separate prompt responsibilities between system and tool specs

Author: Tarquinen

lib/prompts/compress-message.ts

@@ -1,10 +1,5 @@
 export const COMPRESS_MESSAGE = `Collapse selected individual messages in the conversation into detailed summaries.
 
-THE PHILOSOPHY OF MESSAGE COMPRESS
-\`compress\` in message mode transforms specific stale messages into dense, high-fidelity summaries. This is not cleanup - it is crystallization. Your summary becomes the authoritative record of what each selected message contributed.
-
-Think of compression as phase transitions: raw exploration becomes refined understanding. The original message served its purpose; your summary now carries that understanding forward.
-
 THE SUMMARY
 Your summary must be EXHAUSTIVE. Capture file paths, function signatures, decisions made, constraints discovered, key findings, tool outcomes, and user intent details that matter... EVERYTHING that preserves the value of the selected message after the raw message is removed.
 
@@ -20,6 +15,7 @@ You specify individual raw messages by ID using the injected IDs visible in the
 - \`mNNNN\` IDs identify raw messages
 
 Each message has an ID inside XML metadata tags like \`<dcp-message-id priority="high">m0007</dcp-message-id>\`.
+The ID tag appears at the end of the message it belongs to — it identifies the message above it, not the one below it.
 Treat these tags as message metadata only, not as content to summarize. Use only the inner \`mNNNN\` value as the \`messageId\`.
 The \`priority\` attribute indicates relative context cost. Prefer higher-priority closed messages before lower-priority ones.
 
@@ -32,23 +28,10 @@ Rules:
 - Do not invent IDs. Use only IDs that are present in context.
 - Do not target prior compressed blocks or block summaries.
 
-THE WAYS OF MESSAGE COMPRESS
-Compress when an individual message is genuinely closed and unlikely to be needed verbatim again:
-
-Research findings have already been absorbed into later work
-Tool-heavy assistant updates are no longer needed in raw form
-Earlier planning or analysis messages are now stale but still important to retain as summary
-
-Do NOT compress when:
-You may need the exact raw message text, code, or error output in the immediate next steps
-The message is still actively being referenced or edited against
-The target is a prior compressed block or block summary rather than a raw message
-
-Before compressing, ask: _"Is this message closed enough to become summary-only right now?"_
-
 BATCHING
 Select MANY messages in a single tool call when they are independently safe to compress.
 Each entry should summarize exactly one message, and the tool can receive as many entries as needed in one batch.
+When several messages are equally safe to compress, prefer higher-priority messages first.
 
 Because each message is compressed independently:
 

lib/prompts/compress-range.ts

@@ -1,10 +1,5 @@
 export const COMPRESS_RANGE = `Collapse a range in the conversation into a detailed summary.
 
-THE PHILOSOPHY OF COMPRESS
-\`compress\` transforms verbose conversation sequences into dense, high-fidelity summaries. This is not cleanup - it is crystallization. Your summary becomes the authoritative record of what transpired.
-
-Think of compression as phase transitions: raw exploration becomes refined understanding. The original context served its purpose; your summary now carries that understanding forward.
-
 THE SUMMARY
 Your summary must be EXHAUSTIVE. Capture file paths, function signatures, decisions made, constraints discovered, key findings... EVERYTHING that maintains context integrity. This is not a brief note - it is an authoritative record so faithful that the original conversation adds no value.
 
@@ -43,42 +38,23 @@ When you use compressed block placeholders, write the surrounding summary text s
 - Do not write text that depends on the placeholder staying literal (for example, "as noted in \`(b2)\`").
 - Your final meaning must be coherent once each placeholder is replaced with its full compressed block content.
 
-THE WAYS OF COMPRESS
-Compress when a range is genuinely closed and the raw conversation has served its purpose:
-
-Research concluded and findings are clear
-Implementation finished and verified
-Exploration exhausted and patterns understood
-
-Compress smaller ranges when:
-You need to discard dead-end noise without waiting for a whole chapter to close
-You need to preserve key findings from a narrow slice while freeing context quickly
-You can bound a stale range cleanly with injected IDs
-
-Do NOT compress when:
-You may need exact code, error messages, or file contents from the range in the immediate next steps
-Work in that area is still active or likely to resume immediately
-You cannot identify reliable boundaries yet
-
-Before compressing, ask: _"Is this range closed enough to become summary-only right now?"_ Compression is irreversible. The summary replaces everything in the range.
-
 BOUNDARY IDS
 You specify boundaries by ID using the injected IDs visible in the conversation:
 
 - \`mNNNN\` IDs identify raw messages
 - \`bN\` IDs identify previously compressed blocks
 
 Each message has an ID inside XML metadata tags like \`<dcp-message-id>...</dcp-message-id>\`.
+The ID tag appears at the end of the message it belongs to — it identifies the message above it, not the one below it.
 Treat these tags as boundary metadata only, not as tool result content.
 
 Rules:
 
 - Pick \`startId\` and \`endId\` directly from injected IDs in context.
 - IDs must exist in the current visible context.
 - \`startId\` must appear before \`endId\`.
-- Prefer boundaries that produce short, closed ranges.
 - Do not invent IDs. Use only IDs that are present in context.
 
 BATCHING
-Do not call the tool once per range. When multiple independent ranges are ready and their boundaries do not overlap, include all of them as separate entries in the \`content\` array of a single tool call. Each entry should have its own \`startId\`, \`endId\`, and \`summary\`.
+When multiple independent ranges are ready and their boundaries do not overlap, include all of them as separate entries in the \`content\` array of a single tool call. Each entry should have its own \`startId\`, \`endId\`, and \`summary\`.
 `

lib/prompts/system.ts

@@ -5,6 +5,11 @@ The ONLY tool you have for context management is \`compress\`. It replaces older
 
 \`<dcp-message-id>\` and \`<dcp-system-reminder>\` tags are environment-injected metadata. Do not output them.
 
+THE PHILOSOPHY OF COMPRESS
+\`compress\` transforms conversation content into dense, high-fidelity summaries. This is not cleanup - it is crystallization. Your summary becomes the authoritative record of what transpired.
+
+Think of compression as phase transitions: raw exploration becomes refined understanding. The original context served its purpose; your summary now carries that understanding forward.
+
 OPERATING STANCE
 Prefer short, closed, summary-safe compressions.
 When multiple independent stale sections exist, prefer several focused compressions (in parallel when possible) over one broad compression.
@@ -18,12 +23,24 @@ CADENCE, SIGNALS, AND LATENCY
 - Prefer smaller, regular compressions over infrequent massive compressions for better latency and summary quality
 - When multiple independent stale sections are ready, batch compressions in parallel
 
+COMPRESS WHEN
+
+A section is genuinely closed and the raw conversation has served its purpose:
+
+- Research concluded and findings are clear
+- Implementation finished and verified
+- Exploration exhausted and patterns understood
+- Dead-end noise can be discarded without waiting for a whole chapter to close
+
 DO NOT COMPRESS IF
 
-- raw context is still relevant and needed for edits or precise references
-- the target content is still actively in progress
+- Raw context is still relevant and needed for edits or precise references
+- The target content is still actively in progress
+- You may need exact code, error messages, or file contents in the immediate next steps
+
+Before compressing, ask: _"Is this section closed enough to become summary-only right now?"_
 
-Evaluate conversation signal-to-noise REGULARLY. Use \`compress\` deliberately with quality-first summaries. Prefer multiple short, independent compressions before considering broader ones, and prioritize stale content intelligently to maintain a high-signal context window that supports your agency
+Evaluate conversation signal-to-noise REGULARLY. Use \`compress\` deliberately with quality-first summaries. Prioritize stale content intelligently to maintain a high-signal context window that supports your agency.
 
-It is of your responsibility to keep a sharp, high-quality context window for optimal performance
+It is of your responsibility to keep a sharp, high-quality context window for optimal performance.
 `