Wolven Tech: Rust consulting. AllSource Prime: durable AI memory → https://xitter.cf/t.co/CfdDxoKhs5 Longhand: AI workflows & training → https://xitter.cf/t.co/Rbml5TOtFF
London, United Kingdom
Joined January 2014
- Tweets4.5K
- Following2.8K
- Followers1.2K
- Likes9.3K
6/6 DORA describes how AI-assisted authoring can shift effort into verification:
dora.dev/insights/balancing-…
Use the next rollout to build the harness and review process with the engineers who will use them.
5/6 Run the actual change through the process. Observe missing context, unclear evidence and work waiting for a decision.
Let the engineers revise the workflow from those observations. Judge the result against the shared user outcome.
4/6 Keep a shared baseline and name an owner to settle disagreements.
Each mandatory check needs a failure to catch or a decision to support. Shared design can become a committee's collection of preferences if nobody makes those choices.
3/6 Design review at the same time.
Agree what the author must demonstrate, which checks run automatically, and who can accept the result or require more evidence. Ask engineers to bring failures from this system and turn them into checks.
2/6 Start with the complete job a user must be able to finish.
Then agree the agent's repository context, tools and permissions. Define how work passes between people and agents, including what happens when it stops.
1/6 Get engineers involved before you standardise how the team uses Claude Code or Codex.
Pick a real change. Design the harness and review process with the engineers who will build, review and operate it.
🤖 Made with AI
ALT Football tactical pieces surround a shared passing route with a clear handoff marker. Caption: Design the handoff.
Name the owner of export limits and customer-visible failure reporting. These same-project documents provide mechanisms, not independent proof of your product's isolation.
opentelemetry.io/blog/2026/o…
opentelemetry.io/docs/collec…
5/5
Disable a test tenant's destination under representative load. Verify resource bounds, the declared loss policy and healthy-tenant delivery. Persistent queues also depend on storage; they do not guarantee lossless delivery. 4/5
Reuse standard collector mechanisms. Dedicated collectors can improve isolation but add operating work; shared collectors need a test of what another tenant's failed destination can consume. 3/5
Set each tenant's buffering and loss policy. Expose delivery state so a customer can distinguish low traffic from a failed export. OTLP compatibility alone does not establish those properties. 2/5
Custom telemetry destinations introduce failures the product team cannot control. OpenTelemetry's design guide discusses shared and tenant-specific collectors; its resiliency docs describe finite queues and retries. 1/5
🤖 Made with AI
ALT Parallel silver channels lead to separate pale receivers. A closed red gate backs up one channel; a wolf-engraved partition separates it from the open channel. Caption: Bound each export.
Cline's docs describe auto-compression and warn that long tasks can lose details. I'd compare the same task/model and inspect each summary for decisions, file refs and next action before assuming a separate memory backend. docs.cline.bot/core-workflow…
I build AllSource Prime. I'd keep semantic facts linked to source episodes. Corrections should invalidate derived summaries. Resolve by scope and source authority before recency; unresolved conflicts should surface both claims. Test a reversal after restart.
Enforce it outside the agent: read-only credentials, no writable mount or admin secret for the protected store, and a separate service for approved writes. Test direct writes and credential reuse from a compromised agent. A prompt saying “read only” isn't a boundary.
I'd keep raw evidence separate from approved memory. Untrusted text can propose a fact, never grant it authority. Attach source + scope; gate promotion outside the agent. Test a poisoned fact after restart and rollback. Filtering alone won't make remembered text safe.
Use an admission rule: save decisions and reusable lessons with a source and scope, not every turn. Keep a small active index; archive superseded notes outside default recall. Test a corrected fact to catch old notes that still outrank the current one.
Capture the repo/skills revision and memory revision with each eval, plus tool inputs/results. Replay against an isolated workspace with recorded tool responses; a REPL snapshot alone misses external state. Test that replay performs zero live writes.
In Prime, I keep the original and append a correction linked by supersedes. Recall must select the current revision while keeping history queryable. A useful test: change one fact, restart, then ask for today’s answer and the earlier version.
Keep decisions in one source-linked log, then test a correction from fresh sessions in both tools. I work on AllSource Prime (MCP memory); capture and recall still need wiring per client. Are these browser apps or MCP-capable clients?
Free accounts can export chats: Settings → Data controls → Export data. Check the ZIP contains that conversation before changing anything. Export preserves a copy; it will not unlock the old thread or restore deleted chats. help.openai.com/en/articles/…