

• The loop doesn't run end-to-end in the live serving path. Chat emits Convergence Records but nothing grades most of them, so Keystone does not yet learn from ordinary conversations — only the Kalshi trade slice closes Reason→Verify→Converge. [agent-spine note]
• Governors aren't gating live actions. The grounding throttle and the Σ₀ surprise canary exist but aren't yet wired onto every action. [agent-spine §6.5]

Autowork (autonomous issue-fixing)

• Reuses auto/issue-N branches → a second run on the same issue fails to push (non-fast-forward), and a prior bad run poisons re-runs. [observed]
• Runs in the main checkout, so any dirty working tree blocks it (its own error suggests "use a dedicated worktree"). [observed]
• Patch quality is LLM-bound — it self-corrects via a retry loop, but can still produce weak or mis-targeted diffs. [observed]

Skills & the local kernel

• Onlyskills are real — dream_journal, lucid_dreaming, archive_curator, voice_curator; the rest are design contracts, not live. [CLAUDE.md]
• The local kernel (Ouro-1.4B) is weak — ~10% pass@1 on HumanEval — so quality depends on cloud routing; "models are interchangeable" by design. [measured]

Housekeeping

• A couple of autowork runtime logs (data/agi-benchmark.jsonl, data/convergence-autonomous-work.jsonl) are untracked churn — candidates for the gitignore sweep. [observed]

Improved in this pass: the intent-classifier substring bug above. The remaining items are the live backlog; the highest-leverage next ones are wiring §6.5 (grounding gate + Σ₀ canary) so chat interactions actually feed the loop, and isolating autowork in a dedicated worktree.