v0.20.0 — what a trajectory cannot carry across a vendor, and where the capability goes

.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": "0.19.1",
+      "version": "0.20.0",
       "author": {
         "name": "ssheleg",
         "url": "https://x.com/sshlg93"

CHANGELOG.md

@@ -1,3 +1,39 @@
+## 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,
+and what survives when that changes mid-run.** §6 shipped three traps about routing, and all
+three quietly assume the *request* is what moves.
+
+- **A trajectory carries a vendor credential, and it may not be attached to the reasoning.**
+  Tool calls and results are portable — different structure, same meaning, re-render and
+  send. Reasoning is portable *text* plus a **non-portable credential** the vendor attaches
+  to prove the reasoning is its own, and vendors disagree on what they demand: one end
+  validates nothing, the other rejects any credential it did not issue. The credential
+  sometimes sits on the **tool call** rather than the reasoning — which is why *"just strip
+  all reasoning before failing over"*, the policy that sounds safest, is the one that
+  produces a 400. Store trajectories in a neutral internal format, keep the text, discard
+  the credential, re-render per vendor at send time, and put the failover boundary
+  **between turns**. A fallback chain never exercised mid-trajectory has not been tested: a
+  green health probe answers a question about the endpoint, not about your history.
+- **Capability is not spent evenly — the planner is the bottleneck.** *Plan-and-Act*
+  (arXiv:2503.09572) found that with good enough planning a relatively simple executor
+  suffices, and with a wrong decomposition every downstream executor is building on a false
+  premise; their 54% on WebArena-Lite came from improving the **planner**, not the executor.
+  So the strongest model and the most carefully written prompt go to the **manager**. It
+  also says where to look when a multi-agent system underperforms: **a weak plan is
+  invisible in every executor's transcript**, because each one did its own step correctly.
+- **Steps the agent cannot see buy nothing.** Standard agents have no budget awareness, so
+  at **300 steps** they still plateau at roughly what they achieve at **30**. A
+  max-iteration guard is the floor of this rather than the mechanism — it stops the spend
+  and never changes the behaviour that led there.
+
+**And the second displacement in two releases.** The body was at 4609/4750 after v0.19.0
+bought that headroom back; these three lines would have left **5 tokens**. §1's context
+dataclass and sub-agent base class moved to `references/patterns.md`, beside the loop
+listing that went there in v0.19.0, landing the body at **4631/4750**. That is now twice in
+a row that an addition has cost a displacement, which is the auditor's own signal — *the
+answer then is a split, not a trim* — and it is filed rather than absorbed again.
+
 ## v0.19.1 — the class the umbrella had been catching for us, twice
 
 `B-126`'s board row shipped in v0.19.0 with **nine cells against the eight its header

docs/evidence/backlog.md

@@ -32,6 +32,8 @@ row per line, appended, so two runs closing two rows do not land on the same lin
 | B-124 | **`agent-evals` had no `references/` directory and no statistics layer: 315 lines saying what to assert, nothing saying how many times.** A skill whose job is answering *did it get better* could not distinguish a result from noise — a grep over SKILL.md for the vocabulary of repetition and variance returned **one** hit across 315 lines, and it was about escalation. Separately, §5's *"Right tools, right order"* is the exact practice Anthropic calls *"too rigid"*. | 2026-08-31 harness-engineering harvest; the statistics gap was reported independently by **four** of the twenty agents | 3 | 0 | 2 | **1.5** | **closed 2026-08-31, shipped in v0.18.0** — `references/statistics.md`: the noise band with its table, `pass@k` vs `pass^k` and the 91-point gap, the τ-bench data showing trials are correlated so `pass^k` cannot be exponentiated, paired comparison with 3–5 seeds, the harness as a 6 pp variable with a floor AND a ceiling, and the evidence ladder ending on *4/4 on a slice is not 100% system-wide*. Every figure recomputed before it was written; the correlation finding is ours, from recomputing rather than restating. §5 moved **between** its two measured edges — Anthropic's *too rigid* and Claw-Eval's 44% of safety violations missed by trajectory-blind grading — rather than flipping to the other extreme. |
 | B-118 (umbrella) | **The committed social card clipped its eyebrow at the canvas, and the check on it counted bytes.** The eyebrow is generated from the umbrella's `role` cell for this pack — 91 characters — and at the smallest scale the renderer will use it needs **1354px of a 1200px canvas**, so the last **11** characters were never drawn: the card read *"…AND THE WALLET UNDER"* and dropped `" LLM RESALE"`. `test/social_preview.py` checked the PNG signature, the chunk order, the byte count and the 1200x630 dimensions — none of which is a reading about text — so it was green over the clipped card for as long as one existed. | umbrella board row B-118, measured by its v1.11.0 run 2026-08-31; re-measured here before acting | 3 | 0 | 2 | **1.5** | **closed 2026-08-31, shipped in v0.18.2** — the finding was re-derived rather than restated: decoding the committed PNG put accent ink at **x=1199 of 1200**, and replaying the renderer's own metric gave a pen finishing at x=1354 with 11 characters past the edge. `fitScale` was already at its floor of 2, so no scale fixes it and the text had to shorten. The new role string is 60 characters — `orchestration, prompts, evals, protocols, and the LLM wallet` — which names all four skills where the old one named three, and paints to **x=919, 196px (19%) inside the content box**. The sensor is the durable half: `test/card_ink.py` decodes the image and requires every painted pixel to sit inside the content box, with the one deliberate exception (the top-left accent bar) required to be **present**, so a blank or mis-decoded image cannot pass by having nothing in the gutters. Watched failing on the previously committed card — *"accent paints at x=1199, y=190 — 84px past the right edge of the content box … and reaches the canvas edge, so characters were cut off entirely"* — while the old check printed `OK` on that same file. A negative self-test in `validate.yml` plants one pixel in the last column of a repo copy and requires the entry point to refuse it, with `plant_guard.py verify` confirming the plant landed. The decoder lives in its own module deliberately: `test/social_preview.py` is shared by nine repositories and the umbrella stops comparing copies below 0.90 similarity **for all nine at once**, so the shared file gained two lines and measures **0.9133** against its eight siblings. Umbrella-side follow-up is the coordinator's and is named in the ledger. |
 | 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. |
 
 ## Open, and why
 

docs/evidence/verification.md

@@ -10,6 +10,21 @@ This file exists because its absence read as zero exposure. `sshlg-skills` board
 
 ---
 
+## Shipped state — v0.20.0 (2026-08-31)
+
+Measured on the release-candidate tree before the tag exists. Queue positions 3–4 of the
+2026-08-31 harvest, both ARCH class, plus the budget finding that shares their subject.
+
+| REQ | What ships | How it was confirmed | Confirmed |
+|---|---|---|---|
+| MR-1 | §6 names the assumption its three existing traps share, and the reference carries the mechanism | the new fourth bullet states that the three above assume the *request* is portable; `llm-proxy-billing.md` → *What a trajectory cannot carry across a vendor* has the credential mechanism, the tool-call attachment that makes *strip all reasoning* the failing policy, the neutral-format rule and the between-turns boundary | **observed** |
+| MR-2 | Capability allocation is stated, with its source | `references/llm-proxy-billing.md` → *Where the capability goes* cites Plan-and-Act (arXiv:2503.09572) and its 54% on WebArena-Lite from improving the planner; §6 carries the one-line rule | **observed** |
+| MR-3 | Budget awareness sits beside the guard it corrects | §2's bullet list gains it after the iteration refund; the reference explains why a max-iteration guard is the floor and not the mechanism (300 steps performing like 30) | **observed** |
+| MR-4 | The body stayed inside the working limit, and the headroom was bought rather than borrowed | the three lines alone left **4745/4750 — 5 tokens**; §1's context dataclass and sub-agent base class were displaced to `references/patterns.md`, landing at **4631/4750**. Auditor: `0 GAP, 14 PASS` | **observed** |
+| MR-5 | The table-shape guard shipped in v0.19.1 caught its author, on its first use by someone other than its plant | both new board rows were inserted after the last row of *any* table and landed in the three-column *Open, and why* — `python3 test/validate.py` refused with `row has 8 cells against the 3 its own header declares`, twice. **Third occurrence of the class and the first caught by this repository rather than the umbrella**, which is exactly what it was written for | **watched failing** |
+
+---
+
 ## Shipped state — v0.19.1 (2026-08-31)
 
 Measured on the release-candidate tree before the tag exists. A class caught by the umbrella twice

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ssheleg/agent-stack",
-  "version": "0.19.1",
+  "version": "0.20.0",
   "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": "0.19.1",
+  "version": "0.20.0",
   "author": {
     "name": "ssheleg",
     "url": "https://x.com/sshlg93"

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

@@ -40,23 +40,12 @@ OrchestratorAgent.run(AgentContext)
 
 ### Shared Context Object
 
-Pass a single immutable-ish context object to all sub-agents:
+Pass a single immutable-ish context object to every sub-agent. It carries the request
+(`project_id`, `user_question`, `chat_history`), the machinery (`llm_router`, `tracker`,
+`workflow_id`), the resolved provider and model, and one `extra` dict for pipeline flags.
+The full dataclass is in [`references/patterns.md`](references/patterns.md) → *The
+orchestrator's shared context and sub-agent protocol*.
 
-```python
-@dataclass
-class AgentContext:
-    project_id: str
-    user_question: str
-    chat_history: list[Message]
-    llm_router: LLMRouter           # provider abstraction with retry/fallback
-    tracker: WorkflowTracker        # SSE event emitter for real-time UI
-    workflow_id: str                 # unique ID for this request
-    connection_config: ... | None   # external resource config
-    user_id: str | None
-    preferred_provider: str | None  # e.g. "openrouter"
-    model: str | None               # e.g. "<provider>/<model-id>"
-    extra: dict[str, Any]           # pipeline_action, flags, overrides
-```
 
 **Key principles:**
 - Sub-agents never modify context — they return typed results
@@ -65,20 +54,10 @@ class AgentContext:
 
 ### Sub-Agent Protocol
 
-Every sub-agent extends a base class:
-
-```python
-class BaseAgent(ABC):
-    @abstractmethod
-    async def run(self, context: AgentContext, **kwargs) -> AgentResult: ...
-
-    @property
-    @abstractmethod
-    def name(self) -> str: ...
+Every sub-agent extends one base class with a single abstract `run(context) -> Result`,
+so the orchestrator never learns what any of them does internally. The class is in the same
+reference section.
 
-    @staticmethod
-    def accum_usage(total, usage): ...  # merge token counters
-```
 
 Typed result subclasses per agent (e.g. `SQLAgentResult` with `query`, `results`, `attempts`).
 
@@ -108,6 +87,7 @@ was gathered rather than returning nothing. The full listing is in
 - **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
+- **Budget awareness**: tell the model what is left, or 300 steps performs like 30
 
 ---
 
@@ -197,6 +177,11 @@ fallback chain and per-provider retry with exponential backoff, and its
 - **Model selection has three levels** — the request, the tenant, the system
   default — and a tenant override that silently loses to a request parameter is
   how a cheap model ends up billed at a premium one's rate.
+- **Those three assume the REQUEST is portable; the trajectory is not.** Reasoning carries
+  a vendor credential — sometimes on the tool call — so mid-turn failover can 400, and
+  *strip all reasoning* is what causes it. Fail over between turns.
+- **Capability is not spent evenly**: the planner is the bottleneck, so the strongest model
+  goes to the manager, not to whichever agent does the most work.
 ## 7. Multi-Layer Memory System
 
 Four layers, each with a different lifetime and a different reason to exist: