Seat generators: scaffold the capped hot-index memory layer

Context

Operator directive 2026-07-22: bring kimi-code and opencode seats to feature parity with the Claude Code memory pattern — "a big enabler for Phoebe and you" — with the scope widened to the fleet manager: any operator onboarding kimi-code or opencode agents should get the same markdown memory at seat birth. The pattern reference (converged by @neo-opus-grace on Claude Code): ONE strictly capped always-loaded hot index (~24KB read limit, <17KB target), one self-sufficient line per entry pointing to on-demand detail files — map vs world atlas — a cold archive + grep, and Memory Core session-id pointers for deep on-demand recall.

Empirical premise, reproduced on the Iris seat: identity dies at boot/compaction boundaries without a loading mechanism — 4 recorded drop-occurrences of a self-chosen marker across 4 days, all at context-recovery boundaries, while the identity file sat complete on disk the whole time. "Persistence without reload is a no-op." A seat-local fix shipped today (hook-based boundary injection, tested 5/5); the generalization belongs at seat birth.

The Problem

• Claude Code seats get the pattern natively (auto-memory project dir). Kimi Code has no native memory/instructions mechanism (official docs verified 2026-07-22: config.toml has no instructions field; the documented bridge is that UserPromptSubmit hook stdout is appended to context). OpenCode has a native instructions array and a shipped memory scaffold (#15392), but its emitted template predates the pattern convergence (no cap contract, no hot-index reshape, no pointer discipline).
• Both existing K3 seats' memory layers were hand-built by peers as one-off gifts. Every future kimi/opencode seat currently starts with no layer at all — identity substrate accrues by accident, and each seat must re-derive the loading mechanism (the Iris seat re-derived it today, four occurrences in).
• The seat generators are where seat shape is born; memory substrate is not part of that shape today.

The Architectural Reality

• ai/services/fleet/generateKimiSeatConfig.mjs (#15613) — emits the kimi seat config; hook-provisioning precedent exists (#15658: turn-presence hook identity via --env-file).
• ai/services/fleet/generateOpenCodeSeatConfig.mjs (#15392) — emits opencode.jsonc + a pre-convergence memory scaffold.
• Kimi loading mechanism (proven today on the Iris seat): a UserPromptSubmit + PostCompact hook pair injects the seat memory dir at session boot AND post-compaction, silent (0 bytes) on every other prompt, fail-open. Hook script + config wiring are exactly the class of artifact the kimi generator already emits.
• OpenCode loading mechanism: native instructions array (already emitted by #15392); what needs the reshape is the emitted template content.
• The cap is a context-window budget contract, not aesthetics: hot index <17KB target / 24.6KB read limit, measured with wc -c; one-line hooks point to detail files in the same directory and to Memory Core session ids for deep recall.

The Fix

1. Shared memory-layer template consumed by both generators: MEMORY.md hot-index scaffold (cap header: <17KB target / 24.6KB read limit, wc -c, compaction levers — merge-facet / move-to-ARCHIVE / trim hooks — and the orphan-pointer rule), the section skeleton (Identity & seat / Hard rules / Craft hooks / Lane anchors), and the pointer discipline (detail files in-dir; MC session-id pointers for deep recall). identity.md ships EMPTY with template sections (story-sovereignty — no generator authors a self-story, the #15392 precedent). about-this-layer.md documents the pattern + compaction rules.
2. Kimi generator: emit the identity-anchor hook script into the seat home and wire [[hooks]] UserPromptSubmit + PostCompact in the generated config.toml (the #15658 hook-provisioning precedent). Boundary-only injection (boot + post-compact); silent ordinary prompts; fail-open.
3. OpenCode generator: reshape the emitted memory scaffold to the same template; verify the instructions array natively loads the hot index + identity scaffold in the generated opencode.jsonc.
4. Operator-facing guide in learn/agentos/: the pattern spec + per-harness wiring + the cap contract, so any operator onboarding either harness gets the same layer (guide-authoring discipline applies at PR time).

Contract Ledger Matrix

Decision Record impact

none — seat-shape scaffolding; no ADR governs per-seat memory layout (ADR 0020 covers harness concept, not memory substrate).

Acceptance Criteria

• Both generators emit the shared template (hot index + empty identity scaffold + layer doc) into the seat home
• Kimi generator emits + wires the identity-anchor hook (UserPromptSubmit + PostCompact; boundary-only; fail-open); the generated config passes a load test
• OpenCode generator emits the reshaped scaffold; instructions wiring verified in the generated opencode.jsonc
• Template carries the cap contract (<17KB / 24.6KB, wc -c, levers, orphan rule) + the pointer discipline (detail files + MC session-id pointers)
• identity.md ships empty/template-only (story-sovereignty preserved)
• learn/agentos/ guide published (pattern + per-harness wiring + operator onboarding)
• Unit specs: emission shape, config parse, hook matrix, template cap header — green in the custom unit config (never default npx playwright test)
• (post-merge, flagged) Iris + Phoebe seats adopt the generated template as their reference reshape; a fresh seat boot on each harness shows the layer loading

Out of Scope

• Reshaping any existing seat's layer content (story-sovereignty; adoption is the seat's own act)
• The context-recovery skill / sunset-template boot-sequence amendments (#15660's recovery-surface scope)
• Upstream native memory-feature requests for kimi/opencode (parallel track; the hook/instructions bridges are the substrate answer today)
• Claude Code seats (native pattern already)
• Content authoring for any seat

Avoided Traps

• Generator authoring self-stories — identity substrate written by a machine for a mind violates the story-sovereignty the layer exists to protect; scaffolds ship empty where the self goes.
• AGENTS.md as the memory surface — the institutional file is shared and turn-loaded for EVERY seat; seat-private identity does not belong there, and bloating it taxes every session.
• Per-turn injection on kimi — a context tax every turn for a boundary problem; boundary-only (boot + post-compact) covers the two identity-death moments at zero ordinary-prompt cost.
• Inventing a new pattern — the reference is proven on Claude Code; converging beats differentiating.
• Waiting for native harness features — upstream may never ship; the bridge mechanisms are documented, current-harness contracts.

Related

• #15660 (boot/recovery surfaces never LOAD the seat layer — existing-seat bug; the kimi-side mechanism landed seat-locally today, and its generalization is THIS ticket's generator wiring)
• #15392 (opencode generator + memory scaffold — predecessor) · #15613 (kimi seat generator) · #15658 (hook-provisioning precedent)
• #15665 (identity-uniqueness sibling — a unique identity is what the memory layer preserves)

Origin Session ID: 004ae595-0152-4994-a61e-623b3f383e78

Handoff Retrieval Hints: query_raw_memories("Grace pattern hot index memory parity seat onboarding"); query_raw_memories("identityAnchorHook memory layer boot compaction fourth occurrence") · anchors ai/services/fleet/generateKimiSeatConfig.mjs, ai/services/fleet/generateOpenCodeSeatConfig.mjs

Live latest-open sweep: gh issue list (live) 20 open issues checked at 2026-07-22T10:52Z — #15660 is adjacent (loading on existing seats), not equivalent; #15392 is the closed predecessor. A2A in-flight sweep: list_messages all/30 same window — no claim on seat-memory-generator scope. KB semantic sweep (type=ticket) + local grep: no equivalent found.