DSHREM economy swarm: checkpoints, routes, and real cost evidence
Practice a bounded worker and reviewer workflow without making a paid call. Learn which model aliases belong to which providers, how to reconcile an uncertain run, and why a reservation, estimate, settlement record, and invoice answer different questions.
🎯 For Operators planning bounded DSHREM worker tasks and reviewing their receipts
- Keep task caps shared across route changes and retries
- Checkpoint and reconcile an uncertain result before another attempt
- Separate requested model, served model, reservation, recorded cost, and invoice
Practice
Lesson 1 — make a no-spend plan
Goal: prepare a worker request without executing it. The economy profile chooses only before the next task. The explicit aliases keep their existing meaning; quota auto remains the existing Haiku policy.
Exact commands for source inspection only:
dshrem --help
dshrem models
These commands request local help/catalog output; they are not worker prompts. A paid API command such as dshrem ds41f run --read-only --budget 0.05 is included only as a planning example here: do not execute it during this lesson. A numeric budget is a cap, not an operator approval. The root task cap in this request is USD 0.20 total across aliases and retries; the course itself makes no paid call.
Expected result: help and model metadata, if the installed client supports them, without a new worker call. Record the requested alias, requested ceiling, task purpose, expected artifact, and stop condition before asking for an authorized execution.
Common mistake: believing --read-only makes a model call free. Read-only constrains file effects; it does not erase API billing. Also, do not assume the currently deployed subscription auto route creates a new subscription path: the inspected source describes its current automatic route as Haiku.
Lesson 2 — keep aliases and prices precise
Goal: distinguish routes before estimating. The 7 November lane inventory declares ds41f and alias ds41s as DeepInfra DeepSeek V4.1 Flash; ds41f-official is the official DeepSeek route; glm53f is DeepInfra GLM-5.3-Flash. They are not interchangeable. This course treats those source declarations as dated configuration evidence, not proof of current service health.
Exact commands to display current local metadata without a prompt:
dshrem models
For the official DeepSeek schedule, convert current UTC to local seasonal Kyiv time instead of hardcoding an offset:
TZ=Europe/Kyiv date '+%Y-%m-%d %H:%M %Z (%z)'
Expected result: a local catalog and a timestamp with the current Europe/Kyiv zone abbreviation and offset. Official peak windows documented in the lane source are weekdays 01–04 and 06–10 UTC, except Chinese public holidays. DeepInfra's quoted promotion was 0.098/0.294 USD per million input/output tokens and 0.003 cached input; reserve using the declared full bounds 0.20/0.60 instead. Those quotes are date-bound and do not predict actual usage or charges.
Common mistake: substitute the official route because the name looks similar, or display a fixed Kyiv offset all year. Never infer provider from a model nickname alone; use the recorded alias, route, served model, and receipt.
Lesson 3 — checkpoint before retry
Goal: stop after an ambiguous timeout and assemble evidence before a new attempt. A timeout can leave a provider request in flight or a charge unknown.
Exact commands are inspection placeholders; replace paths only with the actual task receipt locations:
dshrem status
dshrem list --json
Do not run find over a broad system path. Inspect the existing task id, session id, requested model, served model, status, held reservation, provider response id, usage, and actual reported cost where available. Preserve the original receipt; append a reconciliation note with UTC time and source. If any required value is missing, mark it unknown or NOT_VERIFIED.
Expected result: a checkpoint that distinguishes input accepted, worker started, response received, artifact written, and cost settled. Only then decide whether another attempt is within the same root cap and has explicit authorization.
Common mistake: retry blindly because no answer appeared on the phone. A missing answer is not proof that no call ran. held, estimated, and settled in NativeMeter are internal accounting states; even a settled row is not a provider invoice. Keep the task cap shared across aliases and retries rather than resetting it when the model changes.
Lesson 4 — keep implementation boundaries visible
Goal: identify which layer is responsible before debugging or claiming an integration works.
Exact commands are local source reads; they do not execute a plan or call a model:
sed -n '1,100p' scripts/tvision.py
sed -n '1,100p' ai_os/tvision/README.md
rg -n "runtime\(|Harness|subprocess.Popen|transport" scripts/dshrem_cli.py scripts/dshrem.py
Expected result: AIOS Python code owns the inspected DSHREM task setup, grants, ledger checks, and receipts; the CLI starts a paired Node executable with the installed JavaScript Harness package as a child. Tvision is a Python entry point and package that validates or compiles a constrained plan. Its documentation says dispatch executes and transport accounts for paid calls, while the current origin/main dispatch integration is not confirmed. JavaScript/TypeScript source in the DSHREM Telegram client or other product services is a separate application surface, not evidence that a swarm scheduler exists. The Tvision specification recommends Go for a possible future validator/adapter, but the observed implementation is Python; Go implementation is not established.
Common mistake: assume the language name proves ownership: “Harness is Node, so it pays”; “Tvision is a plan, so it runs”; or “Go is recommended, so the compiler is Go.” Trace each layer to source and a dated receipt.
Check yourself
Without running a worker, write down alias/provider, one root cap, one stop condition, the fields needed to reconcile a timeout, and the source for each price. Can you tell a subscription charge from API spend? Subscription quota is accounted separately; it must never silently become a paid API fallback. The course sequence starts at phone terminal practice and owner worktree practice. Compare methods for cost per accepted task and AI judging.
When this does not work
If dshrem models makes a network request or fails, do not keep probing; use the dated local lane inventory and record that live metadata was unavailable. If a receipt lacks a provider response ID or cost, keep it unresolved and stop further paid calls for that task. If the ledger is frozen on uncertain spend, reconcile through its established owner; do not create a parallel ledger or manually mark the charge zero. If your installed version lacks economy or ROOT, capture local help/version and use the reviewed installation procedure; do not bypass a hash guard.
Evidence
Evidence date: 8 November 2026 UTC. The read-only lane source dated 7 November 2026 records the alias/provider declarations and cautions that present runtime health and invoices are not proven. It reports an existing USD 2 operator default and a mismatch with an older policy value; neither is permission for this task, whose total paid API cap is USD 0.20. The worker draft dated 8 November 2026 adds that 2026-11-07 NativeMeter ledger rows were historical and not invoice reconciliation; reservation rows absent from that snapshot mean NOT_VERIFIED, not zero. The same draft lists the official peak windows, Kyiv DST instruction, and DeepInfra promotion versus full reservation bounds. No paid call, quota probe, or live provider check was run for this course.
Next
Start with phone terminal practice to learn how to reconnect without assuming a job outcome. Then complete owner worktree practice before any worker receives file-editing scope. For the execution boundary, read Tvision task language and its explicit statement that validation and planning do not execute or spend.
Sources
docs/runbooks/dshrem-lanes-20261107.md— dated alias declarations, price provenance, route limits, and ledger caveats.scripts/dshrem_subscription.pyandscripts/subscription_watchdog.py— subscription-first routing and explicit fallback gate.DSHREM-MOBILE-ECONOMY-20261108/glm53f-result— API documentation worker draft; not a runtime test.- Source verification and cost per accepted task — evidence and accounting concepts.
The economy profile: one root budget and one event
Goal: inspect selection without a model turn. Run dshrem economy plan --purpose coding --budget 0.20. Expected output is a JSON route with role, requested model, provider, API/subscription, selection reason, peak/off-peak, a conservative-price reservation forecast, estimated cost and root remainder. The forecast is not an actual hold; NativeWire computes the real upper bound from serialized request bytes and the actual output limit before NativeMeter reserves it. The plan checks account quota without inference. It does not dispatch a worker or release the operator STOP.
For a separately authorized exercise, create the lifetime budget once in the clean worktree, then keep its root ID across model changes:
DSHREM_ROOT_ID=$(dshrem economy authorize --budget 0.20 | python3 -c 'import json,sys; print(json.load(sys.stdin)["root_id"])')
dshrem economy plan --root-id "$DSHREM_ROOT_ID" --purpose documentation
printf '%s\n' 'Explain this project briefly; do not edit files.' | dshrem economy run --root-id "$DSHREM_ROOT_ID" --event-id lesson-001 --purpose documentation --read-only
The last line is a real model task and may spend API money. Do not run it as a free health check. Expected result: a retained event, controller reservation, worker result, NativeMeter usage and independent acceptance pending. Repeating the same event ID never replays the task. If a request is in flight, its outcome is unknown or money remains held, reconcile before another route. The next eligible request can choose another verified API route or subscription Luna/Terra. Subscription exhaustion stops; it never enables another paid route silently. Writable subscription work requires a separate clean worktree; read-only work stays read-only.
Typical error: interpreting a failed response as permission to rerun a file edit or publication. Preserve the checkpoint, outputs and root budget. A new model does not create a new allowance. Requested model and confirmed served model are separate fields; absent response evidence, served model is unknown.
Observed 2026-11-08 UTC in this continuation: ds41f wrote a mobile layout helper and glm53f drafted lessons, with three provider-accounted requests total. NativeMeter recorded USD 0.004317898, held 0 and pending 0; invoice reconciliation is not claimed. Luna produced the full course through the account CLI, Terra independently reviewed code, and Sol integrates and installs. API and subscription use are recorded separately. AIOS supplies orchestration and metering in Python; the terminal client and Harness run JavaScript/Node, the editor surfaces use TypeScript, and the portal includes Go services. Tvision currently validates/compiles constrained plans in Python; it is not an independent executor or a Go runtime. The separate Twision source directory is not evidence of an installed compatible Tvision runtime.
2026 holiday exceptions are included from the verified State Council calendar; see the provider’s official peak/off-peak terms. Unknown future calendar years are labelled unverified and use a conservative weekday bound. An unmeasured latency is unknown, not a claimed measurement. Account-only continuation has API ceiling zero and does not mint another root budget.
🔗 Article connections
Open the full knowledge graph ↗Published article links. Select a node to open its article at the heading.
Show the connection graph
Article links · keyboard / text view
Same topic
- Tvijo agent system: PRIZMA, RUNLEDGER and acceptance
- Read agent communication receipts without mistaking delivery for a human read
- Audit an AI technology radar against primary sources before promoting a story
- Run an AI coding harness evaluation offline before any live pilot
- DSHREM on a phone: SSH, tmux, and safe task handoff
- DSHREM owner practice: cwd, worktrees, WIP, and root boundaries
Next steps
Related reading
Tvijo agent system: PRIZMA, RUNLEDGER and acceptance
Tvijo is developing a controlled workflow from a repository task to an accepted result, with permissions, resource limits and attributable evidence. PRIZMA describes the intended resource profile; RUNLEDGER connects the work to its checks a…
AI and process foundationRead agent communication receipts without mistaking delivery for a human read
Find the current stage of a swarm message without confusing durable storage, delivery, runtime receipt and viewing. This guide connects local checks, the private observation screen and distinct receipts, so a delivery acknowledgement does not become an unsupported claim that an action completed.
AI and process foundationAudit an AI technology radar against primary sources before promoting a story
Check a technology news item against its primary source, version and event date before deciding to adopt it. This guide follows a fixed ten-item sample, separates facts from interpretation and forecasts, and preserves limitations and the condition for checking the claim again.
💬 Comments