fix(services): Trigger auto-compact from gateway-reported tokens#454
Merged
fix(services): Trigger auto-compact from gateway-reported tokens#454
Conversation
Auto-compact and session rollover gated on tokenizer.EstimateMessagesTokens(entries), which counts only message content. The actual prompt also includes the system prompt (~2K tokens) and tool definitions (~3-5K tokens for 15+ tools), so the gate underestimated by several thousand tokens — sessions never crossed the 80% threshold even when /context showed the window full. - SessionRolloverManager.tokenTriggerFires and ConversationOptimizer.OptimizeMessages now prefer repo.GetSessionTokens().LastInputTokens (the gateway's authoritative count, already shown by /context). Fall back to the entries-only estimate before the first round-trip when LastInputTokens is still 0. - Wire ConversationRepository into the optimizer config in the container. Also in this commit: - Refresh ContextMatchers to track the gateway's /v1/models output. Drop patterns for families the gateway doesn't serve (OpenAI direct, Llama, Cohere). Add Gemini 3 (1M), Gemma 3/3n/4, GPT-OSS, Ministral 3, Devstral, MiniMax M2, GLM 4/5, Nemotron 3, and deep-research (Gemini 3.1 Pro backed, 1M input window). Reorder so specific patterns precede generic ones. - Add LookupContextWindow returning a hit boolean so callers can distinguish a real match from the 8192 fallback. - Model picker (/switch) now shows the context window beside each row — "1M" / "128K" / "?" when no matcher hits — alongside the existing pricing. - Test coverage for Gemini, Gemma, Deep Research, DeepSeek, Mistral family, GPT-OSS / MiniMax / GLM / Nemotron, Qwen3 served variants, and the LastInputTokens trigger path on both the rollover manager and optimizer.
ig-semantic-release-bot Bot
pushed a commit
that referenced
this pull request
Apr 27, 2026
## [0.104.1](v0.104.0...v0.104.1) (2026-04-27) ### 🐛 Bug Fixes * **ui:** Restore typing while agent is busy ([#455](#455)) ([92840d6](92840d6)), closes [#410](#410) * **services:** Trigger auto-compact from gateway-reported tokens ([#454](#454)) ([1fc19dd](1fc19dd)) * **config:** Update model context windows and pricing for current model lineup ([#452](#452)) ([655f9f8](655f9f8)) ### ♻️ Code Refactoring * **config:** Centralize config loading and remove service indirection ([#443](#443)) ([babf173](babf173)) * **config:** Centralize sub-config definitions and consolidate prompts ([#448](#448)) ([9979ac7](9979ac7)) * Extract keybindings configuration to separate file ([#438](#438)) ([6d04195](6d04195)) * **config:** Generate viper defaults via reflection over DefaultConfig() ([#436](#436)) ([85c6e0a](85c6e0a)) * **config:** Make tool prompts configurable via prompts.yaml ([#450](#450)) ([9fc1bb5](9fc1bb5)), closes [#446](#446) * Move channels to separate channels.yaml config ([#444](#444)) ([aa43e0e](aa43e0e)), closes [#441](#441) * Move computer_use to a separate config ([#447](#447)) ([a762c64](a762c64)), closes [#444](#444) * Move prompts to separate prompts.yaml config ([#442](#442)) ([45e4fb6](45e4fb6)) * **config:** Unify sub-configs behind CollectionConfig + utils.{Load,Save}YAML ([#445](#445)) ([aab481c](aab481c)) ### 📚 Documentation * Add directory-structure reference ([#451](#451)) ([95d26f8](95d26f8)) ### 👷 CI/CD * Reduce runs of nix ([acac364](acac364)) ### 🧹 Maintenance * **deps:** Bump modernc.org/sqlite from 1.49.1 to 1.50.0 ([#449](#449)) ([06a535d](06a535d)) * **nix:** Update package to v0.104.0 ([#432](#432)) ([5564697](5564697))
Contributor
|
🎉 This PR is included in version 0.104.1 🎉 The release is available on GitHub release Your semantic-release bot 📦🚀 |
edenreich
added a commit
that referenced
this pull request
Apr 27, 2026
) ## Summary - `GenerateLLMSummary` hardcoded `maxTokens := 200` and ignored `FinishReason`, so when the model hit the cap the truncated mid-sentence text was rendered as a complete `--- Context Summary ---` block. Raise the default to **1024** and expose it as `compact.summary_max_tokens` for tuning. - Warn-log when `response.Choices[0].FinishReason == sdk.Length` so future truncations are visible in logs instead of failing silently. - Bug surfaced after #454 (`fix(services): Trigger auto-compact from gateway-reported tokens`) made auto-compact actually fire on the gateway-reported token count — the 200-cap was always too tight, but compaction rarely triggered before, so the truncation was rarely visible.
ig-semantic-release-bot Bot
pushed a commit
that referenced
this pull request
Apr 27, 2026
## [0.104.2](v0.104.1...v0.104.2) (2026-04-27) ### 🐛 Bug Fixes * **services:** Prevent compact summary truncation at 200-token cap ([#457](#457)) ([36b9612](36b9612)), closes [#454](#454) * Reconcile token displays and persist metadata-only saves ([#459](#459)) ([8bc8767](8bc8767)) ### 📚 Documentation * Update agents MD ([#458](#458)) ([27bfaea](27bfaea)) ### 🧹 Maintenance * **nix:** Update package to v0.104.1 ([#456](#456)) ([784e4bc](784e4bc))
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Closes #453. Auto-compact and session rollover never fired because the trigger gated on
tokenizer.EstimateMessagesTokens(entries)— which counts only message content. The real prompt also includes the system prompt (~2K) and tool definitions (~3–5K for 15+ tools), so sessions blew past the 80% threshold without ever crossing the gate's underestimate.repo.GetSessionTokens().LastInputTokens(the gateway's authoritative count, already shown by/context). Falls back to the entries-only estimate before the first round-trip.ContextMatchersto track the gateway's actual/v1/modelsoutput — added Gemini 3, Gemma 3/3n/4, GPT-OSS, Ministral 3, Devstral, MiniMax M2, GLM 4/5, Nemotron 3, anddeep-research(1M); dropped patterns for families the gateway doesn't serve./switch) now shows the context window beside each row (1M/128K/?when unknown) alongside pricing.