jason.woltje
193479b52d
Mosaic concepts pages now own the adapted content; source/license metadata under docs/reference/concepts. Adds ACT-1 agent-context planning capture, pinned concept test package + preparation utility, foundation observation notes (durability, evidence, federation, onboarding, workflow), and the #1495 consolidation assessment. TOOLS.md updated for the host-dev launcher.
47 lines
2.7 KiB
Markdown
47 lines
2.7 KiB
Markdown
# Owner-reported durability failure and runtime WAL direction
|
|
|
|
Date: 2026-09-06. Recorded by darkwing.
|
|
Status: owner-reported context and direction for later runtime design; not an
|
|
independently investigated incident or an implementation/test receipt.
|
|
|
|
Jason reports that the separate ~/.mosaic Stack v1 deployment is fragile under
|
|
load; agents retain work in memory rather than reliably tracking it in the system,
|
|
and agent state does not reliably survive restarts. He directs that agent actions
|
|
use WAL for durability and expects tasks, missions, goals, workspaces, projects and
|
|
kanban to provide stronger control over system flow.
|
|
|
|
No investigation/intervention in that deployment is requested or performed.
|
|
|
|
## Required follow-through in the later runtime charter
|
|
|
|
- Trusted admission/publisher code enforces write-ahead intent for governed effects.
|
|
Compliance must not depend on a model voluntarily writing a note or remembering
|
|
to update a task. This builds on the accepted pre-effect recording requirement.
|
|
- Specify the durability boundary: flush/sync and acknowledgement ordering, atomic
|
|
publication, recovery of incomplete/torn records, writer ownership and failure
|
|
behavior. Merely appending to a file or calling something WAL is insufficient.
|
|
- Record stable action/request identity, exact scope/assignment/intent revisions,
|
|
and recovery evidence without secrets. Keep work records distinct from private
|
|
transcripts and enforce their visibility rules.
|
|
- Reconstruct selected work and outstanding obligations from authoritative records
|
|
after restart. Chat memory and UI caches cannot be the only source of truth.
|
|
Kanban and other surfaces must view the same work state, not competing task lists.
|
|
- Reconcile uncertain external effects using evidence; WAL does not make arbitrary
|
|
external commands exactly-once or justify blind replay after a crash.
|
|
- Test crash/restart points around intent, acknowledgement, execution and result;
|
|
concurrent writers and load; disk-full/write/sync failures; and loss of a required
|
|
audit path. Verify both preserved work and fail-closed admission. Preserve the
|
|
separately preauthorized fail-safe stopping path.
|
|
|
|
## Scope boundary
|
|
|
|
The accepted map already identifies legacy writeOnce's lack of fsync and the need
|
|
for a trusted durable publisher/recovery boundary. This observation sharpens later
|
|
runtime acceptance requirements; it does not prove a specific v1 root cause.
|
|
|
|
The approved inspector charter remains frozen at
|
|
19b6721128a627a2032ffdb95ece2d50abe69a8f6d521e9eff8bbdaff22798b6.
|
|
FI-ROCKO-3 implements only that offline synthetic inspector. WAL/runtime persistence,
|
|
load remediation and live migration are not added to its scope. A synthetic preview
|
|
cannot establish these durability guarantees. No worker is retasked by this note.
|