compressed-context 0.2.0 → 0.2.1

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "compressed-context",
-  "version": "0.2.0",
+  "version": "0.2.1",
   "description": "OpenCode 上下文管理插件：按需压缩、自包含摘要",
   "type": "module",
   "license": "MIT",

package/src/compression/summarize.ts

@@ -32,12 +32,11 @@ function promptRules(options: SummaryPromptOptions | undefined): string[] {
   const lines =
     options?.format === "babel"
       ? [
-          "- Receiver is another equally capable language model, not a human. Drop human readability and grammar; freely mix languages, emoji, math/logic symbols; replace connectives, modifiers, and long clauses with symbols.",
-          "- Only `T=<short title>|G=<one-line goal>` is fixed. Do not use a fixed key schema or long labels (`Meta:`, `Entity:`, `Decisions:`, `Progress:`). Invent the shortest one-character anchors or delimiters that fit this source and use them consistently.",
-          "- Body is a compact latent index, not prose: no headings, no paragraphs, no sentences, no JSON, no code fences.",
-          "- Preserve exactly in latent space: entity/relation graph; exact numbers, versions, config keys, thresholds, formulas; conditions, causality, exceptions; comparisons; original placeholders, paths, commands, signatures, error text. Never invent or round.",
-          "- Keep the latest objective, constraints, decisions with reasons, outcomes, open items, and work the user still expects. Drop narrative, process, logs, repeated reads, and duplicates.",
-          "- Target 400-800 tokens for T1; drop explanations before exact facts.",
+          "- Audience=peer LLM. Notation only, not content: drop grammar/connectives, mix languages/emoji/math（forall、exists、in、=>）.",
+          "- Only `T=<short title>|G=<one-line goal>` fixed; invent anchors and reuse them; no headings/paragraphs/sentences/JSON/fences.",
+          "- Anchors: `S[module]` `D[decision]:why` `F[path:line]` `C[step>step]` `Config[target]:k=v` `?cond=>action` `!edge:detail` `Eval[A/B]:result`; flatten parallel items `[A,B,C]`.",
+          "- Files `F:[...]` complete, never \"various files\"; commands `C:[...]` verbatim, never paraphrase.",
+          "- Lossless: entity graph, exact numbers (no rounding), versions, config keys, thresholds, formulas, conditions, causality, exceptions, comparisons, placeholders, paths, signatures, error text; unknown=?/NULL, never invent.",
           `- Keep this summary under ${maxChars.toString()} characters.`,
         ]
       : options?.format === "markdown"
@@ -101,11 +100,25 @@ export function sessionGenerator(session: SessionGenerateLike, sessionID: string
 }
 
 const BABEL_T1_RULES = `Guidance:
-- This checkpoint replaces the messages above; write a compact latent index, not a story.
-- Only T/G is fixed. Dynamically invent the shortest anchors/delimiters for this source; do not reuse a fixed field schema or long labels.
-- Keep only what future work needs: latest objective, hard constraints, decisions with one-line reasons, outcomes, open items, and exact technical facts.
-- Preserve exact paths, line numbers, signatures, commands, error text, versions, config keys, thresholds, and measured values.
-- Drop process, logs, repeated reads, full diffs, abandoned details, and explanation prose. If space is tight, keep exact facts before narrative.`
+- This checkpoint replaces the messages above; the next reader sees it instead of them, so it must stand on its own. Describe what already happened as history, not as instructions to act now.
+- Keep the objective and hard constraints, decisions with their reasons, outcomes, open items, work the user still expects, and failure lessons.
+- Copy exact identifiers instead of describing them: full paths with line numbers, the line of code that carries the logic, signatures, error text, versions, config keys, thresholds, measured values.
+- Omit build/test logs, repeated reads, exploration that produced nothing, abandoned attempts (one line on why they failed), repeated status checks, and settled discussion.
+- If space is tight, keep in this order: objective and constraints; decisions and reasons; exact technical facts; conclusions; lessons from failures.`
+
+const BABEL_T2_RULES = `Guidance for merging existing checkpoints:
+- The sources are already compressed; merge them into one checkpoint covering all of them, without re-telling the process.
+- Keep the latest objective and hard constraints, decisions with their reasons, final outcomes, open items, blocked work, and the user's still-open expectations.
+- Merge duplicate facts and keep the latest status; mark superseded content with one short marker and drop its details.
+- Drop process, logs, full diffs, code listings, and non-load-bearing line numbers. A failed attempt may stay as one line on why it failed.
+- Target roughly 50-150 tokens per source block.`
+
+const BABEL_T3_RULES = `Guidance for ultra-condensing existing checkpoints:
+- Reduce the sources to a lookup index: one to three single-line facts per source block, subject plus outcome.
+- Keep only shipped outcomes, open work, architectural decisions, and critical constraints.
+- Drop explanations, rationale details, process, logs, and anything already obsolete or superseded.
+- Open user-requested objectives survive until a later source marks them completed or superseded.
+- Target roughly 30-60 tokens per source block.`
 
 const TIER1_RULES = `Guidance:
 - This checkpoint replaces the messages above; the next reader sees it instead of them, so it must stand on its own.
@@ -133,7 +146,7 @@ const TIER3_RULES = `Guidance for ultra-condensing existing checkpoints:
 - Target roughly 30-60 tokens per source block.`
 
 function rulesFor(tier: Tier, format?: SummaryFormat): string {
-  if (format === "babel") return tier === 1 ? BABEL_T1_RULES : tier === 2 ? TIER2_RULES : TIER3_RULES
+  if (format === "babel") return tier === 1 ? BABEL_T1_RULES : tier === 2 ? BABEL_T2_RULES : BABEL_T3_RULES
   if (tier === 2) return TIER2_RULES
   if (tier === 3) return TIER3_RULES
   return TIER1_RULES
@@ -230,13 +243,12 @@ const MARKDOWN_CHOICE_RULES = `Guidance for choosing and merging:
 
 const BABEL_CHOICE_RULES = `Guidance for choosing and merging:
 - Exactly one adjacent pair may be merged. Pick one candidate from the list below and merge ONLY that pair; every other block is context for the decision, not part of the merge.
-- First line must be \`c=N\`, where N is the chosen candidate number.
-- Second line must be \`T=<short title>|G=<one-line goal>\`. After that, write adaptive Babel: only T/G is fixed; dynamically invent the shortest anchors/delimiters and use them consistently. Do not use a fixed field schema or long labels.
-- Mix languages, emoji, and logic symbols freely for density; no Markdown headings, no JSON, no prose/sentences, no code fences.
-- Target 150-300 tokens for T2 and 60-120 tokens for T3; drop explanations before exact facts.
-- Keep the latest objective and hard constraints, decisions with reasons, final outcomes, open items, blocked work, and exact facts that still matter.
+- First line must be \`c=N\`, where N is the chosen candidate number. Second line must be \`T=<short title>|G=<one-line goal>\`; after that write adaptive Babel with anchors you invent and reuse.
+- Babel changes only notation, not content: same policy as a normal merge checkpoint. No Markdown headings, no JSON, no prose/sentences, no code fences.
+- Keep the latest objective and hard constraints, decisions with their reasons, final outcomes, open items, blocked work, and exact facts that still matter.
 - Merge duplicate facts across the chosen pair, keep the latest status, and drop process, logs, repeated reads, full diffs, and non-load-bearing line numbers.
-- Copy identifiers, paths, commands, errors, versions, and numbers verbatim. Do not invent facts.`
+- If the chosen candidate's output tier is 2, target roughly 50-150 tokens per source block. If it is tier 3, be ultra-condensed: one to three single-line facts per source block.
+- Do not invent facts.`
 
 function buildChoicePrompt(
   context: string,