integrate: Discord-sourced fixes (#286–#298)

AGENTS.md

@@ -1862,8 +1862,8 @@ The probe is **lazy / turn-boundary driven** — `pickProvider()` reads `Date.no
 
 A fallover can cross transports — the common `cloud (native_tools) → local (grammar)` default (`appendLocal`) does exactly that. Both the request shape and the response parse are decoupled from the primary:
 
-- **Request.** Each attempt re-resolves `{ transport, adapter }` for the chosen link via `resolveActiveLlmSlice(providerId)`, so the wire shape is correct for whoever serves. `buildLlmStreamParams` keeps `grammar` **populated even on the native path** (it used to blank it) so a grammar-only link handed the request still has its GBNF; native providers ignore `grammar` and read `tools`, so carrying both is safe.
-- **Response.** The completion is stamped with `servedTransport` — the transport of the link that actually answered — by the fallback seams in [src/runtime/llm-fallback-seam.ts](src/runtime/llm-fallback-seam.ts) (`createFallbackCompleter` / `createFallbackStreamer`). `step-executor.parseDepsFor(completion, deps)` prefers `servedTransport` over the caller's configured `toolTransport` for every parse decision (`tryParseToolCalls`, the empty-completion recovery gate). Without this, a native primary that fell over to a grammar link parsed the grammar reply as OpenAI `tool_calls` and silently broke tool-calling. The stamp is pinned directly on the real seam factories by [src/runtime/llm-fallback-seam.test.ts](src/runtime/llm-fallback-seam.test.ts) (deleting either stamp turns it red) and end-to-end through the loop by [src/llm/fallback/fallback-e2e.integration.test.ts](src/llm/fallback/fallback-e2e.integration.test.ts).
+- **Request.** Each attempt re-resolves `{ transport, adapter }` for the chosen link via `resolveActiveLlmSlice(providerId)`, so the wire shape is correct for whoever serves. `buildLlmStreamParams` keeps `grammar` **populated even on the native path** (it used to blank it) so a grammar-only link handed the request still has its GBNF; native providers ignore `grammar` and read `tools`, so carrying both is safe. On a think-tag profile the prompt shape is per-link too: the main prompt for a native-tools primary is built **prefill-suppressed** (issue #283 — a literal `<think>` shipped to a chat endpoint is at best noise, at worst corrupted server-side), and `LlmStreamParams.grammarPrompt` carries a lazy, memoized prefill-carrying variant that the seams substitute for grammar links (`promptFor`), so a llama-server link still gets the shape its template + GBNF prelude expect. The one-shot repair retry rebuilds both variants repair-shaped.
+- **Response.** The completion is stamped with `servedTransport` — the transport of the link that actually answered — by the fallback seams in [src/runtime/llm-fallback-seam.ts](src/runtime/llm-fallback-seam.ts) (`createFallbackCompleter` / `createFallbackStreamer`). `step-executor.parseDepsFor(completion, deps)` prefers `servedTransport` over the caller's configured `toolTransport` for every parse decision (`tryParseToolCalls`, the empty-completion recovery gate). Without this, a native primary that fell over to a grammar link parsed the grammar reply as OpenAI `tool_calls` and silently broke tool-calling. The streamer additionally stamps `servedTransport` on **every chunk** — the return-value stamp only exists after the last delta, which is too late for the live stream parser: `consumeStream` creates its parser lazily off the first chunk's stamp so a grammar-served stream (whose GBNF output starts mid-`<think>`) keeps emitting live `reasoning_delta`s under a native primary. `completionAssumesOpenReasoning` keys purely off the served/parse transport: grammar-served output always continues an open think block; a chat completion never does (so even in the unsupported grammar-primary → native-link ordering a clean chat reply is no longer swallowed whole as reasoning). The stamps are pinned directly on the real seam factories by [src/runtime/llm-fallback-seam.test.ts](src/runtime/llm-fallback-seam.test.ts) (deleting either stamp turns it red) and end-to-end through the loop by [src/llm/fallback/fallback-e2e.integration.test.ts](src/llm/fallback/fallback-e2e.integration.test.ts), including think-tag profile cases in both directions of the parse decision.
 
 The remaining asymmetry: `tools` is populated only when the **primary's** transport is `native_tools`. Placing a native-tools provider **below** a grammar-only primary would reach it without a `tools` payload — an unusual ordering; order native-tools links at or above the first grammar-only link. Slot affinity is decided pre-request from the primary, so a cloud→local fallover runs the local link without slot-cache reuse for that turn (correctness-neutral). The grammar string itself is always built for the primary model.
 
@@ -1876,7 +1876,7 @@ The remaining asymmetry: `tools` is populated only when the **primary's** transp
 5. **Whole-chain exhaustion rethrows the last (already-humanized) error** so `loop_failed` classification is unchanged. Pinned by [src/llm/fallback/run-with-fallback.test.ts](src/llm/fallback/run-with-fallback.test.ts).
 6. **Exactly one switch notice per state transition** (away / back), none on sticky turns. Pinned by [src/llm/fallback/provider-fallback-chain.test.ts](src/llm/fallback/provider-fallback-chain.test.ts).
 7. **`appendLocal` appends the local provider when configured, nothing when not.** Pinned by [src/llm/fallback/fallback-config.test.ts](src/llm/fallback/fallback-config.test.ts).
-8. **A cross-transport fallover parses the response with the served link's transport, not the primary's**, and the turn reaches the fallback's answer instead of `loop_failed`. Pinned by [src/llm/fallback/fallback-e2e.integration.test.ts](src/llm/fallback/fallback-e2e.integration.test.ts) (real `AgentLoop` + `step-executor`, both unary and streaming).
+8. **A cross-transport fallover parses the response with the served link's transport, not the primary's**, and the turn reaches the fallback's answer instead of `loop_failed`. On a think-tag profile the grammar link also receives the prefill-carrying `grammarPrompt` variant and its streamed reasoning stays classified live (per-chunk `servedTransport` stamp). Pinned by [src/llm/fallback/fallback-e2e.integration.test.ts](src/llm/fallback/fallback-e2e.integration.test.ts) (real `AgentLoop` + `step-executor` + the real seam factories; both unary and streaming, plain and think-tag profiles).
 9. **Breaker state is partitioned by session** — one partition's success does not clear another's armed cooldown, and a keyless call shares one default partition. Pinned by [src/llm/fallback/provider-fallback-chain.test.ts](src/llm/fallback/provider-fallback-chain.test.ts) ("partition isolation").
 10. **The cooldown ladder must be non-decreasing** — a decreasing `cooldownMs` is rejected at parse time so "escalating" stays true. Pinned by [src/config/llm-config.test.ts](src/config/llm-config.test.ts).
 

MEMORY.md

@@ -2,6 +2,7 @@
 
 This document is the source-of-truth for how cross-session memory works in `atomic-agent`. It complements:
 
+- `MEMORY_GUIDE.md` — the operator-facing walkthrough with worked examples; start there if you want to *use* memory rather than modify it.
 - `ARCHITECTURE.md` — overall runtime topology and invariants.
 - `AGENTS.md` — short engineering summary for automated contributors.
 - `PROMPT.md` — full anatomy of the stable prefix and variable tail, including where the memory channels render in the prompt.