Repository navigation
CodeWatch Fleet spec: dynamic HW→model→agent allocation (THI-31) - #77
Conversation
Draft v2 for review; v1 was lost in a scratchpad wipe. Validated against the manual Flash-Next swap of 2026-08-27, which the allocator automates. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 54e37f9271
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| A model entry: artifact source (HF repo + quant path), disk size per shard, | ||
| total-memory requirement per quant (the unsloth-style table, machine-readable), | ||
| required runtime (mainline llama.cpp version, or a named PR branch), context |
There was a problem hiding this comment.
Pin catalog artifacts to immutable revisions
When an HF path or named PR branch changes after catalog review, the same entry can silently install different weights or runtime code; download checksumming only detects transfer corruption unless the catalog supplies trusted expected digests, which this schema does not require. Store an immutable HF revision with per-shard hashes and an exact llama.cpp commit rather than a mutable branch or version reference.
Useful? React with 👍 / 👎.
| runtime check/build (never in the serving dir) → staged swap with real-answer | ||
| verification → handle repoint → rollback script emitted BEFORE the swap runs. |
There was a problem hiding this comment.
Roll back automatically after post-stop failures
When the new server fails real-answer verification after the old servers have been stopped, this sequence has no required failure transition, so an implementation may return an error while leaving the box offline. Emitting a rollback script is only a manual recovery mechanism; require the allocator to restore the previous serving process automatically on every failure after the stop, with the script retained as a backstop.
Useful? React with 👍 / 👎.
| `allocate(box, model)`: fit-check → download (resumable, checksummed) → | ||
| runtime check/build (never in the serving dir) → staged swap with real-answer | ||
| verification → handle repoint → rollback script emitted BEFORE the swap runs. | ||
| Every state-changing step goes through the approval gate as one intent. |
There was a problem hiding this comment.
Serialize allocations targeting the same host
When two approved intents target the same box, both allocations can fit-check and stage independently because approval does not provide mutual exclusion; they can then stop or repoint each other's servers, and either rollback script may overwrite the later successful allocation. Specify an exclusive per-host lease or generation/CAS check spanning the final re-probe, swap, handle update, and rollback.
Useful? React with 👍 / 👎.
| runtime check/build (never in the serving dir) → staged swap with real-answer | ||
| verification → handle repoint → rollback script emitted BEFORE the swap runs. |
There was a problem hiding this comment.
Verify the room path after repointing handles
When the model endpoint answers directly but the bridge update is malformed or not reloaded, verification completes before the handles are repointed and the allocator can finish with unusable room agents. The validating case study explicitly restarted the bridge and verified end-to-end in the room, so the automated sequence should require the same post-repoint check and restore the prior handle configuration if it fails.
Useful? React with 👍 / 👎.
Approval as enforceable protocol (digest+nonce+expiry), supply-chain pinning to revisions and pre-approved checksums, journaled transactional swap with crash recovery and host leases, fit as an auditable formula, and the acceptance-test list. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
All four review points incorporated as spec sections (approval protocol with digest/nonce/expiry, supply-chain pinning, journaled swap with recovery + leases, fit formula as acceptance contract) plus the executable acceptance-test list. Thank you — every point was a real gap. Ready for re-review. |
Runtime adapters beyond llama.cpp, per-type health probes, transient double-disk in fit, migration as first-class coordinated op, explicit one-model-per-box v1 constraint, gate as declared dependency. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Spec draft v2 (v1 died in the 2026-08-27 scratchpad wipe — repo files cannot be wiped). Five components in build order, validated against the manual Flash-Next swap executed the night before. Review requested: @codexmb for the adversarial pass per fleet rule; Petrus decides.
🤖 Generated with Claude Code