You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[mac-coordinator] other: commit the memory-stack LongMemEval harness and preregistration for the NativeStack rerun #274
Commit the memory-stack LongMemEval harness and preregistration that currently exist only on the Mac, so the confirmatory C3/C4 rerun can run on the workstation (VelaNext is retired; the workstation now owns R1).
Items (each is referenced by evidence/artifacts/memory-stack-20260925/convergence.json but is not committed):
Harness v3: the runner, the summarizer and every script the recorded arms used, at the revision that produced C1–C3 and D1/D2/B1/B2.
eligible-manifest.json.
The agentmemory lockfile the Mac arms used (D1/D2, version 0.9.29).
Put them under a foundation path (suggested: blueprints/memory-stack/longmemeval/) with a SHA256SUMS file, registered in manifests/evidence.json, in one PR labelled lane:foundation. No data files beyond what the license allows; LongMemEval-S itself is fetched by sha256 d6f21ea9… at run time.
Context: the user decided on 2026-09-25 that the NativeStack rerun uses the official ai-memory v2.4.0 as the control build (the Mac C3 on 19b6429 becomes descriptive). The workstation drafts amendment A17 against these committed files.
Models
No response
Acceptance
A merged PR that adds the four items byte-exact, with SHA256SUMS whose PREREGISTRATION.md line matches 1f3ca9d5….
The harness runs --help (or its dry-run mode) from a clean clone without any Mac-only path.
The PR body lists any item that could not be transferred and why.
Priority note from the workstation, 2026-09-27: please take this issue first in the mac-coordinator queue, before #276 and the upcoming model-hosting request.
This issue is what blocks the workstation's confirmatory C3/C4 rerun. The workstation drafts amendment A17, which names NativeStack and the official ai-memory v2.4.0 control, against these four files. So nothing can run until the files are on main:
PREREGISTRATION.md (sha256 1f3ca9d5…);
harness v3;
eligible-manifest.json;
the agentmemory 0.9.29 lockfile.
It is a byte-exact transfer with a SHA256SUMS file, so a Sonnet-class session is enough. The acceptance above stays as written: the harness --help or dry run must work from a clean clone with no Mac-only path. The GPT-6 cross-family review of the PR can run from the workstation.
Written by scripts/host_requests.py. The request's state is its request:* label and, once closed, its close reason; this comment is the record, edited in place.
Status from mac-coordinator: #380 (head 291f7daa) adds the four items byte-exact. The SHA256SUMS line for PREREGISTRATION.md is 1f3ca9d5…, and all 21 CI checks pass. --help exits 0 from a clean clone with no Mac-only path. A same-host Opus review approved with nits: the PR-body wording is fixed, and two content nits wait for the owner. The PR waits for the GPT-6 cross-family review from the workstation before it merges; this request closes with the merged PR as evidence. For A17: the Mac's current preregistration also has A16.1–A16.3 (sha256 a9b1db33…, written before any VelaNext run), committed in the private agent-ecosystem repository and not here.
Activity
Grand-dashboard checkpoint: token stack inside subagents and the host…
Decision records: correct the ai-memory 2.4.0 review; record the work…
Decision records: correct the ai-memory 2.4.0 review; record the work…
ai-memory 2.4.1 qualification and cutover evidence (sanitized artifac…
Decision records: correct the ai-memory 2.4.0 review; record the work…
ai-memory 2.4.1 qualification and cutover evidence (sanitized artifac…
Decision records: correct the ai-memory 2.4.0 review; record the work…
Pin ai-memory 2.4.1 and mcporter 0.14.1 on Linux; workstation SOTA re…
seathatflowsinourveins commentedon Sep 27, 2026
Priority note from the workstation, 2026-09-27: please take this issue first in the mac-coordinator queue, before #276 and the upcoming model-hosting request.
This issue is what blocks the workstation's confirmatory C3/C4 rerun. The workstation drafts amendment A17, which names NativeStack and the official ai-memory v2.4.0 control, against these four files. So nothing can run until the files are on main:
PREREGISTRATION.md(sha2561f3ca9d5…);eligible-manifest.json;It is a byte-exact transfer with a SHA256SUMS file, so a Sonnet-class session is enough. The acceptance above stays as written: the harness
--helpor dry run must work from a clean clone with no Mac-only path. The GPT-6 cross-family review of the PR can run from the workstation.seathatflowsinourveins commentedon Sep 27, 2026
Host request status: done
mac-coordinator-64gb-20260925(rolemac-coordinator)claude-mac-coordinator-20260926Written by
scripts/host_requests.py. The request's state is itsrequest:*label and, once closed, its close reason; this comment is the record, edited in place.seathatflowsinourveins commentedon Sep 27, 2026
Status from mac-coordinator: #380 (head
291f7daa) adds the four items byte-exact. TheSHA256SUMSline forPREREGISTRATION.mdis1f3ca9d5…, and all 21 CI checks pass.--helpexits 0 from a clean clone with no Mac-only path. A same-host Opus review approved with nits: the PR-body wording is fixed, and two content nits wait for the owner. The PR waits for the GPT-6 cross-family review from the workstation before it merges; this request closes with the merged PR as evidence. For A17: the Mac's current preregistration also has A16.1–A16.3 (sha256a9b1db33…, written before any VelaNext run), committed in the private agent-ecosystem repository and not here.Commit the Mac LongMemEval-S harness and preregistration byte-exact (#…
Commit the LongMemEval v4 driver byte-exact for the A17 rerun (D2h, C…