feat(providers): add context-compression proxy plugin with tokenSavingMode#271
Open
feat(providers): add context-compression proxy plugin with tokenSavingMode#271
Conversation
…gMode Implements an intelligent context manager (ICM) that compresses older conversation messages before forwarding to the LLM API, reducing token usage when tokenSavingMode is enabled in the profile. - SmartCrusher: anchor-selection + token-budget compression (50% ratio) - Content-aware routing: log/diff/search/generic compressors - Adapters for Anthropic, OpenAI, and OpenAI-compatible formats - ICM phases: proactive non-tail compression, emergency tail compression, and message dropping as last resort - Savings tracker persists cumulative token savings to disk - /codemie:token-saver slash command to enable the feature - `codemie profile set token-saving-mode on/off` CLI support - Integration test script: scripts/test-compression.mjs Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
8nevil8
force-pushed
the
feat/claude-prompt-caching-autocompact-defaults
branch
from
4dbc809 to
24757d9
Compare
…of solution-architect
…perpowers routing - Rewrites SKILL.md v0.1 → v0.2: full SDLC flow (requirements → worktree → complexity scoring → brainstorming/writing-plans → spec-reviewer → subagent-driven-development → code review → qa-lead) - Replaces solution-architect references with superpowers:brainstorming + superpowers:writing-plans; 2-tier routing (5-7 simple, 8-15 brainstorming) - Rewrites branch-workflow.md: removes manual git checkout -b sections, delegates branch/worktree creation to superpowers:using-git-worktrees, fixes Node.js commands (npm run lint / npm test) - Updates complexity-assessment-guide.md: 3-tier routing → 2-tier routing aligned with SKILL.md; removes 11-12 borderline case Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
Orchestrates automated-tests → ui-tests (conditional) → spec-refinement (conditional) → /memory-refresh reminder. Invoked by tech-lead as Phase 5 after all implementation tasks complete and code review passes. Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
- automated-tests: lint → build → unit tests pipeline, stops at first failure - ui-tests: conditional frontend test runner, three-state outcome (PASS/FAIL/SNAPSHOT) - spec-refinement: aligns spec with implementation post-drift, commit deferred to user when invoked by qa-lead Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
…lowed-tools - Add YAML frontmatter (description + allowed-tools) to codemie-catchup, codemie-init, codemie-subagents, memory-refresh - Remove hardcoded absolute path from handoff.md (dev artifact) - Add allowed-tools to report-issue SKILL.md Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
…ia skills CLI - Add Phase 0: installs obra/superpowers project-level via `npx skills add obra/superpowers --all -a claude-code` - Flatten phase headers (## not nested ###) - Simplify Critical Rules section (remove emoji noise) - Update Next Steps to reference tech-lead as SDLC entry point Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
Centralize keyword sets and regex patterns into error-detection.ts, mirroring Python headroom error_detection.py. Fix CCR ratio threshold from 0.4 to 0.6 in diff/log/search compressors to match Python source. Also allowlist docs/superpowers/plans/ in .gitleaks.toml to suppress false-positive detection of well-known public JWT test vectors in plan docs. Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
…ceptor Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
…hex, version, req_id)
Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
…Router Add computePressureMinRatio(), expand ContentRouterConfig with three fields, and add routeWithPressure() method. 6 tests passing. Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
- Add JSDoc to routeWithPressure clarifying compressionRatio=1.0 means threshold not met - Add JSDoc to skipUserMessages config field (consumed by ICM layer, not ContentRouter) - Replace stub routeWithPressure test with 3 behavioral tests covering fill=0, fill=1, and result shape Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
Replace position/bias-based Phase 3 drop with score-based dropping (recency + error keyword signals). Export ContextStrategy, scoreMessage, buildDropCandidates. Add icm-strategy.test.ts (6 tests). Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
…gthen sort test - Phase 3 drop loop now captures message references before mutation, preventing stale-index filtering on subsequent iterations - Added JSDoc to ContextStrategy enum explaining it as preparatory infrastructure for future caller-controlled phase selection - Strengthened buildDropCandidates sort assertion to full ascending-order check; fixed misleading comment on index 0 score Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
…n ICM Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
…room 3e.5 parity) Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
… modules Core modules that were created but not yet committed: - ccr/types.ts: CcrStore interface contract - ccr/bm25.ts: BM25 scoring for semantic retrieval - ccr/store.ts: CompressionStore with CCR get/set and BM25 indexing - transforms/config.ts: CompressConfig interface and buildCompressConfig factory - transforms/hooks.ts: CompressionHooks and CompressionEvent interfaces - compressors/types.ts: add cacheKey field to CompressionResult Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
Remove skills added on this branch (automated-tests, qa-lead, spec-refinement, ui-tests) and restore spec-reviewer, tech-lead, and codemie-worktree to their main branch versions. Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
- Add `codemie profile set features.<group>.<flag> <value>` command - FEATURE_GROUPS registry with full contextCompression flag set - Value coercion: booleans (true/false/on/off), numbers, nullable (none→null) - FlagDef.min/max for generic range validation (targetRatio 0–1) - Replace old token-saving-mode set command with typed feature-flag command - Migrate interceptor gate from features.tokenSavingMode to features.contextCompression.enabled - Migrate cacheAligner read to ccFeatures.cacheAligner - buildCompressConfig now typed ProfileFeatures, reads features.contextCompression.* - AgentCLI --enable-token-saving sets contextCompression.enabled instead of tokenSavingMode Generated with AI Co-Authored-By: codemie-ai <codemie.ai@gmail.com>
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
Adds an Intelligent Context Manager (ICM) proxy plugin that compresses older conversation messages before forwarding requests to the LLM API, reducing input token usage when
tokenSavingModeis enabled in the user's profile.Changes
src/providers/plugins/sso/proxy/plugins/context-compression/): Full plugin with SmartCrusher, content-aware routing (log/diff/search/generic compressors), provider adapters (Anthropic/OpenAI/OpenAI-compatible), ICM three-phase pipeline (proactive non-tail, emergency tail, message dropping), and savings trackercodemie profile set token-saving-mode on/offCLI command +features.tokenSavingModeconfig field/codemie:token-saverslash command: One-command setup that enables tokenSavingMode in the active profile/handoffslash command: Session handoff with context preservation across compactionscripts/test-compression.mjs— runscodemie-claude --task, scans logs, verifies pipelineImpact
Token savings for multi-turn conversations with older messages containing logs, diffs, or repetitive content. ICM Phase 2 (proactive) compresses non-tail messages at ~50% ratio; Phase 1 (emergency) triggers only when over the context limit, preserving current-turn messages.
Checklist