v0.21.0 — the risk one tool cannot show you, and the money a refund does not cover

.claude-plugin/marketplace.json

@@ -12,7 +12,7 @@
       "displayName": "Agent Stack",
       "source": "./plugins/agent-stack",
       "description": "Four skills: agent-orchestrator — tool-calling loops, pipelines with checkpoints, provider routing with fallback, memory architecture, plus the wallet side of reselling LLM access; agent-evals — run/trace/thread evals, LLM judges, and fixtures grown from production; agent-interop — MCP servers and clients, A2A agent cards, the MCP Registry, and gateways; agent-harness — system prompts, tool shaping, workflow-vs-agent, and auditing an agent system.",
-      "version": "0.20.0",
+      "version": "0.21.0",
       "author": {
         "name": "ssheleg",
         "url": "https://x.com/sshlg93"

CHANGELOG.md

@@ -1,3 +1,41 @@
+## v0.21.0 — the risk one tool cannot show you, and the money an iteration refund does not cover
+
+Two findings, both of them about a rule that is right on one axis and silently assumed to
+cover a second.
+
+**Tool annotations are a risk vocabulary, and their defaults are asymmetric on purpose.**
+`readOnlyHint` **false**, `destructiveHint` **true**, `idempotentHint` **false**,
+`openWorldHint` **true** — so a server author who omits annotations entirely has declared
+the *most dangerous* shape. That is the correct fail-closed choice and the opposite of what
+most authors think they are doing: annotation **narrows** an assumption, silence widens it.
+And every one is a hint rather than a contract — the specification says a client must treat
+them as untrusted unless the server is, so an annotation may inform a UI or a policy default
+and may never be the thing that decides whether a destructive call runs.
+
+**The lethal trifecta now has a name here.** Private data, untrusted content, and the
+ability to communicate externally: any two are safe, all three in one session are an
+exfiltration path that no prompt-level instruction reliably closes. It lands in
+`agent-harness/references/tools.md` rather than in a permission section for a structural
+reason — **it is a property of the tool set assembled in a session, so per-tool analysis
+cannot see it by construction.** Every tool can pass its own review and the combination
+still be unsafe; adding a `fetch(url)` beside a private-data reader closes the triangle and
+will not look like a security change. `agent-interop/references/gateway.md` gains the
+composer's half: a gateway's whole value is assembling many servers into one surface, which
+is exactly the operation that can create a risk none of its inputs had, so **granting a role
+one more server is a trifecta question, not only a least-privilege one.**
+
+**Refund the iteration, charge the money.** Our loop refunds an iteration on a recoverable
+provider error — a 502 should not consume one of the ten attempts. That is right, and it
+says nothing about money: the provider still billed the call. A response that arrived and
+then failed to parse was generated, metered and charged. The observed failure is precise —
+a harness charged its cost on the **success path** of `query()`, so a format error meant the
+cost was never added, and the code compensated inside the exception handler. Where a cost
+ceiling is the *primary* bound (that harness ships `cost_limit = 3.0` with the step limit
+**off**), a leak there is a leak in the only guard. Accounting belongs in a `finally`, and
+**a budget that under-counts is worse than one that over-counts**: over-counting stops a run
+early and visibly, under-counting is invisible until the invoice and biases toward the
+failing runs, which are the expensive ones.
+
 ## v0.20.0 — what a trajectory cannot carry across a vendor, and where the capability goes
 
 Three findings from the harvest, all landing on the same question: **which model does what,

docs/evidence/backlog.md

@@ -34,6 +34,8 @@ row per line, appended, so two runs closing two rows do not land on the same lin
 | B-126 | **The pack had zero coverage of the prompt cache — the harvest's most corroborated gap, 14 findings from 9 independent bundles** — and SKILL.md §10 told the reader to rebuild the prompt per request without saying what that costs. Separately, four of the compaction ladder's five rungs were labelled `free` when every one of them edits history and invalidates the prefix from its replacement point. | 2026-08-31 harness-engineering harvest, cluster #1 by computed priority | 3 | 0 | 2 | **1.5** | **closed 2026-08-31, shipped in v0.19.0** — `references/kv-cache.md`: the directional rule, the 100k-conversations timestamp incident (TTFT 0.5s→3–5s, bill nearly doubled), the 2^N cache populations a runtime condition in front of the boundary creates, the named invalidators including the role-switch and hot-plugin trades, and four optimisations that pay cache for nothing. §10 corrected; the ladder gained a Cache column. Headroom bought by **displacement** — the body was at 4749/4750 and the Data Verification Protocol moved to where §10's own text sends prompt content. **The ceiling was enforced in CI and invisible locally**: the *House skill audit* job failed this very release on the headroom GAP, while `npm test` measured nothing — `validate.py` now hard-fails at 5000 and reports at 4750 on every local run, watched failing on a planted 1,200-char body. The §2 listing was displaced to `patterns.md`, taking the body from 4749/4750 to **4609/4750**.  |
 | B-128 | **§6's three routing traps all assume the REQUEST is what moves, and a mid-run failover moves a trajectory.** Reasoning is portable text plus a non-portable vendor credential — sometimes attached to the tool call rather than the reasoning — so the safest-sounding policy, strip all reasoning before failing over, is the one that 400s. Separately, nothing said where model capability should go: the planner is the bottleneck, not the executor. And a max-iteration guard stops a spend without changing the behaviour that led there. | 2026-08-31 harness-engineering harvest, queue positions 3-4 by computed priority (both ARCH) | 2 | 0 | 2 | **1.0** | **closed 2026-08-31, shipped in v0.20.0** — `llm-proxy-billing.md` gains three sections: what a trajectory cannot carry across a vendor (with the neutral-format rule and the between-turns failover boundary), where the capability goes (Plan-and-Act arXiv:2503.09572, 54% on WebArena-Lite from the planner), and budget awareness (300 steps performing like 30). §6 gains a fourth trap naming the assumption the other three share; §2 gains the budget line beside the iteration refund. |
 | B-129 | **`agent-orchestrator`'s SKILL.md body is structurally full: two releases running, an addition has cost a displacement.** v0.19.0 displaced the §2 loop listing to buy room for the cache correction (4749 → 4609); v0.20.0 displaced §1's context dataclass and base class for three routing lines that would otherwise have left 5 tokens (→ 4631). The house auditor states the remedy in its own GAP text — *the answer then is a split, not a trim* — and each pass has instead paid with someone else's listing. | 2026-08-31, observed across v0.19.0 and v0.20.0 | 1 | 0 | 3 | **0.3** | **open.** The displacements were each defensible on their own (a listing belongs in a reference, the body carries what is read every time), but the pattern is a body at capacity rather than three good edits. A split means deciding which of the 13 numbered sections are one skill and which are another — a routing decision, not a formatting one, so it is filed rather than improvised at the end of an unrelated release. |
+| B-130 | **Annotations were mentioned only as untrusted input, with no hint, no default and no name for the risk a single tool cannot express.** `tools.md:152` said descriptions and annotations are untrusted and stopped there; nothing stated that omitting them declares the most dangerous shape, and the family had **no word for the lethal trifecta** — a property of the tool set in a session that per-tool review cannot see by construction. | 2026-08-31 harness-engineering harvest, queue position 5 (ARCH); corroborated independently by the permissions bundle | 2 | 0 | 1 | **2.0** | **closed 2026-08-31, shipped in v0.21.0** — `tools.md` gains the four hints with their asymmetric defaults, the hint-not-a-contract rule, and the trifecta with the reason it lives beside tool design rather than in a permission check; a seventh trap names per-tool review as unable to see it. `gateway.md` gains the composer's half: granting a role one more server is a trifecta question, not only least-privilege. |
+| B-131 | **We shipped `iteration refund` and nothing said the money is not refunded with it.** A recoverable provider error does not consume an attempt — correct — but the provider still billed the call, and a response that arrived and then failed to parse was generated, metered and charged. Cost accounting written on the success path under-counts on exactly the runs that go worst. | 2026-08-31 harness-engineering harvest, queue position 6 (ARCH) | 2 | 0 | 1 | **2.0** | **closed 2026-08-31, shipped in v0.21.0** — `llm-proxy-billing.md` gains *Refund the iteration, charge the money* with the observed failure (cost added after a successful parse, compensated inside the exception handler, in a harness whose `cost_limit` is the primary bound with the step limit off), the `finally` rule, and why under-counting is the worse asymmetry. §2's bullet now carries the pairing in one clause. |
 
 ## Open, and why
 

docs/evidence/verification.md

@@ -10,6 +10,20 @@ This file exists because its absence read as zero exposure. `sshlg-skills` board
 
 ---
 
+## Shipped state — v0.21.0 (2026-08-31)
+
+Measured on the release-candidate tree before the tag exists. Queue positions 5–6, both ARCH.
+
+| REQ | What ships | How it was confirmed | Confirmed |
+|---|---|---|---|
+| TR-1 | The four annotation hints and their asymmetric defaults are stated, with the consequence | `tools.md` → *Annotations, and the risk one tool cannot show you* carries the table (`readOnlyHint` false, `destructiveHint` true, `idempotentHint` false, `openWorldHint` true) and the reading that follows: an unannotated tool has declared the most dangerous shape | **observed** |
+| TR-2 | A hint may inform a default and may never decide a destructive call | the same section states the specification's untrusted-unless-the-server-is rule as a limit on what an annotation is allowed to gate | **observed** |
+| TR-3 | The lethal trifecta has a name and lives where per-tool review can be shown not to reach it | `grep -c 'lethal trifecta' tools.md` → **2**; a seventh entry in *Traps* names per-tool review as unable to see it by construction; `gateway.md` gains the composer's half before its own *Traps* | **observed** |
+| TR-4 | The iteration refund no longer implies a money refund | `llm-proxy-billing.md` → *Refund the iteration, charge the money* with the observed success-path accounting failure and the `finally` rule; SKILL.md §2's bullet carries the pairing in one clause; `grep -c refund` across the skill still resolves to one home for the mechanism | **observed** |
+| TR-5 | The body stayed inside the working limit without a third displacement | auditor `0 GAP, 14 PASS`, body **4655/4750** — the two additions were written into references and one clause, deliberately, because B-129 records that this body is at capacity | **observed** |
+
+---
+
 ## Shipped state — v0.20.0 (2026-08-31)
 
 Measured on the release-candidate tree before the tag exists. Queue positions 3–4 of the

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ssheleg/agent-stack",
-  "version": "0.20.0",
+  "version": "0.21.0",
   "scripts": {
     "test": "python3 test/validate.py && python3 test/plant_guard_test.py && node test/installer_test.js"
   },

plugins/agent-stack/.claude-plugin/plugin.json

@@ -3,7 +3,7 @@
   "name": "agent-stack",
   "displayName": "Agent Stack",
   "description": "Four skills: agent-orchestrator — tool-calling loops, pipelines with checkpoints, provider routing with fallback, memory architecture, plus the wallet side of reselling LLM access; agent-evals — run/trace/thread evals, LLM judges, and fixtures grown from production; agent-interop — MCP servers and clients, A2A agent cards, the MCP Registry, and gateways; agent-harness — system prompts, tool shaping, workflow-vs-agent, and auditing an agent system.",
-  "version": "0.20.0",
+  "version": "0.21.0",
   "author": {
     "name": "ssheleg",
     "url": "https://x.com/sshlg93"

plugins/agent-stack/skills/agent-harness/references/tools.md

@@ -126,6 +126,48 @@ the mistake cannot be made**, rather than documenting the mistake.
 - A required `confirm: true` on a destructive action, so a partially-formed call fails
   closed.
 
+## Annotations, and the risk one tool cannot show you
+
+MCP tools carry four hints, and their defaults are asymmetric on purpose:
+
+| Hint | Default | So an unannotated tool is assumed to be |
+|---|---|---|
+| `readOnlyHint` | `false` | one that writes |
+| `destructiveHint` | `true` | destructive |
+| `idempotentHint` | `false` | unsafe to repeat |
+| `openWorldHint` | `true` | reaching outside your system |
+
+**A server author who omits annotations entirely has declared the most dangerous shape**,
+which is the correct fail-closed choice and the opposite of what most authors assume they
+are doing. Annotate to *narrow* the assumption; silence widens it.
+
+**Every one is a hint, not a contract.** The specification is explicit that a client must
+treat descriptions and annotations as untrusted unless the server itself is trusted — so an
+annotation informs a UI and a policy default, and may never be the thing that decides
+whether a destructive call runs.
+
+### The lethal trifecta — a property of the session, not of a tool
+
+Three capabilities that are individually ordinary and jointly an exfiltration path:
+
+1. access to **private data**,
+2. exposure to **untrusted content**,
+3. the ability to **communicate externally**.
+
+Any two are safe. All three in one session mean untrusted content can instruct the agent to
+read private data and send it out, and no prompt-level instruction reliably prevents it.
+
+**The reason it belongs here rather than in a permission check:** the trifecta is a property
+of *the tool set assembled in a session*, so **per-tool analysis cannot see it**. Every tool
+can pass its own review and the combination still be unsafe — which is why a review that
+walks a server's tools one at a time answers a different question than "what can this
+session do". Look at the set, and at what a gateway composes into it.
+
+**The practical consequence for a tool author:** a tool that only reads private data is
+fine; adding a `fetch(url)` beside it is what closes the triangle, and it will not look like
+a security change in review.
+
+
 ## Evaluating tools
 
 Tools deserve **thorough documentation and testing**, and testing means running the agent
@@ -150,5 +192,7 @@ Fix the interface, not the prompt, when the fault is in this table.
   can do nothing with.
 - **Treating tool output as trusted.** It is attacker-controlled input if the server is; the
   specification says descriptions and annotations are untrusted unless the server is.
+- **Reviewing tools one at a time.** The lethal trifecta is a property of the assembled set;
+  a per-tool review cannot see it by construction.
 - **Adding a tool to fix a prompt problem.** The set grows, selection degrades, and the
   original defect is still there.

plugins/agent-stack/skills/agent-interop/references/gateway.md

@@ -147,6 +147,25 @@ authorization** — the last being the per-tool control that a generic gateway l
 That is the strongest single argument for it in a cluster that already runs Gateway API: the
 routing objects are ones your platform team already reviews.
 
+## The gateway is where the lethal trifecta gets assembled
+
+A gateway's whole value is composing many servers into one surface, and that is exactly the
+operation that can create a risk none of its inputs had. The **lethal trifecta** — private
+data, untrusted content, and the ability to communicate externally — is a property of the
+**tool set in a session**, so every upstream can pass its own review and the composition
+still be unsafe.
+
+Two consequences for whoever runs the gateway:
+
+- **A role is a trifecta decision, not only a least-privilege one.** Granting a role one
+  more server is the moment to ask which of the three corners it just completed — not
+  whether that server is individually trustworthy.
+- **Per-tool annotations do not answer it.** They are hints about one tool (`readOnlyHint`,
+  `destructiveHint`, `idempotentHint`, `openWorldHint`, and their defaults assume the
+  dangerous shape — `agent-harness/references/tools.md`), and no combination of per-tool
+  hints computes a session-level property.
+
+
 ## Traps
 
 - **Introducing a gateway and leaving the direct paths open.** The policy is then advisory.

plugins/agent-stack/skills/agent-orchestrator/SKILL.md

@@ -86,7 +86,8 @@ was gathered rather than returning nothing. The full listing is in
 - **In-loop trimming**: At ~80% capacity, collapse older assistant+tool pairs into one-liner summaries
 - **Token limit recovery**: On `LLMTokenLimitError`, compress to 60% and retry once. If still fails, return partial answer
 - **Max iterations guard**: Always have a hard limit. On exhaustion, compose best-effort answer from data gathered so far
-- **Iteration refund**: a recoverable provider error is not charged to that guard
+- **Iteration refund**: a recoverable provider error is not charged to that guard — the
+  money still is, and accounting on the success path alone under-counts the worst runs
 - **Budget awareness**: tell the model what is left, or 300 steps performs like 30
 
 ---

plugins/agent-stack/skills/agent-orchestrator/references/llm-proxy-billing.md

@@ -24,6 +24,7 @@ the patterns hold for any upstream that issues per-tenant keys with limits.
 - [Discovering spend you do not control](#discovering-spend-you-do-not-control)
 - [Guardrails: budgets, loops, auto-pause](#guardrails-budgets-loops-auto-pause)
 - [Key lifecycle and healing](#key-lifecycle-and-healing)
+- [Refund the iteration, charge the money](#refund-the-iteration-charge-the-money)
 - [The refund waterfall](#the-refund-waterfall)
 - [Model routing and fallbacks](#model-routing-and-fallbacks)
 
@@ -226,6 +227,41 @@ handled three different ways.
 
 ---
 
+## Refund the iteration, charge the money
+
+These are two axes and it is easy to ship one rule for both.
+
+`agent-orchestrator`'s loop **refunds the iteration** on a recoverable provider error: a
+502 should not consume one of the ten attempts the agent has to finish its work. That is
+right, and it says nothing about money.
+
+**The provider still billed the call.** A response that arrived and then failed to parse
+was generated, metered and charged upstream; so was the one that arrived truncated, and the
+one whose tool call was malformed. The tempting symmetry — *the attempt did not count, so
+it did not cost* — is how a spend guard under-counts on exactly the runs that go worst.
+
+The failure mode is specific and worth naming, because it hides where nobody looks:
+
+> A harness charged `self.cost` on the **success path** of its `query()`. When parsing the
+> response raised a format error, `query()` never returned, so the cost was never added —
+> and the code compensated for it explicitly inside the exception handler. Where a cost
+> ceiling is the *primary* bound — that harness ships `cost_limit = 3.0` with the step
+> limit **off** — a leak in that accounting is a leak in the only guard there is.
+
+The rule, in one line each:
+
+- **Iteration:** refunded on a recoverable provider error, never on a misconfiguration.
+- **Money:** charged whenever the provider generated tokens, including on every path that
+  raises after the response arrived.
+- **Therefore:** accounting belongs in a `finally`, or in the exception handler as well as
+  the success path. If the only place your cost is added is the line after a successful
+  parse, the guard is quietly optimistic.
+
+A budget that under-counts is worse than one that over-counts: over-counting stops a run
+early and is visible immediately; under-counting is invisible until the invoice, and it
+biases toward the failing runs, which are the expensive ones.
+
+
 ## The refund waterfall
 
 A payment refund has to come out of somewhere, and the money has usually moved.