cache trace mvp

Added a standalone cache tracing module and wired it into the embedded
runner so you can capture message flow and the exact context sent to
  Anthropic in a separate JSONL file.

  What changed

  - New tracing module: src/agents/cache-trace.ts (self‑contained,
env‑gated, writes JSONL, computes per‑message digests).
  - Hook points in src/agents/pi-embedded-runner/run/attempt.ts: logs
stage snapshots (loaded/sanitized/limited/prompt/stream/after) and wraps
the
    stream fn to record the real context.messages at send time.

  How to enable

  - CLAWDBOT_CACHE_TRACE=1 enables tracing.
  - CLAWDBOT_CACHE_TRACE_FILE=~/.clawdbot/logs/cache-trace.jsonl
overrides output (default is
$CLAWDBOT_STATE_DIR/logs/cache-trace.jsonl).
  - Optional filters:
      - CLAWDBOT_CACHE_TRACE_MESSAGES=0 to omit full messages (still
logs digests).
      - CLAWDBOT_CACHE_TRACE_PROMPT=0 to omit prompt text.
      - CLAWDBOT_CACHE_TRACE_SYSTEM=0 to omit system prompt.

  What you’ll see

  - One JSON object per line with stage, messagesDigest, per‑message
messageFingerprints, and the actual messages if enabled.
  - The most important line is stage: "stream:context" — that is the
exact payload pi‑mono is sending. If this diverges from earlier stages,
you’ve
    found the mutation point.
This commit is contained in:
Andrii
2026-01-21 10:03:24 +01:00
committed by Peter Steinberger
parent 63d017c3af
commit 5392fa0dfa
2 changed files with 303 additions and 0 deletions

View File

@@ -16,6 +16,7 @@ import { normalizeMessageChannel } from "../../../utils/message-channel.js";
import { isReasoningTagProvider } from "../../../utils/provider-utils.js";
import { isSubagentSessionKey } from "../../../routing/session-key.js";
import { resolveUserPath } from "../../../utils.js";
import { createCacheTrace } from "../../cache-trace.js";
import { resolveClawdbotAgentDir } from "../../agent-paths.js";
import { resolveSessionAgentIds } from "../../agent-scope.js";
import { makeBootstrapWarn, resolveBootstrapContextForRun } from "../../bootstrap-files.js";
@@ -367,6 +368,17 @@ export async function runEmbeddedAttempt(
throw new Error("Embedded agent session missing");
}
const activeSession = session;
const cacheTrace = createCacheTrace({
cfg: params.config,
env: process.env,
runId: params.runId,
sessionId: activeSession.sessionId,
sessionKey: params.sessionKey,
provider: params.provider,
modelId: params.modelId,
modelApi: params.model.api,
workspaceDir: params.workspaceDir,
});
// Force a stable streamFn reference so vitest can reliably mock @mariozechner/pi-ai.
activeSession.agent.streamFn = streamSimple;
@@ -379,6 +391,15 @@ export async function runEmbeddedAttempt(
params.streamParams,
);
if (cacheTrace) {
cacheTrace.recordStage("session:loaded", {
messages: activeSession.messages,
system: systemPrompt,
note: "after session create",
});
activeSession.agent.streamFn = cacheTrace.wrapStreamFn(activeSession.agent.streamFn);
}
try {
const prior = await sanitizeSessionHistory({
messages: activeSession.messages,
@@ -388,12 +409,14 @@ export async function runEmbeddedAttempt(
sessionManager,
sessionId: params.sessionId,
});
cacheTrace?.recordStage("session:sanitized", { messages: prior });
const validatedGemini = validateGeminiTurns(prior);
const validated = validateAnthropicTurns(validatedGemini);
const limited = limitHistoryTurns(
validated,
getDmHistoryLimitFromSessionKey(params.sessionKey, params.config),
);
cacheTrace?.recordStage("session:limited", { messages: limited });
if (limited.length > 0) {
activeSession.agent.replaceMessages(limited);
}
@@ -564,6 +587,10 @@ export async function runEmbeddedAttempt(
}
log.debug(`embedded run prompt start: runId=${params.runId} sessionId=${params.sessionId}`);
cacheTrace?.recordStage("prompt:before", {
prompt: effectivePrompt,
messages: activeSession.messages,
});
// Repair orphaned trailing user messages so new prompts don't violate role ordering.
const leafEntry = sessionManager.getLeafEntry();
@@ -633,6 +660,12 @@ export async function runEmbeddedAttempt(
}
}
cacheTrace?.recordStage("prompt:images", {
prompt: effectivePrompt,
messages: activeSession.messages,
note: `images: prompt=${imageResult.images.length} history=${imageResult.historyImagesByIndex.size}`,
});
// Only pass images option if there are actually images to pass
// This avoids potential issues with models that don't expect the images parameter
if (imageResult.images.length > 0) {
@@ -660,6 +693,10 @@ export async function runEmbeddedAttempt(
messagesSnapshot = activeSession.messages.slice();
sessionIdUsed = activeSession.sessionId;
cacheTrace?.recordStage("session:after", {
messages: messagesSnapshot,
note: promptError ? "prompt error" : undefined,
});
// Run agent_end hooks to allow plugins to analyze the conversation
// This is fire-and-forget, so we don't await