Pi plugin: memory protocol is lost on turns started without before_agent_start (breaks pi-claude-bridge)

[AlejandroTatum]

📝 Bug Description

The Pi plugin injects MEMORY_INSTRUCTIONS (and the post-compaction recovery notice) by returning { systemPrompt } from before_agent_start (plugin/pi/index.ts, still present on main, around L1961–1982).

Pi (0.87.1) turns a handler-returned systemPrompt into a run-scoped forceSystemPrompt (extensions/runner.js) and clears it when the run ends (agent-session.js, _runSystemPromptOptions = undefined). A turn started by pi.sendMessage(..., { triggerTurn: true }) while Pi is idle — for example a background subagent result — runs without before_agent_start, so that turn's system prompt reaches the provider without the Engram memory protocol.

• With most providers this is silent: the model just runs without the memory protocol.
• With pi-claude-bridge the turn fails, because the bridge recorded the prompt including Engram's text and receives one that is an exact prefix of it.

The missing tail is exactly "\n\n" + MEMORY_INSTRUCTIONS (2673 chars), observed on two different prompts.

🔄 Steps to Reproduce

1. Pi 0.87.1 with the Engram Pi plugin and pi-claude-bridge, using a claude-bridge/* model.
2. Launch any subagent in background mode (for example gentle-pi's subagent_run with mode: "background") and let the parent turn end.
3. When the result is delivered (sendMessage with triggerTurn: true), the new turn fails.

✅ Expected Behavior

The memory protocol is present on every turn, including turns started by triggerTurn messages. No provider error.

❌ Actual Behavior

Error: prompt-capture: no capture for this 62770-char system prompt, and it embeds none of the 12 known.
Closest known match diverges at offset 62770 (65443-char key, last recorded at turn_start).

65443 − 62770 = 2673 = "\n\n" + MEMORY_INSTRUCTIONS.

Operating System

Linux (Arch/Manjaro)

Engram Version

2.2.1 (Pi plugin from a source checkout; the same code is on main)

Agent / Client

Other

📋 Relevant Logs

No response

💡 Additional Context

• Client detail: Pi coding agent 0.87.1 + pi-claude-bridge.
• Proposed fix (same pattern as gentle-shell#1485 / bug(ci): existing policy and Obsidian regression suites are not run #1497): append the protocol idempotently to event.systemPromptOptions.appendSystemPrompt instead of returning a replacement systemPrompt; keep the current return only as a fallback when systemPromptOptions is absent (older Pi).
• Verified locally: 4 new tests (RED 3/4 → GREEN 4/4), full plugin suite 131/131, and the real repro (background subagent result arriving while Pi is idle) no longer fails.
• Related, same symptom from other layers: bug(prompt): background subagent completion turn fails on claude-bridge (prompt-capture miss; appended harness likely missing) gentle-shell#1528, prompt-capture throws on a triggerTurn follow-up turn: failing prompt is an exact prefix of the last recorded key (tail missing) elidickinson/pi-claude-bridge#144. Neither identifies Engram as the source of the missing tail.

[dnlrsls]

Thanks for the detailed report. The Engram-side correction is already merged in #1589, commit 4cb56ac085f53088c7058ea17e4c398f739f1a5b, and was first published as gentle-engram@0.2.0 on October 1. The npm package points to the pi-v0.2.0 release commit, which contains the fix; pi-v0.1.16 still uses the old replacement strategy. The PR closed #1588 rather than this report.

We confirmed the prompt-lifetime bug in an isolated Pi 0.87.1 SDK session using the current hook legacy fallback as a controlled baseline and an offline provider. An idle pi.sendMessage(..., { triggerTurn: true }) made two provider requests but invoked before_agent_start only once. The replacement strategy lost exactly 2673 characters, "\n\n" + MEMORY_INSTRUCTIONS; the structured appendSystemPrompt path retained the complete protocol on both requests. Limits: no real background child or Claude call was executed. Using bridge 0.9.0 PromptCaptures and its public capture boundaries did not reproduce the reported exception, so the experimental assertion expecting that exception failed. Focused plugin tests passed 3/3; the local full suite was 313/314 because our temporary-root configuration caused Go to ignore go.mod, not an observed code regression.

The PR Plugin Tests check passed, and the 0.2.0 publishing workflow successfully ran the plugin suite before publishing. Those checks do not cover a full Pi + Claude bridge E2E. Please update the Pi package to gentle-engram@0.2.0 (distinct from the Engram core version) and restart/reload Pi. If the original bridge failure persists, the exact bridge version and extension load order would help distinguish a remaining integration issue.