Compaction summaries were limited to 0.8 * compaction.reserveTokens (13107 by default), turn-prefix summaries to 0.5 * reserveTokens, and branch summaries to a hardcoded 4096. On providers where reasoning shares the output limit (OpenAI Responses, adaptive Anthropic), high thinking levels used most of that limit and compaction failed with "generation hit the token cap", even though the model allows far more output.
Tying the limit to reserveTokens assumed the summary request is as large as the full context, so its output had to fit in the reserved space. It is usually much smaller: recent messages kept by keepRecentTokens are not sent, a split turn goes to a separate request, tool results are truncated, and tool definitions and the normal system prompt are not included. Manual compaction can also run long before the context is full. The provider layer already sizes output from the actual request, using the model's output limit reduced to fit the remaining context, as for normal turns.
Summary requests now set no maxTokens and rely on that. compaction.reserveTokens only controls when compaction starts.
generateSummary() and generateSummaryWithUsage() now ignore the reserveTokens parameter.
closes#9512, closes#9075, closes#10908
Replayed thinking is often not part of the live context (for example,
radius drops it, and Anthropic drops blocks that do not match their
request), but serializeConversation included all of it. A Pi Mobile
session at ~996K of 1M context tokens produced 1.38M-token summarization
requests that failed on every attempt. Providers only return summaries of
the reasoning anyway; visible text, tool calls, and tool results record
what happened.
Make SessionManager projections authoritative for provider requests, add append-only context edits and actionable turn boundaries, and preserve existing queue scheduling during continuation and recovery.
Run next-turn preparation only when the agent loop continues, allowing threshold compaction before the next provider request. Restore interactive progress when the same run resumes.
Closes#6879
Important to notice:
this changes when prepareNextTurn & prepareNextTurnWithContext are invoked:
the now run only when a next turn actually happens, and after shouldStopAfterTurn.
Add provider-context prefix preparation and optional summary request settings needed by cache-friendly compaction. Existing AgentSession callers continue to pass undefined, so standalone summarization remains the default until the feature is wired in.
session_before_compact and session_compact now carry
reason ("manual" | "threshold" | "overflow") and willRetry, matching
what the RPC protocol already exposes via auto_compaction_start/end.
Extensions can now distinguish manual /compact from threshold
auto-compaction and overflow recovery through the public API instead
of monkeypatching AgentSession internals.
- New compaction.md covers auto-compaction and branch summarization
- Explains cut points, split turns, data model, file tracking
- Documents session_before_compact and session_before_tree hooks
- Rewritten hooks.md matches actual API (separate event names)
- Correct ctx.ui.custom() signature (returns handle, not callback)
- Documents all session events including tree events
- Adds sessionManager and modelRegistry usage
- Updates all examples to use correct API
pi-ai:
- Fixed usage.input to exclude cached tokens for OpenAI providers
- Previously input included cached tokens, causing double-counting
- Now input + output + cacheRead + cacheWrite correctly gives total context
coding-agent:
- Session header now includes branchedFrom field for branched sessions
- Updated compaction.md with refined implementation plan
- Updated session.md with branchedFrom documentation
