Hippo Feature

You are entering hippo-feature mode to build a new memory mechanic from RESEARCH.md. Follow the eval pyramid strictly. Skipping tiers wastes hours of LoCoMo time.

The eval pyramid (non-negotiable)

If a feature shows no Tier 1 signal, do not proceed to Tier 2. If Tier 2 shows a regression, do not run Tier 3.

The loop

0. Pick the feature

If $ARGUMENTS names a feature (e.g. acc-evc, vmpfc-value, dlpfc-goals, vlpfc-gate, pineal-salience-v2), use it. Otherwise read the PFC priority table in RESEARCH.md (lines ~459-466) and propose the top three by effort × benchmark delta. Wait for user confirmation.

1. RED: Write the failing micro fixture FIRST

• Drop a fixture at benchmarks/micro/fixtures/<feature>.json with shape:

  {
    "name": "<feature>-<aspect>",
    "mechanic": "<feature>",
    "description": "<one sentence: what behaviour this proves>",
    "remembers": [...],
    "queries": [{"q": "...", "must_contain_any": [...], "top_k": N}]
  }
  

• The fixture must encode behaviour the current system cannot satisfy. If it passes on main today, the fixture is wrong; make it harder.
• Run python benchmarks/micro/run.py --filter <feature> and confirm it fails.
• Save the failing baseline: python benchmarks/micro/run.py --out benchmarks/micro/results/baseline-<feature>.json.

2. PLAN: Outside voice on non-trivial features

Per global CLAUDE.md outside-voice rule: if the feature touches schema, retrieval ranking, or storage (i.e. anything in the PFC priority table), run /plan-eng-review on the implementation plan before writing code. Skip only for one-line tweaks.

3. GREEN: Minimum implementation

• Smallest diff that makes the fixture pass.
• No new abstractions, no speculative config flags, no "while I'm here" cleanups.
• After each change, python benchmarks/micro/run.py --filter <feature> until it passes.

4. REGRESSION CHECK: Run all micro fixtures

• python benchmarks/micro/run.py: every existing fixture must still pass.
• If any fixture regresses, revert and reconsider. Do not "improve" the regressed fixture to make it pass.

5. TIER 2 SMOKE: Stratified LoCoMo subsample

$env:HIPPO_BIN='node C:/Users/skf_s/hippo/bin/hippo.js'
python benchmarks/locomo/run.py `
  --data benchmarks/locomo/data/locomo10.json `
  --output-dir benchmarks/locomo/results `
  --output-name hippo-smoke-<feature> `
  --conversations 1 --sample 10 `
  --score-mode evidence

• Compare mean_score to the most recent smoke baseline.
• Required: non-negative delta on the affected categories. Variance dominates absolute scores at N=50, so trust the direction, not the level.
• If delta is negative, return to step 3.

6. COMMIT

• One logical commit per feature: code + fixture + result baseline.
• Reference the RESEARCH.md section in the commit body (e.g. RESEARCH.md §4.3 ACC EVC-adaptive recall).
• Never use --no-verify.

7. STOP: Do not run Tier 3

Full LoCoMo only on explicit user request (release gate). Even with a green Tier 2, do not run the 5-8 hour LoCoMo full unless the user asks.

Hard rules

• One feature at a time. Don't bundle ACC + vmPFC into one branch even if RESEARCH.md groups them.
• Real DB for tests (project memory rule). No mocks where the real SQLite store is feasible.
• Power models: DO NOT TOUCH (project memory rule, applies cross-repo to skf_s).
• Salience gate. The v1 60% lexical-overlap gate destroyed LoCoMo from 0.28 to 0.02. Any salience work must be default-off and prove a positive delta on Tier 2 before being enabled.

Pre-flight checks (run before step 1)

• python benchmarks/micro/run.py: baseline must currently be 1.00 pass rate
• node bin/hippo.js --version: record the starting version for the commit

Proactivity: DEFAULT MODE (full-power)

This command runs in full-power proactive mode. Decide, act, report the decision in one line.

• Dirty WIP on master? Commit it to a wip/<feature>-pre branch (preserve, don't lose), reset master, then create feat/<feature> clean. Never overwrite WIP. Never block on it.
• Already on a feature branch? Continue on it.
• On master with clean tree? Create feat/<feature> and switch.
• Outside-voice plan review? Spawn a senior-code-reviewer sub-agent inline against the plan. Don't pause and ask the user to invoke /plan-eng-review.
• Tier 2 smoke (5-10 min)? Run in background; keep working on docs/commit prep; report on completion.
• Real stop-points (failing fixture after honest attempt, plan rejected by reviewer, Tier 2 negative delta, hippo CLI broken): surface those. Housekeeping is not a stop-point.

The user has authorised proactive execution. Ask only at real forks in the road.

Task / feature: $ARGUMENTS

If $ARGUMENTS is empty, list the top three features from RESEARCH.md PFC priority table and ask which to build.