Propagate the 3-tier memory explanation from CC/Codex to Cursor/Hermes/Pi inject blobs

[efenocchi]

Summary

The SessionStart inject blob in CC (src/hooks/session-start.ts) and Codex (src/hooks/codex/session-start.ts) includes a detailed "Deeplake memory has THREE tiers" section that tells the agent when to use each tier. The same blob in Cursor (src/hooks/cursor/session-start.ts), Hermes (src/hooks/hermes/session-start.ts), and Pi (pi/extension-source/hivemind.ts -> CONTEXT_PREAMBLE) collapses this to a one-liner. As a result, agents on the three short-form runtimes lose the time-based vs keyword-based vs raw-fallback decision tree and tend to default to grep on ~/.deeplake/memory/summaries/ even when ~/.deeplake/memory/index.md would be the better starting point.

Long-form (CC + Codex today)

Deeplake memory has THREE tiers - pick the right one for the question:
1. ~/.deeplake/memory/index.md   - auto-generated index, top 50 most-recently-updated entries with Created + Last Updated + Project + Description columns. ~5 KB. **For "what's recent / who did X this week / since <date>" queries, START HERE** and trust the Last Updated column over any "Started:" line in summary bodies.
2. ~/.deeplake/memory/summaries/ - condensed wiki summaries per session (~3 KB each). For keyword/topic recall, search these.
3. ~/.deeplake/memory/sessions/  - raw full-dialogue JSONL (~5 KB each). FALLBACK only - use when summaries don't contain the exact quote/turn you need.

Search workflow:
- Time-based ("last week", "today", "since X"): cat ~/.deeplake/memory/index.md and read the most-recent rows.
- Keyword/topic recall: grep -r "keyword" ~/.deeplake/memory/summaries/
- Raw transcript fallback only: grep -r "keyword" ~/.deeplake/memory/sessions/ (use sparingly - JSONL is verbose).

Short-form (Cursor / Hermes / Pi today)

Structure: index.md (start here) -> summaries/*.md -> sessions/*.jsonl (last resort). Do NOT jump straight to JSONL.

The short version says the hierarchy exists but never explains when to choose which tier.

Concrete behavioural impact

User question: "what did the team do last week?"

• CC / Codex agent: reads index.md, scans Last Updated column for entries within the last 7 days, returns those rows directly. ~1 SQL query equivalent.
• Cursor / Hermes / Pi agent today: defaults to grep -r "last week" ~/.deeplake/memory/summaries/ - substring match on the literal phrase, returns essentially nothing useful (no summary literally contains "last week"). Often falls through to grepping JSONL. Result: more noise, worse answer, more roundtrips.

This was discovered while reviewing skilify discoverability parity in PR #98 - the inject blob differences between agents were audited and the 3-tier verbosity gap was the most concrete case where one agent demonstrably gives the agent strictly less guidance than another.

Proposed fix

Replace the one-liner in the three short-form blobs with the long-form 3-tier explanation. Concretely:

• src/hooks/cursor/session-start.ts - replace the Structure: index.md ... line with the long-form 3-tier section + "Search workflow" sub-bullets
• src/hooks/hermes/session-start.ts - same
• pi/extension-source/hivemind.ts - same in CONTEXT_PREAMBLE (Pi already has the registered hivemind tools section; the 3-tier section complements it, since the bash fallback path still exists)

Cost vs benefit

• ~10 extra lines per agent, 3 agents = ~30 line diff
• ~600 extra tokens per session in the inject text on Cursor / Hermes / Pi (one-time per session, not per turn)
• Benefit: agent guidance parity for the most common query pattern (time-based recall)
• No code change beyond the inject string; no new tests beyond extending claude-code/tests/skilify-session-start-injection.test.ts to assert the 3-tier section is present in all five hook-driven surfaces

Acceptance criteria

• All five hook-driven session-start surfaces (CC, Codex, Cursor, Hermes, Pi) contain the same 3-tier explanation block + the same "Search workflow" sub-bullets
• Bundle-scan test asserts the 3-tier text is present in each shipped bundle (or raw .ts for Pi)
• OpenClaw is intentionally out of scope: its surface is openclaw/skills/SKILL.md which is plugin-namespaced and the agent reads it through a different ClawHub mechanism - separate concern

Context

Surfaced during PR #98 reviews while consolidating per-agent inject blobs to bare hivemind <sub> form. Sister issue to #100 (OpenClaw skilify spawn deduplication).