Runtime recall wiring

Wiring that lets every governed harness read the durable KB with bounded automatic recall. Writes are always explicit (,ai-kb remember); hooks never auto-write capsules. Skill contract: ~/.agents/skills/k-ai-kb/SKILL.md.

Persistence vs injection

Path	Behavior
Explicit	,ai-kb search / get / remember — any harness with CLI or agent-pull
Automatic	Only where a safe context-injection channel exists; gated and capped

Runtimes without injection (e.g. Cursor cloud) retain explicit search/write only.

Harness matrix

Shared scripts: session_context.py (start), perturn_recall.py (per-turn), worklog_dispatcher.sh (post-tool). Adapters in repo under home/dot_* / exact_hooks/.

Runtime	Session start and warm-up	Per-turn recall	Worklog	Exception
Cursor	sessionStart; AI_EMBED_WARM=1	beforeSubmitPrompt	shell/tool/edit hooks	Top-level additional_context; 10,000-character carrier
Claude	SessionStart; AI_EMBED_WARM=1	UserPromptSubmit	PostToolUse*	Local llama.cpp settings excluded
Codex	SessionStart; AI_EMBED_WARM=1	UserPromptSubmit	PostToolUse	Absolute paths, shell wrappers, hook-specific output; re-trust after hook changes
Copilot	SDK onSessionStart; payload warm-up	onUserPromptSubmitted	onPostToolUse*	Parent-session env affects subagent writes and status/note; startup recall stays blind
OpenCode	system transform; payload warm-up	chat.message	tool.execute.after	Adapter synthesizes the shared payload shape
Pi	session_start ensure; first before_agent_start	before_agent_start	tool_result	Uses session-aware ,agent-memory status --json
Gemini	SessionStart; AI_EMBED_WARM=1	BeforeAgent	AfterTool
Cursor cloud	none	none	none	Explicit agent-pull only

Shared prefix source: prefix.txt. Custom subagent profiles render it directly; manual tmux prompt wrapping uses the same text.

Without warm-up signal: session context includes ### Recall Notice (self-search mid-task). AI_AGENT_DEPTH=fast skips warm-up and per-turn retrieval.

Startup warm-start (BM25 only)

Gates in session_context.py — all must pass:

Gate	Rule
Topic	Named via .session-topic-<id>.txt; not current or session-*
Spec	Non-empty <topic>.txt becomes query
Lane	BM25 only; no embedder in hook timeout
Scope	workspace-local, domain, or universal
Output	≤3 capsules under ### Relevant Learnings (,ai-kb)

Review/unbound topics: no warm-start. Separate from resident embedder warm-up (per-turn path only).

Per-turn recall gates

perturn_recall.py / Pi ai-kb-recall.ts — hybrid mode, dedupe via .recall-seen-<session-key>.json:

Gate	Value
Query	current prompt
Absolute cosine	≥ 0.55 (best row, not rank-0 — RRF order is relevance-blind)
Relative tail floor	0.85 of best cosine (0.60 for BM25 warm-start)
Scope	workspace-local, domain, universal via --workspace-gate
Connect-only	AI_EMBED_CONNECT_ONLY=1; unavailable worker → omit block, continue

Queries travel over stdin and are never written to process arguments. Tail trimming drops weak rows without reordering BM25 or fused/MMR results. Correction patterns may inject an anti-pattern note directive; durable writes still require verified remember.

Depth profiles (AI_AGENT_DEPTH)

Unset/invalid → balanced.

Depth	BM25 startup	Resident warm-up	Fetch/inject	Prompt cap (chars)	Body cap (chars)	Timeout
fast	yes	skipped	disabled	—	—	—
balanced	yes	requested	6 / 3	600	240	6s
deep	yes	requested	12 / 5	1200	360	9s

fast removes per-turn retrieval, not thresholds. Budgets fixture-backed: recall_worklog_state_machine.py. AI_KB_RECALL_TIMEOUT can raise per-turn timeout only.

Store boundaries

,ai-kb is the sole durable semantic store. Codex auto-memory pinned off (memories = false). /tmp/specs archived daily to ~/.local/share/agent-specs-archive/ (raw preservation; no auto-KB write).

Sources and verification

• exact_k-ai-kb/readonly_SKILL.md
• ai-kb-recall.ts, ~/.agents/hooks/perturn_recall.py
• scripts/tests/recall_worklog_state_machine.py