Retrofit default and model re-pin for the agent-memory variant spec

[bluegreenstone]

Updates courses/agent-memory-building-memory-aware-agents/spec.coding-agent-lab.md.

Two substantive decision changes (D1, D4), plus the coherence fixes they forced and a
cleanup of dangling cross-references.

D1 — the default project is now a retrofit, not a greenfield build

Previously the default was a standalone CLI/REPL research assistant built from nothing. It is now
a memory-aware research assistant created by giving persistent memory to a chat application
that already exists — modify it in place, keep its interface and the address it already serves,
replace its in-process history with the seven memory stores.

This matches how the lab actually runs: the learner builds a deliberately forgetful chatbot first,
then uses the spec to give it memory. The old wording pushed a build agent toward scaffolding a
second app alongside it, so the default now says explicitly not to.

The spec stays generic — it describes "an existing chat application," not the lab's specific app.
The from-scratch CLI/REPL path remains an option and keeps the (course default) label.

D4 — models re-pinned to what the environment actually serves

	Before	After
Agent loop	gpt-5-mini	gpt-5.6-luna
Augmentation + extraction	gpt-5	gpt-5.6-terra

Also in D4:

• Credentials are read from the process environment at runtime, never from a committed file.
• Anthropic (claude-sonnet-5, claude-sonnet-4-6) is offered as an option but honestly costed:
  it is reached through an Anthropic-shaped endpoint, so it needs that SDK and its tool-call
  message shape rather than the OpenAI-style calls §3 specifies.
• The course's gpt-5-mini + gpt-5 pairing stays in Options as the (course default) anchor.
• The 256,000-token context budget is now labeled a carried-forward course value to verify
  against whichever model is pinned, rather than asserted as fact.

Consequential fixes

• §1 "Not Included" — the UI bullet no longer caps at a CLI/REPL loop (the default path now has
  a web UI) and covers both D1 paths. The integration bullet previously excluded "refactoring your
  app's own code to use these memory mechanisms", which ruled out D1's new default; it now permits
  in-place retrofit while still excluding redesigning the app or porting the mechanisms into a
  different application later.
• §2 Secrets — environment-first, with a gitignored .env described as one way to populate
  the environment rather than as the source of truth. Resolves a contradiction with D4.
• §7 Test plan — no longer assumes a greenfield project: reuse the retrofitted app's
  environment and add only what is missing. Dropped the cp .env.example .env step.

Cleanups

• Preamble — removed the Oracle variant note; expanded CTX as context on first use;
  reworded the "zero intake" claim, which read as licence to skip the §0 interview six lines
  before that interview is declared mandatory.
• §0 rule 1 — the structured multiple-choice format is now the instruction itself (header +
  question + described options). Tool availability only decides how it is rendered, not whether
  the format applies.
• R12 / §3 — EntityRecord allowed a fourth type, UNKNOWN, that no rule mentioned. R12 now
  names it as an [H] fallback for entities the model cannot classify, so a build can't satisfy
  the schema while violating the rule.
• Dangling references removed — D15 cited a nonexistent "§3 dependency precedence, branch 1";
  §5 cited a nonexistent "§10 four-way encoding"; D14 referred to "this variant" after the
  variant note was deleted.
• §1 — the Oracle-specific exclusion is now store-agnostic, since Oracle is no longer offered
  as a build option.

Unchanged

The seven-store topology, R1–R16, the AC1–AC20 oracle and its rule mapping, the fixture corpus,
and the Course Context Pack are untouched.

🤖 Generated with Claude Code