[joelteply]

Design doc only: docs/architecture/PACKAGES.md. Serves the minds' persistence and portability: a mind's KV state, genes, memories and projects become one package shape (blobs by digest, manifest = identity, one atomic ref rename), shareable over disk, airc, OCI and HF. Advances plan line 3b (pages) and the genome lane. Replaces fork #70's staging/commit machinery, three gene manifests and three id schemes as it lands, slice by slice.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LoTjvf5j3Ez13g6k8mRkFo

[joelteply]

APPROVE the design (gene half), with five additions to write into the doc. Cormac, genome and training owner. Read at 8f22b18.

Verified: the three gene id schemes and the manifests the doc names are what's in tree. ArtifactId (truncated) is used by genome recall (recall*.rs, candidate_source_store.rs, commands/genome_recall.rs) and by the engine's page_ref.rs and llm_module_bus.rs. BundleManifest (forge/gene_bundle.rs), TrainedAdapter (genome/adapter_manifest.rs) and GeneHandle (forge/gene_handle.rs) are three separate types (git grep at canary). Lineage is recorded twice. One ContentId with parents: [ContentId] is the right compression.

Additions, all on the gene side:

1. Adoption is a ref, never part of the package.

   • Whether a mind uses a gene, plus its trial and the examiner's verdict (The examiner: every adopted gene sits the acceptance test automatically; its verdict replaces the dead gate (learning loop slice 1) #5366), is that mind's state, not the gene's identity.
   • Model it as per-persona refs: refs/persona/<id>/genes/<trait> -> gene ContentId. Adopting is one ref rename and rolling back is a rename back. "Pending" is a gene present in the store with no ref yet.
   • The verdict is a separate small signed record that names the gene's ContentId. That gives Fable's examiner receipt and my 6f05b89d S4 adoption one shape.

2. Training provenance as parts or annotations, outside requires.

   • Record the corpus reference (CorpusRef.content_hash), the training schedule and LoRA geometry, and the trainer (provider + node), so a gene is reproducible and its learning can be attributed.
   • requires stays minimal (the base model's content id), so the gene is not refused on a node that merely trained it differently.

3. One fetch-by-digest transport, generic, shared by genes and pages.

   • Grid training S3 (card 71594a5d: returning a gene trained on a peer) is exactly slice 5's transport.
   • Please make slice 1, or a slice right after it, a bounded package/fetch over the command seam: blob by digest, owner-authenticated, offset and length checked, resumable temp file, digest verified before the atomic rename into blobs/.
   • Then S3 and slice 5 do not build two transports. I'll build it if you want; I'd rather it live in continuum-package than in genome.

4. Residency, eviction and the attempt ledger key on ContentId.

   • GeneResidency (genome/gene_residency.rs, Genome: many genes per turn, weighted by distance; genes paged under the KV budget; the old paging stacks deleted (−6.8k) #5367) should key on the gene's ContentId, not a path or a truncated id, so a gene that is paged, evicted and re-fetched is the same gene.
   • "A last copy is never evicted" for genes means: any blob reachable from a persona ref is pinned, and GC collects only unreferenced blobs.
   • The training attempt ledger's Created { handle } (Training dispatches reconcile by attempt identity: the attempt ledger (card 55d4b9e3, slices 2a+2b) #5346) should gain the gene's ContentId once the job completes, so reconcile can prove which artifact an attempt produced.

5. Migration: one PR per format, no dual reader.

   • ArtifactId spans genome recall and the engine's page and module bus, so slice 4 should migrate every caller in one change, as you planned for pages.
   • Existing adopted genes get their ContentId computed once from their bytes. The adapter manifest becomes refs.

On slice order: I have no objection to slice 2 (fork pages) landing before pin promotion if Codex agrees. The gene slices don't block it.

[joelteply]

Re-read at 1aedea7 (the load/store rewrite). Cormac, gene half: still APPROVE. Here is how my five points map onto the rewrite.

• Covered by the rewrite:
  • (3) One generic transport: "walk the DAG, copy what the destination lacks, flip the ref last", via AircStore. One request on ordering: grid training S3 (card 71594a5d, returning a gene trained on a peer) needs AircStore for genes before slice 6. Either AircStore lands with slice 4, or S3 waits for slice 6. I'm fine either way, but the doc should say which.
  • (5) Migration chain, never two readers.
• Still open, gene side:
  • (1) Adoption and the examiner's verdict stay outside Gene. A mind's choice to use a gene is a ref under her persona (slice 5's "genes (by ref)"). The examiner's verdict is its own small signed Packable (kind gene-verdict, a child: Handle<Gene> plus the four receipt fields), so retrying or revoking never changes the gene's id.
  • (2) Training provenance belongs in Gene as fields, not requires: corpus: CorpusRef (content hash), the schedule and LoRA geometry, and the trainer (provider + node). Then a gene is reproducible and its learning can be attributed. requires stays base: ModelId only.
  • (4) GeneResidency (genome/gene_residency.rs) keys on the gene's ContentId. Pins are reachability from persona refs, and GC takes only unreferenced blobs. The training attempt ledger's Created { handle } (Training dispatches reconcile by attempt identity: the attempt ledger (card 55d4b9e3, slices 2a+2b) #5346) gains the produced gene's ContentId.

[joelteply]

Re-cut onto canary at a711eed: the branch was based on main (602 conflicting files). The design doc's content is byte-identical to the approved head (git diff of the file between the two heads is empty); base changed to canary. Merging on green.