sea-learn-loop — exercise the SEA self-learning completion

Driver: .agents/skills/sea-learn-loop/learn.sh (paths relative to repo root). CLI check — evidence is the RESULT: line and both passes' metadata lines.

SEA is always available in the Rust workspace. The default run exercises the real SEA path. The first pass starts with no evidence; the second must recall the first pass's persisted record after reopening the scratch store. Both must report successful persistence and adaptation. --sea is accepted for compatibility and does not change the build path.

Prerequisites

• Pinned stable Rust 1.99.1 via ./tools/cargo.sh (never bare Homebrew cargo).

Run (agent path)

.agents/skills/sea-learn-loop/learn.sh                       # default SEA-on path
.agents/skills/sea-learn-loop/learn.sh "my custom prompt"    # custom input
.agents/skills/sea-learn-loop/learn.sh --sea                 # accepted; no-op on build path

It builds the locked CLI into the same target directory it executes, creates a unique scratch store under target/skill-scratch/ (never the live ~/.abi/), overrides inherited persistence settings and runs abi complete --learn -- "<input>" in two separate processes. It requires learn=true, local generation provenance, persisted=true, adapted=true, and nonzero second-pass evidence_count on the metadata line, not inside the generated response. Its own scratch directory is removed on exit; other runs' stores are untouched. Prints RESULT: PASS — SEA learning persisted, adapted and recalled across processes. (exit 0) or RESULT: FAIL (nonzero). Audit-vetoed prompts intentionally fail this acceptance smoke because they must not learn.

Gotchas

• First-pass evidence_count=0 is expected in a new scratch store, but a zero second-pass count fails the cross-process recall check.
• The requested-model line shows the catalog default (claude-fable-5) while provider=local and generation_engine=persona-template establish that --learn is not a live API call; no credentials are needed.
• Every metadata field rides that one first line, including learn=true evidence_count=N adapted=B self_learn_cycle=... recall_ready=.... self_learn_cycle reports what actually happened (persistence+adaptation+recall only when the learning write committed, adaptation-only when it was skipped or disabled, adaptation-only+write-unknown when the outcome is unknown) — never a completed cycle that the adjacent persisted= and wdbx_status= fields contradict. Do not move the learn fields onto their own line: the smoke reads only the first line, which is the boundary that stops markers repeated in generated response text from satisfying a persistence check.
• First build may compile the workspace; later ./tools/cargo.sh build -p abi-cli runs are incremental.
• --learn is local persona-template learning, not live-provider or bridge inference. Combining it with --live or a local-bridge model is rejected; this does not train language-model weights.

Troubleshooting

Symptom	Fix
build FAIL	./tools/cargo.sh build -p abi-cli, then ./tools/check.sh.
missing learn=true	complete grammar drifted — check crates/abi-cli complete handler + --learn.