ghost

ghost — the meta-harness

ghost is the orchestrating layer: it decides how to run a task and learns from the outcome. It is a consumer of two services, not a substrate itself:

• kypp (~/code/kypp) — the governed memory engine (ghost's recall/brief).
• pillbox — the substrate (sandbox, §0 log, the verifiable session score, cost).

Naming: ghost = meta-harness, kypp = its memory engine. Distinct components.

Two modes today, both reusing ../eval/gate.py's proven run→score→frozen-task substrate:

ghost.py — router (v1)

Per-task model selection. Policies always:<model> / cascade:<m1,m2,…>; metric = cost-adjusted quality; computes an oracle ceiling (is a learned router even worth building on this set?). See the file header. The cost-router in ../router/ is the memory-backed sibling (learns which model clears the bar per task-class).

ace.py — ACE loop over kypp

ACE (Agentic Context Engineering): a Generator→Reflector→Curator loop that grows a playbook from execution feedback rather than rewriting a prompt. Every stage is a kypp verb, and the playbook is kypp's governed memory (no second store):

ACE stage	ghost does	kypp verb
inject playbook	prepend task-relevant claims to the prompt	kypp recall <task> (compose)
Generator	run the worker, grade it	pillbox run + session score
Reflector	mine the failure trajectory → lessons	kypp capture --distill
Curator	dedup / promote-corroborated / supersede	kypp consolidate
helpful/harmful	attribute the score to the claims a run saw	kypp usage + score

ACE bullets become kypp claims, inheriting governance (authority, corroboration, staleness) that AxACE's flat playbook lacks.

Injection is task-conditioned recall by default, not briefing (dump-all): dumping the full store pollutes a cheap model (scores below baseline — kypp handoff HANDOFF-kypp-kimi.md §2.1). --inject {recall,briefing,none} selects compose / dump-all / baseline for the ablation; semantic targeting needs KYPP_EMBED_MODEL.

What it measures: held-out quality as the playbook grows over iterations — the accrual question ("does remembering lessons help?"), kept honest by a fixed held-out split the loop never reflects on. This is not a quality-lift claim against the σ̂ wall (the parked optimization gate); it's the runtime-memory question the memory matrix (../eval/memory/) validated at the single-task level, now in a loop.

Curator remove-harmful (credit-assignment #2): ace.py computes a per-claim helpful/harmful signal ghost-side from kypp usage (which claims a run saw) + the run score, flags harmful candidates, and kypp reject <handle> (landed) does the demote (status=rejected → dropped from recall/briefing, row preserved). Gated behind --prune-harmful (off by default — the attribution is correlational, so don't auto-prune unsupervised on one round's evidence).

Run

python3 ghost/ace.py --self-test          # attribution math, no agent

scripts/lk-build.sh                        # codesigned libkrun binary (from repo root)
KYPP_DISTILL_MODEL=zai-coding-plan/glm-5.1 \
PILLBOX=./target/debug/pillbox \
  python3 ghost/ace.py --train aider --iters 3 \
    --worker-model zai-coding-plan/glm-4.5-air --project ace-aider

Prereqs (same as gate.py): frozen aider/{train,held-out}/* bookmarks in the evals pillbox (../eval/freeze-split.sh), opencode authed, runner image present, kypp on PATH. The Reflector's distiller model is kypp's KYPP_DISTILL_MODEL env (a frontier reflector over a cheap worker = the teacher→student split).