@gotgenes/pi-subagents 6.18.2 → 6.18.3

package/CHANGELOG.md

@@ -5,6 +5,21 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [6.18.3](https://github.com/gotgenes/pi-packages/compare/pi-subagents-v6.18.2...pi-subagents-v6.18.3) (2026-05-24)
+
+
+### Performance Improvements
+
+* reorder append-mode prompt for KV cache reuse ([#180](https://github.com/gotgenes/pi-packages/issues/180)) ([5f688bd](https://github.com/gotgenes/pi-packages/commit/5f688bd1d008e20987d28626c5f5d0df0f66b854))
+
+
+### Documentation
+
+* plan reorder append-mode prompt for KV cache reuse ([#180](https://github.com/gotgenes/pi-packages/issues/180)) ([bb0ddec](https://github.com/gotgenes/pi-packages/commit/bb0ddec8a7beb37baace5698e4fa4d09e61497d6))
+* **retro:** add planning stage notes for issue [#180](https://github.com/gotgenes/pi-packages/issues/180) ([3413158](https://github.com/gotgenes/pi-packages/commit/341315898baa09652df18731ad318c89861ec62c))
+* **retro:** add retro notes for issue [#166](https://github.com/gotgenes/pi-packages/issues/166) ([fae30ce](https://github.com/gotgenes/pi-packages/commit/fae30cec3dd99bbac490a2764a8340aa12fc171c))
+* **retro:** add TDD stage notes for issue [#180](https://github.com/gotgenes/pi-packages/issues/180) ([1560f2d](https://github.com/gotgenes/pi-packages/commit/1560f2d6f7029cbbe0cc7b1efe1aba2a243e8357))
+
 ## [6.18.2](https://github.com/gotgenes/pi-packages/compare/pi-subagents-v6.18.1...pi-subagents-v6.18.2) (2026-05-24)
 
 

package/docs/plans/0180-reorder-append-prompt-for-kv-cache.md

@@ -0,0 +1,100 @@
+---
+issue: 180
+issue_title: "perf(pi-subagents): reorder append-mode system prompt to enable KV cache reuse"
+---
+
+# Reorder append-mode system prompt for KV cache reuse
+
+## Problem Statement
+
+In append mode, `buildAgentPrompt()` places varying, agent-specific content (the `<active_agent>` tag and env block) *before* the large shared inherited system prompt (~8k tokens).
+LLM KV caching works on prefixes — the cache is only reusable when the beginning of the prompt matches.
+Every subagent spawn reprocesses the entire inherited prompt from scratch because the prefix differs per agent.
+
+## Goals
+
+- Reorder the append-mode system prompt so shared/stable content comes first and varying content follows.
+- Preserve the `<active_agent>` tag at any position — pi-permission-system's `ACTIVE_AGENT_TAG_REGEX.exec()` searches the full string.
+- Keep replace-mode prompt ordering unchanged (it has no shared inherited content to cache).
+- Update tests and JSDoc to reflect the new ordering.
+
+## Non-Goals
+
+- Changing replace-mode prompt assembly (no shared prefix to cache).
+- Modifying pi-permission-system (its regex parsing is already position-independent).
+- Changing the *content* of any prompt section — only reordering.
+
+## Background
+
+`buildAgentPrompt()` in `src/session/prompts.ts` assembles the system prompt for subagents.
+In append mode, the current ordering is:
+
+```text
+1. <active_agent name="${name}"/>     ← VARIES per agent
+2. # Environment ...                  ← VARIES per runtime
+3. <inherited_system_prompt>          ← SHARED (~8k tokens)
+4. <sub_agent_context>                ← SHARED (static)
+5. <agent_instructions>               ← VARIES per agent
+6. memory / skills                    ← VARIES
+```
+
+pi-permission-system's `getActiveAgentNameFromSystemPrompt()` in `src/active-agent.ts` uses `ACTIVE_AGENT_TAG_REGEX.exec(systemPrompt)` — a regex search that finds the tag at any position, confirmed by reading the source.
+
+## Design Overview
+
+Move shared/stable sections to the front of the append-mode prompt:
+
+```text
+1. <inherited_system_prompt>          ← SHARED (~8k tokens, NOW CACHEABLE)
+2. <sub_agent_context>                ← SHARED (static)
+3. <active_agent name="${name}"/>     ← VARIES (after cached prefix)
+4. # Environment ...                  ← VARIES
+5. <agent_instructions>               ← VARIES per agent
+6. memory / skills                    ← VARIES
+```
+
+This is a pure reordering — no content changes.
+The `<active_agent>` tag remains in the system prompt for pi-permission-system to find via regex.
+The env block and agent instructions still provide context to the model; their position relative to the inherited prompt is not semantically significant.
+
+## Module-Level Changes
+
+### `src/session/prompts.ts`
+
+1. Reorder the return statement in the `config.promptMode === "append"` branch to place `identity` (wrapped in `<inherited_system_prompt>`) and `bridge` before `activeAgentTag` and `envBlock`.
+2. Update the JSDoc comment on `buildAgentPrompt()` — replace "Both modes prepend" language with a description that notes the tag is included (not necessarily prepended) in append mode.
+
+### `test/session/prompts.test.ts`
+
+1. Update "prepends `<active_agent>` tag in append mode" — change from asserting `prompt.startsWith()` to asserting the tag appears *after* the inherited system prompt.
+2. Update "active_agent tag appears before envBlock in both modes" — the append-mode assertions change: the tag should still appear before the env block, but no longer at index 0.
+   The replace-mode assertions remain unchanged (`tagIdx === 0`).
+
+## Test Impact Analysis
+
+- Two existing tests assert `<active_agent>` is prepended (index 0) in append mode — these must change to assert the new ordering.
+- All other prompt tests use `toContain()` and are position-independent — they pass without changes.
+- No new test files or test surfaces are needed; the existing test suite covers the reordering adequately once the positional assertions are updated.
+
+## TDD Order
+
+1. **Red: update positional assertions for append mode.**
+   Change the two append-mode tests to assert the new ordering: `<inherited_system_prompt>` appears before `<active_agent>`, and the tag appears before the env block but not at index 0.
+   Commit: `test: assert cache-friendly prompt ordering in append mode (#180)`
+
+2. **Green: reorder the append-mode return statement.**
+   Move `identity` + `<inherited_system_prompt>` wrapper and `bridge` before `activeAgentTag` + `envBlock` in the return expression.
+   Update the JSDoc on `buildAgentPrompt()`.
+   Commit: `perf: reorder append-mode prompt for KV cache reuse (#180)`
+
+## Risks and Mitigations
+
+| Risk                                         | Mitigation                                                                                                                                         |
+| -------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------- |
+| pi-permission-system depends on tag position | Confirmed `ACTIVE_AGENT_TAG_REGEX.exec()` searches the full string — position-independent.                                                         |
+| Model behavior changes with reordered prompt | The same content is present; only ordering changes. The inherited system prompt as the "base" followed by specialization is arguably more natural. |
+| Replace mode accidentally affected           | Replace mode has its own code path and is not touched by this change.                                                                              |
+
+## Open Questions
+
+None — the design is straightforward and confirmed safe by code inspection.

package/docs/retro/0166-extract-parent-session-info.md

@@ -37,3 +37,30 @@ Type check and lint both clean after all steps.
 - `RunOptions` in `agent-runner.ts` needed a new import of `ParentSessionInfo` from `agent-manager.ts`; no circular dependency since `agent-runner.ts` already imports from `agent-manager.ts`.
 - `agent-tool.ts` still imports `AgentSpawnConfig` (needed by `AgentToolManager` interface) — the new `ParentSessionInfo` import was added alongside it.
 - All 5 commits are clean `refactor:` messages; architecture doc update is a separate `docs:` commit.
+
+## Stage: Final Retrospective (2026-05-24T18:00:00Z)
+
+### Session summary
+
+Planning, TDD implementation (5 steps), shipping, and CI verification all completed in a single session.
+Released as `pi-subagents-v6.18.2`.
+Zero rework — every TDD step went green on first attempt.
+
+### Observations
+
+#### What went well
+
+- The planning session's identification of the deep-merge trap in `background-spawner.test.ts`'s `makeParams` factory paid off — the TDD implementation handled it without friction because the risk was anticipated.
+- The 5-step inside-out TDD order (manager → runner → background → foreground → agent-tool) was the right sequence.
+  Each step only introduced type errors in files that subsequent steps would fix, with no circular breakage.
+- Clean mechanical execution — 805 tests before and after, zero rework commits, lint and type-check clean throughout.
+
+#### What caused friction (agent side)
+
+- `missing-context` — The plan repeated the issue body's stale "13 fields" count without verifying against the actual `AgentSpawnConfig` interface (which had 15 fields after `bypassQueue` was added in a prior issue).
+  The plan also inconsistently claimed the extraction would reduce the count to both "11" and "10" in different places.
+  Impact: required corrections in the architecture doc update, but no implementation rework.
+
+#### What caused friction (user side)
+
+- None observed — the user let the session run autonomously through all stages without intervention.

package/docs/retro/0180-reorder-append-prompt-for-kv-cache.md

@@ -0,0 +1,33 @@
+---
+issue: 180
+issue_title: "perf(pi-subagents): reorder append-mode system prompt to enable KV cache reuse"
+---
+
+# Retro: #180 — Reorder append-mode system prompt for KV cache reuse
+
+## Stage: Planning (2026-05-24T20:00:00Z)
+
+### Session summary
+
+Produced a plan to reorder the append-mode system prompt in `buildAgentPrompt()` so the shared inherited content (~8k tokens) comes before the varying `<active_agent>` tag and env block, enabling LLM KV cache prefix reuse across subagent invocations.
+
+### Observations
+
+- Confirmed pi-permission-system's `ACTIVE_AGENT_TAG_REGEX.exec()` is position-independent — no changes needed in that package despite the `pkg:pi-permission-system` label on the issue.
+- Only two tests assert positional ordering in append mode (`startsWith` and `tagIdx === 0`); all other prompt tests use `toContain()` and are unaffected.
+- Replace mode is a separate code path and is not touched.
+- The TDD cycle is minimal: one red step (update two positional assertions), one green step (reorder the return statement + update JSDoc).
+
+## Stage: Implementation — TDD (2026-05-24T20:15:00Z)
+
+### Session summary
+
+Completed both TDD cycles in `buildAgentPrompt()` in `src/session/prompts.ts`.
+Two positional assertions in `test/session/prompts.test.ts` were updated to expect the new ordering (red), then the append-mode return statement was reordered and the JSDoc updated (green).
+Test count unchanged at 805 across 50 files.
+
+### Observations
+
+- The JSDoc bullet for append mode also described the old ordering ("env header + parent system prompt + ...") and was corrected as part of the green step.
+- The `<active_agent>` tag is followed by a `\n\n`, so when it moves after `<sub_agent_context>`, a `\n\n` separator between the bridge and the tag was needed to maintain clean section boundaries.
+- No deviations from the plan; both steps were exactly as described.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@gotgenes/pi-subagents",
-  "version": "6.18.2",
+  "version": "6.18.3",
   "type": "module",
   "exports": {
     ".": "./src/service.ts"

package/src/session/prompts.ts

@@ -17,12 +17,14 @@ export interface PromptExtras {
  * Build the system prompt for an agent from its config.
  *
  * - "replace" mode: env header + config.systemPrompt (full control, no parent identity)
- * - "append" mode: env header + parent system prompt + sub-agent context + config.systemPrompt
+ * - "append" mode: parent system prompt + sub-agent context + env header + config.systemPrompt
  * - "append" with empty systemPrompt: pure parent clone
  *
- * Both modes prepend an `<active_agent name="${config.name}"/>` tag so downstream
+ * Both modes include an `<active_agent name="${config.name}"/>` tag so downstream
  * extensions (e.g. `@gotgenes/pi-permission-system`) can resolve per-agent policy
  * inside the child session by parsing the system prompt.
+ * In replace mode the tag is prepended; in append mode it follows the shared
+ * inherited content so the stable prefix is cacheable by the LLM's KV cache.
  *
  * @param parentSystemPrompt  The parent agent's effective system prompt (for append mode).
  * @param extras  Optional extra sections to inject (memory, preloaded skills).
@@ -76,13 +78,17 @@ You are operating as a sub-agent invoked to handle a specific task.
       ? `\n\n<agent_instructions>\n${config.systemPrompt}\n</agent_instructions>`
       : "";
 
+    // Place shared/stable content first so the LLM's KV cache can reuse the
+    // inherited prefix across all subagent invocations. The <active_agent> tag
+    // and env block vary per call and are placed after the cacheable prefix.
     return (
-      activeAgentTag +
-      envBlock +
-      "\n\n<inherited_system_prompt>\n" +
+      "<inherited_system_prompt>\n" +
       identity +
       "\n</inherited_system_prompt>\n\n" +
       bridge +
+      "\n\n" +
+      activeAgentTag +
+      envBlock +
       customSection +
       extrasSuffix
     );