Stage annex — Mira's two lanes
Stage annex — Mira's two lanes
The plan of record is 02-architecture/PLAN.md (Min, status: authority) — 8
stages, 3 arcs, 14 adopted essences. This annex does not re-plan it. It carries
only what my lanes add: the per-step detail, the traps, and the proofs.
Owner assignment, per Min's §5 division:
| Lane | Owner | Stages |
|---|---|---|
| COST & PROVIDERS | Mira | 6 (cost + retention) |
| SKILLS & TOOLS STORE | Mira | 4 (store steps, with Min's registry steps) |
Young's instruction: "register implementation plan, break it into stages and assign owner. We've live Hermes as our go to sample and references."
So every stage below names the Hermes reference it is checked against — Hermes is our sample, not our dependency.
Stage 6 — Cost and retention
Deliverable: a ledger that disaggregates, and retention as an enforced routing input. Owner: Mira (per Min's §5).
The rule this stage exists to enforce
Min's what-else-to-look-into.md Tier 1: every Tier 1 item is a control that
exists and is not consulted on the path that matters. Stage 6 is the stage where
that rule is mine to satisfy — my retention rule is currently a policy footnote
that no code reads.
| # | Step | Proof | Hermes reference |
|---|---|---|---|
| 6.1 | The ledger: billing_instrument (none · go_allowance · zen_payg), cost_state, window_state | every model call records its instrument | none — Hermes has no GO/ZEN. This is ours |
| 6.2 | E3 — per-requester attribution, spend cap at the boundary | spend attributable per agent; the cap stops the call | A2A_PEER_TOKENS — "nothing in the request body can assert identity" |
| 6.3 | E12 — trace-id threading through the boundary | a call traceable end to end | LiteLLM's require_trace_id_on_calls_to_agent; Hermes' contextId-keyed conversations |
| 6.4 | Retention as a routing input — task classification (does this touch vault content?) | a zero-retention task cannot route to a non-zero-retention model | no Hermes equivalent — ours |
| 6.5 | Window exhaustion ≠ spend exhaustion | GO's windowed allowance and ZEN's PAYG float are distinguishable | no Hermes equivalent — ours |
The proof I would defend hardest
A zero-retention task physically cannot reach a non-zero-retention model — the call fails by declaration, not by policy.
Min adopted this verbatim. It is the only rule in the tree where the cost ladder and the privacy ladder currently agree — the cheapest model is the zero-retention one. Stage 6.4 is what catches the day they disagree.
Three traps, each with a falsifier
| Trap | Falsifier |
|---|---|
| An unpriced run reads as free | inject an unpriced run; the ledger must read cost_state: unknown, never $0.00. V2 recorded 30,721 rows at $0.00000 |
| A prose refusal records as success | a model that refuses in prose must record refusal. V2's classifier called it completed — a cost tool that hides defects is the same failure as a ledger that hides cost |
| Window exhaustion reads as spend | exhaust a window; the ledger must read window_state: blocked while spend is $0 — because the model blocks, not the bill |
🔴 One trap neither plan records — and it sits in Stage 3.1, not mine
a2a-sdk's TaskState is a protobuf enum, and its JSON serialisation is
TASK_STATE_CANCELED — a different string from the lowercase canceled wire
value. The same state, two strings, and only one is the wire value.
So cortex/tasks/states.yaml must record both forms, or Stage 3.1's mapping
encodes the wrong one. It is Min's lane — reported here because my ledger's
window_state axis and his state vocabulary must agree on the spelling.
And a second, cheaper one: TaskState is not iterable —
for s in TaskState raises TypeError. Read TaskState.keys(). A conformance
probe written the obvious way fails for the wrong reason, which is exactly how a
green gets trusted.
Open, and mine
| # | Question | Status |
|---|---|---|
| 6A | Is GO's $60 hard or soft? | ✅ CLOSED — hard, per model. See mira-go-allowance-answer.md. Ladder then ruled in mira-rollover-ladder-ruling.md |
| 6B | Tier-2 ordering: GO before ZEN? | ✅ CLOSED — GO paid → GO free → ZEN. ZEN last. See mira-rollover-ladder-ruling.md. A jump go_allowance → zen_payg skipping GO free is a defect |
| 6C | a2a-sdk 1.2.2, or Hermes' plugin? | ✅ RULED — official SDK. Cost noted: +26 packages. See mira-litellm-a2a-gateway-delta.md |
| 6D | Retention classification input | OPEN — needs a task-classification input nobody has designed. Mine + Min's. My strongest rule; no teeth |
| 6E | billing_instrument carry the model id | ✅ now REQUIRED — Haiku's $3/5h cap makes which model a routing decision the ledger must record |
| 6F | Luna/Haiku at $15/mo = $3 per 5 h | ✅ CLOSED — reframed. Haiku 5.5 is the complex-task escalation, scarce by design. Bulk escalation → GLM/MiMo at $12/5h |
| 6G | Prompt caching as headroom | OPEN, mine — GLM caches reads at $0.03 vs $0.15 input. A 5× reduction cheaper than any model switch, and nothing mentions it |
Stage 4 — Skills and tools store
Deliverable: one source, loaded on demand. Owner: Mira (store) with Min (registry) and Sol (load probes) — per Min's §5.
The load model is settled — adopt, do not design
Min's LEARNED-AND-PLAN.md and PLAN.md Stage 4 adopt it, and Hermes ships the
attern already. Three filter points at three boundaries, and V2 filtered at none
of them — which is why all 67 skills reached every profile:
| Boundary | What it filters | Source |
|---|---|---|
| the card — what a peer can see | advertised toolsets | Hermes — A2A_ADVERTISED_TOOLSETS, card built from the live tool registry |
| the index — what an agent may load | plane / capability | Min's cortex/registry |
| the context — what actually enters | task-driven search + explicit load | Moyai E1 / Hermes progressive disclosure |
| # | Step | Proof | Hermes reference |
|---|---|---|---|
| 4.1 | Declare plane/capability per skill | every skill declares a plane | Hermes skill_utils.get_skill_search_roots() with tiers |
| 4.2 | Generate card skills from cortex/skills | card matches the declared set | the card is built from the registry — same shape |
| 4.3 | Filter the profile projection by plane | a heart agent receives no brain skills | — |
| 4.4 | E1 — search by keywords → ≤5 as metadata only → content only after an explicit load | a search adds no instruction text; a load does | skills_tool_plugin.py — "Anthropic-recommended progressive-disclosure limits" |
| 4.5 | E9 — an agent that cannot load its declared tools refuses to start | a 0-tool agent raises | the same shape as Hermes' own load failures |
| 4.6 | Move the MCP tool list into the registry; declare per-agent permission | a refused tool call fails by declaration | MCP is config-driven for them — the shape, not a licence to build |
| 4.7 | The drift probe — the guard exists in prose, no probe specified | cortex and card disagree → the probe fails | evals/heartbeat_idle_wire.py's header is the contract: real components, fake externals, isolated state, no network |
The two proofs I would not drop
skills: []must be a hard failure, not a valid card. V2 served eight empty cards and read them as healthy. Absence must announce itself.- A search adds no instruction text; a load does. That single assertion is the difference between Moyai's load model and V2's copy-all-67 — everything else in Stage 4 is bookkeeping around it.
Also mine, and not in any plan
A curator. Hermes ships agent/curator.py — "background maintenance for
agent-created skills — usage tracking, staleness, archival, and LLM-driven review."That is the second mechanism, and it is the same shape as Moyai's E7 rule for
memory: capture must not depend on the agent remembering. A store that nothing
prunes is V2's demise in slow motion — 67 skills where a handful are ever loaded.
Recommendation: cortex/skills/ needs a curator step in Stage 4, and the proof
is that a skill never loaded is archived rather than left to rot. Not currently in
any stage.
What I am NOT taking on
- Stage 5 (memory) — flagged unowned by Sol, recommended to Min by Min himself. Still unassigned until Young rules. I am not claiming it.
- Stage 0 and 1 — root guard, prover, registry. Min's and Sol's.
- Stages 2, 3, 7 — transport, task contract, executor. Sol's.
Every stage above is a design until its proof runs. No probe has run in any of
them — every terminal call in my session has failed on a stuck cwd, and no code
exists (find . -name '*.py' returns 0).