pi-condense 2.0.0 → 2.0.1

package/CHANGELOG.md

@@ -1,73 +1,81 @@
 # Changelog
 
-Format follows sibling pi packages (e.g. [`pi-cohort`](https://github.com/jjuraszek/pi-cohort/blob/main/CHANGELOG.md)):
-one entry per `vX.Y.Z` tag, newest first, terse bullets, dated.
+Keep-a-Changelog style (`## [X.Y.Z] - <date>`, newest first), matching sibling
+pi packages (e.g. [`pi-cohort`](https://github.com/jjuraszek/pi-cohort/blob/main/CHANGELOG.md)).
 
 Published to npm as [`pi-condense`](https://www.npmjs.com/package/pi-condense) (`pi install npm:pi-condense`).
 Pushing a `vX.Y.Z` tag triggers `.github/workflows/release.yml`, which runs the tests and
 publishes via OIDC trusted publishing. See `.agents/skills/release/SKILL.md`.
 
-## v2.0.0 — 2026-07-02
+## [Unreleased]
+
+## [2.0.1] - 2026-07-03
+
+- **Prune summaries are now hidden from Pi's main window** (`display: false` at both injection sites in `index.ts`). They stay in LLM context and session history (recoverable via `context_tree_query`) but no longer print the full markdown block into the TUI. Mirrors upstream `pi-context-prune` `2fd6127`.
+- **Sharpened the summarizer prompt** (`src/summarizer.ts`): the key-outcome bullet now requires copying file paths, identifiers, signatures, and error strings *verbatim* (never reworded), and the summarizer skips tool calls that succeeded with nothing reusable to record. Scoped to short tokens - values/full output stay summarized, so the ~8x median compression and the oversized guard are unaffected.
+- **`context_tree_query` description now shows a literal call example** (`{ toolCallIds: ["t12", "t3"] }`) and consolidates the two near-duplicate parameter descriptions into one that marks the field required. Reduces the empty-first-call retry seen in real sessions. No schema or behavior change.
+
+## [2.0.0] - 2026-07-02
 
 - **Renamed the package `pi-context-prune` -> `pi-condense`** and switched distribution from git-tag pins to npm. Install with `pi install npm:pi-condense` (was `git:github.com/jjuraszek/pi-context-prune@vX.Y.Z`). Migrate pinned `settings.json` entries; `release.sh sync-presets` reports stale pins.
 - **Added npm release machinery:** `.github/workflows/release.yml` (tag-triggered, OIDC + provenance, `tag == package.json` gate) and `test.yml` (bun, Ubuntu + Windows matrix). Ported `release.sh` to the shared sibling skeleton (`propose` / `patch|minor|major` / `current` / `verify` / `sync-presets`), replacing the old tag-pin `release.mjs`.
 - **Added** `LICENSE` (MIT), `files` allowlist, `engines`, and `author` to `package.json`; removed dead `pi.skills` / `pi.prompts` manifest entries (those directories never existed).
 - **Renamed** the cost-event producer id `EXTERNAL_COST_SOURCE` `"pi-context-prune"` -> `"pi-condense"`. Aggregators keyed on the old `source` string see it as a new producer.
 
-## v1.0.0 — 2026-05-31
+## [1.0.0] - 2026-05-31
 
 - **Removed three `pruneOn` modes**, leaving `agent-message` (default) and `on-demand`:
-  - `every-turn` — debugging-only trigger with the worst prompt-cache churn.
-  - `on-context-tag` — depended on the external `ttttmr/pi-context` extension and overlapped its `context_compact`.
-  - `agentic-auto` — the scaffolded DCP-style model-driven `context_prune` tool was never wired to range compression (see `PRUNING.md § Future Work`).
+  - `every-turn` - debugging-only trigger with the worst prompt-cache churn.
+  - `on-context-tag` - depended on the external `ttttmr/pi-context` extension and overlapped its `context_compact`.
+  - `agentic-auto` - the scaffolded DCP-style model-driven `context_prune` tool was never wired to range compression (see `PRUNING.md`, Future Work).
 - **Removed** the `context_prune` tool, the agentic-auto system prompt, the `<pruner-note>` unpruned-count reminder, and the `remindUnprunedCount` setting. Deleted `src/reminder.ts`, `src/context-prune-tool.ts`, `src/progress-text.ts`.
 - **Migration:** none required. Configs pinned to a removed mode fall back to `agent-message` via `isPruneOn()`. A stale `remindUnprunedCount` key in `settings.json` is ignored.
 
-## v0.11.1 — 2026-05-28
+## [0.11.1] - 2026-05-28
 
 - **Release flow:** `release.sh` now rewrites every `~/.pi/agent*/settings.json` pin of `git:github.com/jjuraszek/pi-context-prune@<ref>` to the new `@vX.Y.Z` automatically after pushing the tag. Opt out with `--no-update-pins`. Aligns this fork's release workflow with sibling pi-* packages.
 - **Docs:** `README.md` install section leads with the jjuraszek tag-pin (was upstream npm/sha references). `AGENTS.md` release blurb updated to reflect tag pins + automatic settings rewrite. `.agents/skills/release/SKILL.md` documents the new flow + flags. Adds this `CHANGELOG.md` for parity with sibling pi-* packages.
 
-## v0.11.0 — 2026-05-28
+## [0.11.0] - 2026-05-28
 
-- **Pre-flush pipeline:** content-hash dedup (re-reads of identical `(toolName, content)` pairs alias the original instead of going through the LLM), trivial-batch skip (`minBatchChars`), protected tools allowlist, stub-replace rather than delete. See `PRUNING.md § Pre-flush Pipeline & Safeguards`.
+- **Pre-flush pipeline:** content-hash dedup (re-reads of identical `(toolName, content)` pairs alias the original instead of going through the LLM), trivial-batch skip (`minBatchChars`), protected tools allowlist, stub-replace rather than delete. See `PRUNING.md`, Pre-flush Pipeline & Safeguards.
 - **Settings:** moved to `<agent-dir>/settings.json#contextPrune` namespace (was a separate file). Honors `$PI_CODING_AGENT_DIR`.
 
-## v0.10.0 — 2026-05-11
+## [0.10.0] - 2026-05-11
 
 - `quietOversizedSkips` setting to suppress `skipped-oversized` notifications.
 - Demote oversized-skip notification severity to info.
 - Use short refs in pruned summaries (e.g. `t1`, `t2` rather than full toolCallIds) so the model can pass them back through `context_tree_query` more reliably.
 
-## v0.9.x — 2026-05-05 → 2026-05-11
+## [0.9.x] - 2026-05-05 to 2026-05-11
 
 - `0.9.3`: spinner animation fix for `/pruner now`.
 - `0.9.2`: replace footer progress with aboveEditor widget during `/pruner now`.
 - `0.9.1`: allow `ESC` to cancel `context_prune` tool call.
 - `0.9.0`: agentic-auto mode (`pruneOn: "agentic-auto"`), `context_prune` tool surfaced to the LLM, `remindUnprunedCount` setting.
 
-## v0.8.x — 2026-05-04 → 2026-05-05
+## [0.8.x] - 2026-05-04 to 2026-05-05
 
 - `0.8.1`: bug fixes around session-start index rebuild.
 - `0.8.0`: `agent-message` trigger mode + batching, footer status widget.
 
-## v0.7.0 — 2026-05-04
+## [0.7.0] - 2026-05-04
 
 - `on-context-tag` trigger mode (integrates with `pi-context` `context_checkpoint`).
 
-## v0.6.x — 2026-05-02
+## [0.6.x] - 2026-05-02
 
 - Tree browser (`/pruner tree`) + summary overlay (`Ctrl-O`).
 - `dedupByContentHash` cross-flush dedup.
 
-## v0.5.0 — 2026-05-01
+## [0.5.0] - 2026-05-01
 
 - Cumulative summarizer token/cost stats (`/pruner stats`).
 
-## v0.4.0 — 2026-05-01
+## [0.4.0] - 2026-05-01
 
 - Configurable summarizer model + thinking level (`/pruner model`, `/pruner thinking`).
 
-## Earlier (v0.1.x – v0.3.x)
+## Earlier (v0.1.x - v0.3.x)
 
 Initial extension scaffolding, `context_tree_query` tool, base summarization loop, session-JSONL index persistence. See `git log` for granular history.

package/index.ts

@@ -194,7 +194,7 @@ export default function (pi: ExtensionAPI) {
 
     const appendEntry = (customType: string, data?: unknown) => sessionManager!.appendCustomEntry(customType, data);
     const appendSummaryMessage = (content: string, details: unknown) =>
-      sessionManager!.appendCustomMessageEntry(CUSTOM_TYPE_SUMMARY, content, true, details);
+      sessionManager!.appendCustomMessageEntry(CUSTOM_TYPE_SUMMARY, content, false, details);
 
     // Routes alias persistence through whichever delivery is active so the
     // dedup pre-flush pass writes CUSTOM_TYPE_DEDUP_ALIAS entries via the
@@ -400,11 +400,14 @@ export default function (pi: ExtensionAPI) {
 
         try {
           if (!shouldSkipOversized) {
-            // Write one summary message per turn and index its tool calls.
+            // Write one hidden summary message per turn and index its tool calls.
+            // `display: false` keeps the summary in future LLM context (convertToLlm
+            // ignores `display`) while suppressing the full markdown block from Pi's
+            // main window; rebuild keys on customType, not display.
             const batchToolCallIds = batch.toolCalls.map((tc) => tc.toolCallId);
             if (delivery === "runtime") {
               pi.sendMessage(
-                { customType: CUSTOM_TYPE_SUMMARY, content: summaryText, display: true, details: batchDetails },
+                { customType: CUSTOM_TYPE_SUMMARY, content: summaryText, display: false, details: batchDetails },
                 { deliverAs: "steer" }
               );
               indexer.registerSummaryRefs(summaryRefs);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-condense",
-  "version": "2.0.0",
+  "version": "2.0.1",
   "description": "Pi extension that summarizes completed tool-call batches, replaces raw outputs with short stubs in future context, and recovers any original on demand via context_tree_query.",
   "author": "Jacek Juraszek",
   "license": "MIT",

package/src/chain-range-prune.test.ts

@@ -50,7 +50,7 @@ function summaryMsg(timestamp: number, toolCallIds: string[]): any {
     role: "custom",
     customType: "context-prune-summary",
     content: "summary text",
-    display: true,
+    display: false,
     details: { toolCallRefs: toolCallIds.map((id, i) => ({ shortId: `t${i + 1}`, toolCallId: id })) },
     timestamp,
   };

package/src/query-tool.ts

@@ -9,14 +9,14 @@ export function registerQueryTool(pi: ExtensionAPI, indexer: ToolCallIndexer): v
     name: "context_tree_query",
     label: "Query Original Tool History",
     description:
-      "Retrieve original tool call results that have been pruned from active context. Pass the short refs from a pruner-summary message to get back the full original outputs.",
+      "Retrieve original tool call results that have been pruned from active context. Pass the short refs listed in a pruner-summary message, e.g. context_tree_query({ toolCallIds: [\"t12\", \"t3\"] }), to get back the full original outputs.",
     promptSnippet: "Retrieve original pruned tool outputs by short ref",
     promptGuidelines: [
       "When you need the full output of a tool call that was summarized and pruned from context, use context_tree_query with the short refs listed in the relevant pruner-summary message.",
     ],
     parameters: Type.Object({
-      toolCallIds: Type.Array(Type.String({ description: "One or more short refs or tool call IDs to retrieve" }), {
-        description: "List of short refs or toolCallIds to look up",
+      toolCallIds: Type.Array(Type.String(), {
+        description: 'Required. One or more short refs (e.g. "t12") or raw tool call IDs from a pruner-summary message.',
       }),
     }),
 

package/src/summarizer.ts

@@ -14,10 +14,10 @@ import { serializeBatchForSummarizer } from "./batch-capture.js";
 const SYSTEM_PROMPT = `You are summarizing a batch of tool calls made by an AI coding assistant.
 For each tool call provide:
 - Tool name and a one-sentence description of what it did
-- Key outcome: success/failure and the most important data returned
+- Key outcome, plus any file paths, identifiers, signatures, or error strings copied verbatim - never reword these
 - Any findings the future conversation needs to remember
 
-Keep each tool call to 1-3 bullet points. Be concise.`;
+Keep each tool call to 1-3 bullet points. Skip calls that succeeded with nothing reusable to record. Be concise.`;
 
 const RANGE_SYSTEM_PROMPT = `You are fusing several per-step summaries of one CLOSED sub-task from an AI coding assistant's history into a SINGLE cohesive summary.
 - Merge overlapping or repeated information; do not restate each step separately.