Include split-turn prefix in real-conversation gate

The early cancel check only scans preparation.messagesToSummarize, so split-turn runs can be cancelled even when the actual user request is present in preparation.turnPrefixMessages and would be summarized later in this same handler. Fresh evidence in this commit is that the new content-aware gate at messagesToSummarize.some((message, index, messages) => isRealConversationMessage(...)) was added here, while split-turn prefix handling still occurs later (preparation.isSplitTurn path), which can now incorrectly skip compaction and let history keep growing in heartbeat-heavy sessions.

Useful? React with 👍 / 👎.

Include split-turn prefix in the real-conversation gate

This cancellation check only scans preparation.messagesToSummarize, but split-turn compaction later summarizes preparation.turnPrefixMessages as well, so a real user request living in the prefix can be ignored here. In heartbeat-heavy windows where messagesToSummarize is boilerplate, this now returns cancel: true and skips compaction even though meaningful content exists in the same compaction operation, which lets session history keep growing instead of being summarized.

Useful? React with 👍 / 👎.

Ignore reasoning-only blocks when classifying meaningful turns

This branch treats every non-text block except tool-call variants as meaningful, so assistant turns that only contain provider reasoning/thinking metadata are counted as “real conversation.” Because transcript sanitization preserves thinking blocks for most providers by default (dropThinkingBlocks is usually false), heartbeat/NO_REPLY sessions on reasoning-capable models can still bypass this safeguard and continue compacting on non-conversational traffic.

Useful? React with 👍 / 👎.

Include turnPrefixMessages in real-conversation gate

The new cancellation guard only checks preparation.messagesToSummarize, so split-turn compactions can now be cancelled even when the actual user request lives in preparation.turnPrefixMessages. This is a regression because the same handler later summarizes prefix messages when preparation.isSplitTurn is true (see the turnPrefixMessages summarization path), so heartbeat-heavy windows in messagesToSummarize will incorrectly return { cancel: true } and skip compaction despite meaningful content being part of the current compaction run.

Useful? React with 👍 / 👎.

Strip markup before checking NO_REPLY sentinel

hasMeaningfulText rejects silent turns only through isSilentReplyText(trimmed), but that matcher is exact-token only and does not handle wrapped forms like **NO_REPLY** or <b>NO_REPLY</b>. In heartbeat sessions where models format the sentinel, these assistant turns are misclassified as meaningful conversation, so the new guard still allows compaction on boilerplate traffic instead of canceling. This path already normalizes markup for HEARTBEAT_OK, so silent-token detection needs the same normalization to keep behavior consistent.

Useful? React with 👍 / 👎.

Include split-turn prefix messages in real-conversation gate

This early-cancel check only scans preparation.messagesToSummarize, but the same handler later summarizes preparation.turnPrefixMessages when preparation.isSplitTurn is true. In split-turn sessions where the meaningful user request is in the prefix and the summarize window contains heartbeat boilerplate (HEARTBEAT_OK/NO_REPLY), this condition now returns { cancel: true } and skips compaction even though real conversation is present in the current compaction operation, allowing history growth instead of summarization.

Useful? React with 👍 / 👎.

Let safeguard mode write its fallback compaction boundary

When agents.defaults.compaction.mode is safeguard, buildEmbeddedExtensionFactories() always registers compactionSafeguardExtension, and that handler deliberately returns a minimal compaction result instead of cancelling when there is no real conversation because otherwise _checkCompaction re-triggers in a loop (src/agents/pi-extensions/compaction-safeguard.ts:711-723). This new early return skips session.compact() entirely for heartbeat-only histories, so the safeguard never gets a chance to write that boundary, and overflow recovery later treats the result as a failed compaction. In safeguard-mode heartbeat sessions, this reintroduces the stuck/retrigger behavior the extension was added to prevent.

Useful? React with 👍 / 👎.

Ignore Anthropic toolResult blocks in meaningful-content checks

Anthropic transcript repair keeps tool results inside role: "user" messages as content[].type === "toolResult" (src/agents/pi-embedded-helpers/turns.ts:49-58), but this branch treats every non-text block other than tool-call variants as meaningful. That means Claude sessions whose "user" turns are only tool-result payloads still count as real conversation, so the new guard will keep compacting heartbeat/tool-only traffic instead of skipping it. toolResult blocks need to be excluded here the same way tool-call blocks are.

Useful? React with 👍 / 👎.

Link tool results to summarized asks, not only in-window users

After one compaction cycle, the original user ask is no longer present as a message entry (session-truncation.ts removes summarized messages before firstKeptEntryId). A later tool-only span from that same task therefore has no role === "user" inside this lookback, so every toolResult here is misclassified as non-conversational. In compact.ts that turns overflow recovery into compacted: false, and in safeguard mode it falls back to previousSummary, dropping the new tool output from future context. Any long-running task that compacts once before the assistant replies can now lose state or fail overflow recovery.

Useful? React with 👍 / 👎.

Treat input/output_text blocks as textual content

OpenAI session history already accepts input_text and output_text blocks as normal message text (openai-ws-stream.ts handles both in contentToText()/contentToOpenAIParts()), but this classifier only recognizes literal type === "text". A boilerplate message like [{type:"output_text", text:"NO_REPLY"}] or [{type:"input_text", text:"HEARTBEAT_OK"}] therefore falls into the non-text branch and counts as real conversation, so the new safeguard still compacts boilerplate-only OpenAI/Responses transcripts instead of skipping them.

Useful? React with 👍 / 👎.