log: model identity on transform lines

devlog/2026-08-17_log-model-identity/REQ.md

@@ -0,0 +1,16 @@
+# REQ — log: model identity on transform lines
+
+## Background
+
+Analyzing a user log (omp issue #47 family, 2026-08-17): the compress-rejection loop could not be attributed to a model from the log alone — required environment forensics. All sibling adapters (omp #78, pi #156, proxy #164) are adding model identity in the same sweep.
+
+## Requirements
+
+- R1: The per-turn debug line (`transform-in`) must carry the model identity when the host provides it (`event.model`).
+- R2: Format: `providerID/model.id` (matches the plugin's own model-limit resolution key). `null` when absent (V1 path has no model ref in scope).
+- R3: Debug-only (existing `debug()` gate), no behavior change, no new dependency.
+
+## Acceptance
+
+- AC1: `transform-in` debug line includes `model` field on the V2 context path.
+- AC2: typecheck clean; full suite green (54/54).

devlog/2026-08-17_log-model-identity/WORKLOG.md

@@ -0,0 +1,3 @@
+# WORKLOG
+
+- 2026-08-17: Added `model: event.model ? providerID/id : null` to the `transform-in` debug line in `runPipelineV2` (packages/billion-context-opencode/src/index.ts). V1 path untouched (no model ref in that scope; the V1 line already lacks per-model resolution). tsc clean, 54/54.

packages/billion-context-opencode/src/index.ts

@@ -253,7 +253,9 @@ async function runPipelineV2(
   const coveredIds = collectCoveredMessageIds(state)
   const tokenCount = estimateTokens(cores, coveredIds)
   const resolved = runtime.configFor(runtime.getModelLimit(sessionID))
-  debug("transform-in", { sid: sessionID, msgs: msgs.length, cores: cores.length, tokens: tokenCount, limit: resolved.modelContextLimit, blocks: state.blocks.length })
+  // Model identity on every transform line: weak-model compress-rejection
+  // loops must be attributable from logs alone (omp #78 family).
+  debug("transform-in", { sid: sessionID, model: event.model ? `${event.model.providerID}/${event.model.id}` : null, msgs: msgs.length, cores: cores.length, tokens: tokenCount, limit: resolved.modelContextLimit, blocks: state.blocks.length })
 
   const turn = runtime.core.processTurn({
     messages: cores,