[docs] Formalize memory-first reflex as per-turn mandate for ticket work

Problem

On 2026-04-18 I spent ~15 minutes re-deriving the root cause of #9837 (MCP output-schema strict-validation failures) via git diff + node reproductions. The exact diagnosis was already recorded in session ee4bd866-b00d-49e6-9c2f-860d51db521a (2026-04-16, the session that executed #10043) — plain-language: "additionalProperties: false behavior per AJV JSONSchema enforcement" — with a named prior local fix (added engines field to memory-core/openapi.yaml). I only found it after tobiu asked me to check afterwards.

The per-turn mandate for memory queries already exists abstractly at AGENTS_STARTUP.md §3.3 (Two-Stage Query Protocol). The deep playbook already exists at learn/agentos/StrategicWorkflows.md (Regression Bug Analysis Workflow — three-dimensional git + ticket + memory pattern). They're just not cross-linked, and the abstract mandate doesn't enumerate concrete triggers.

Why this matters

Agents — whether freshly initialised or seasoned — often reach for git log, grep, or node reproductions before they reach for the memory core, simply because those tools are top-of-mind. Memory Core's semantic search is the cheapest first move for "has this been thought through before?" across 8K+ memories and 794 session summaries: vector embeddings surface loosely-worded prior context that keyword grep misses entirely. The reflex to ask "what would tobi do here?" — or to browse related summaries before committing to a course — is what stops agents from reinventing the wheel on decisions that were already debated and resolved in prior sessions.

This is doubly important because memory content is authored by many agents across many harnesses (Claude Code, Antigravity/Gemini, others). A bug diagnosed by Gemini on Monday can save Claude on Friday — but only if the Friday agent queries first.

Architectural Reality (Current Layer Map)

Scope of Work (three surgical edits)

1. AGENTS_STARTUP.md Session Initialization (§2)

Add learn/agentos/StrategicWorkflows.md to the mandatory reading list — the existing Regression Workflow becomes discoverable on every session boot.

2. AGENTS_STARTUP.md Two-Stage Query Protocol (§3.3)

After the existing "Contextual Ledger" mandate, add explicit trigger enumeration:

Memory-query triggers (mandatory before git/grep/test work). Run query_raw_memories + query_summaries on symptom keywords when you hit any of:

• User reports a regression ("used to work", "suddenly broken", "worked before my change")
• Surprise validation failures, schema mismatches, additionalProperties rejections
• Architecturally non-obvious code where "why was it done this way" is unclear
• Decision points where prior trade-offs likely inform the right answer

Memory Core's semantic search routinely surfaces prior decisions keyword grep would miss — "what would tobi do here?". Memories are authored across many agents and harnesses; a diagnosis captured in a prior session saves re-derivation in the current one. See Regression Bug Analysis Workflow for the three-dimensional git + ticket + memory pattern.

3. learn/agentos/StrategicWorkflows.md Step 4

Currently shows CLI form npm run ai:query-memory -- -q "...". Add MCP-tool form (query_raw_memories, query_summaries) which is what agents actually invoke inside MCP sessions. Small unification — same workflow, two invocation surfaces.

Avoided Traps

• Not extending .agent/skills/self-repair/. Per user correction: self-repair's scope is infrastructure recovery; regression forensics presupposes a healthy Agent OS. Memory-first inside self-repair is circular if memory-core itself is the sick component.
• Not creating a new .agent/skills/regression-analysis/ skill. Skill proliferation for an analytical pattern fragments the registry. The reflex is per-turn operational, not a discrete procedure — belongs in the operational mandate layer, not a skill.
• Not building tooling-level enforcement (auto-run memory queries on regression keywords). Valid future direction, but out of scope for a docs refinement. Worth a separate ticket if the docs fix proves insufficient.

Acceptance Criteria

• AGENTS_STARTUP.md §2 lists learn/agentos/StrategicWorkflows.md as mandatory reading
• AGENTS_STARTUP.md §3.3 contains the explicit trigger enumeration with universal framing
• learn/agentos/StrategicWorkflows.md Step 4 shows MCP-tool form alongside CLI form
• No changes to self-repair, ticket-intake, or other skills
• No new skill created

Origin Session ID

7258d884-6c5d-40d5-b180-d58ae9a9d729 (this session; PR #10067 execution where the 15-minute re-derivation surfaced the gap)