v1.6.0 - Docstring stripping, TS condensation, function truncation

v1.6.0

What's new

Python docstring stripping — Docstrings are stripped from rendered symbol bodies. The implementation and signature remain; the prose description does not. Recovers 10–100+ lines per function/class in docstring-heavy codebases.

TypeScript/JavaScript class condensation — Large TS/JS classes (> 40 lines) now get the same treatment as Python classes: first 40 lines verbatim, remaining methods as signature stubs.

Standalone function truncation — Any function/method body exceeding 60 lines is truncated with a # ... (N lines omitted) note. Prevents single large functions from crowding out other symbols.

Benchmark (cold start, 4-task study)

Task	top-k (uncompressed)	v1.6 cold	reduction
add rate limiting to auth API	125,701	958	-99.2%
refactor database connection pooling	53,586	12,888	-75.9%
fix memory leak in request handler	113,466	5,134	-95.5%
add JWT token validation middleware	130,934	4,913	-96.2%
average	423,687	23,893	-94.4%

Cumulative cold-start progression

Version	Feature	Avg cold reduction
v1.2	baseline	~37%
v1.3	threshold tuning, test file detection, stub scoring	~41%
v1.4	utility file detection, cross-file import dedup	~45%
v1.5	Python class body condensation	~95% (single task)
v1.6	docstring stripping + TS condensation + func truncation	94.4% (4-task avg)

v1.5.0 - Faza 3: Class Body Condensation

v1.5.0 - Faza 3: Class Body Condensation

What's new

• Large class bodies (> 40 lines) are now condensed in symbol extraction
• First 40 lines are kept verbatim; remaining methods are rendered as signature-only stubs
• Classes with fewer than 40 lines are unaffected

Benchmark (cold start, "add rate limiting to auth API")

Strategy	Tokens	vs top-k
top_k_selection (uncompressed)	125,703	baseline
compressed_pack cold	6,416	-94.9%
compressed_pack warm	387	-99.7%

Cumulative cold-start improvements

Version	Feature	Cold reduction
v1.2	baseline	~37%
v1.3	threshold tuning, test file detection, stub scoring	~41%
v1.4	utility file detection, cross-file import dedup	~45%
v1.5	class body condensation	~95%*

*measured against uncompressed top-k on the Redcon repo itself

v1.4.0 — Phase 2 Cold-Start

Phase 2 cold-start compression improvements

Changes

1. Utility file detection
Files matching config.*, helpers.*, utils.*, validators.*, constants.*, settings.*, types.*, exceptions.*, errors.* now bypass the full-file path and always use symbol extraction — same mechanism as test file handling in v1.3. Bodies of symbols with no keyword hits are stubbed to signature-only lines.

2. Cross-file import deduplication
import and from ... import lines already emitted by an earlier file are stripped from subsequent files' compressed output. Eliminates repeated boilerplate imports across route handlers, services, and models.

Per-file impact (Redis caching task)

File	Before	After	Reduction
src/utils/helpers.py	268	55	79%
src/utils/validators.py	249	123	51%
src/config.py	392	177	55%
tests/test_users.py	463	66	86%

Cumulative benchmark (cold cache, 15-file Python microservice)

Version	Avg tokens	Avg reduction
v1.2	7,749	37%
v1.3	7,272	41%
v1.4	6,750	45%

Best case (JWT auth): 7,512 → 5,283 tokens (57% reduction cold start)

v1.3.0 — Cold-Start Compression

Cold-start compression improvements

Redcon v1.3 reduces token usage on first run (no prior cache) with four targeted optimizations to the compression pipeline.

Changes

1. Lower full-file threshold (600 → 300 tokens)
Files between 300-600 tokens now go through symbol/slice extraction instead of being sent verbatim. Previously, small utility files (config, validators, helpers) were always sent in full.

2. Tighter keyword-window snippets

• snippet_hit_limit: 8 → 6
• snippet_context_lines: 2 → 1

3. Test file detection
Files matching test_*, *_test.*, */tests/*, *spec* patterns are always compressed via symbol extraction — even at small sizes — because fixture-heavy structure wastes tokens when sent in full.

4. Signature-only stubs for low-relevance symbols
Symbols with no keyword hits in the body (score < 3.5) are now rendered as a single signature line instead of the full function body. Symbols with keyword hits are unaffected.

Benchmark (cold cache, 15-file Python microservice)

	v1.2	v1.3
Average tokens	7,749	7,272
Average reduction	37%	41%
JWT auth task	7,512 (39%)	5,992 (51%)
Refactor task	7,831 (36%)	7,232 (41%)

v1.2.0

What's new

Dual-license model

• MIT for the core engine (context optimization, compressors, scanners, CLI, SDK, benchmarks)
• Proprietary license for commercial components (gateway, control plane, agent middleware, LLM integrations, runtime, telemetry)
• Added LICENSE-COMMERCIAL and SPDX headers to all proprietary files

Benchmark corrections

• Fixed inflated savings numbers — previous figures were measured with warm cache from sequential runs
• Honest cold-start numbers: 37% reduction (first run, no prior cache)
• Warm-cache numbers: 92% reduction (subsequent runs with summary reuse)
• Updated README, landing page, docs, and wiki

Infrastructure

• Fixed CI no-merge-base error (--depth=1 removed from git fetch)
• Fixed Dockerfile: renamed contextbudget to redcon, installs [gateway] extra
• Added openai, anthropic, all extras to pyproject.toml
• Added CI matrix workflow (Python 3.10 / 3.11 / 3.12 x core / gateway / redis)
• Added docker-compose.yml for gateway + Redis stack

Demo and benchmarks

• Added end-to-end demo (demo/run_demo.py) covering pack, policy, cost analytics, gateway, and LLM adapters
• Added 1,000-task benchmark corpus, runner, and pre-generated results (redcon-benchmarks/)

Redcon v1.1.0

Redcon v1.1.0

Project renamed from ContextBudget → Redcon.
All packages, env vars (RC_*), API key prefix (rck_), and Prometheus metrics (rc_*) have been updated accordingly.

Enterprise features

P0 — Security & Observability

• Prometheus metrics endpoint (/metrics) with request latency, event ingest, and rate-limit counters
• Per-org ingest rate limiting (RC_CLOUD_EVENTS_RATE_LIMIT / RC_CLOUD_EVENTS_RATE_WINDOW)
• Admin bootstrap token (RC_CLOUD_ADMIN_TOKEN) required for POST /orgs

P1 — Scalability

• Redis-backed session store for horizontal gateway scaling (RC_GATEWAY_REDIS_URL)
• PgBouncer connection pooling (deploy/pgbouncer.ini, redcon-cloud/docker-compose.prod.yml)
• Per-org usage quotas (migrations/005_quotas.sql) and Stripe billing hooks (migrations/006_billing.sql)
• TypeScript/Node.js SDK (@redcon/sdk) with CloudClient and GatewayClient

P2 — Enterprise polish

• Multi-region Postgres streaming replication (deploy/postgres-primary.conf, deploy/postgres-replica.conf, redcon-cloud/docker-compose.multiregion.yml)
• Read replica routing for analytics queries (READ_DATABASE_URL)
• OIDC SSO skeleton (RC_CLOUD_OIDC_*)
• Slack and PagerDuty webhook adapters
• 5-step onboarding wizard UI

Migration notes

Old	New
pip install contextbudget	pip install redcon
CB_GATEWAY_* env vars	RC_GATEWAY_*
CB_CLOUD_* env vars	RC_CLOUD_*
cbk_... API keys	rck_...
@contextbudget/sdk	@redcon/sdk
contextbudget-cloud/	redcon-cloud/

Existing API key values in the database are unaffected — only the prefix format in new keys issued after this release changes.

ContextBudget v1.0

ContextBudget v1.0

First stable release of the ContextBudget platform.

Core Engine

• File scanning, ranking, and compression pipeline
• Token-budget enforcement with configurable policy
• Delta-context propagation for multi-turn agent sessions
• On-disk summary cache with pluggable backends (local file, Redis)

Gateway

• FastAPI runtime gateway (/prepare-context, /run-agent-step, /report-run)
• Remote policy fetch from cloud control plane
• Fire-and-forget audit push after each request

Cloud Control Plane (contextbudget-cloud)

• Multi-tenant org / project / repository hierarchy
• API key issuance with SHA-256 hashing and optional expiry
• Append-only audit log with gateway push endpoint
• Versioned policy management with scope inheritance (repo > project > org)
• PostgreSQL row-level security on the events table
• Dashboard endpoints: overview, repositories, savings, heatmap
• Cost analytics: baseline vs optimised tokens, savings by repo and date
• Agent run tracking

CLI

• cb pack, cb scan, cb score, cb plan, cb cost-analysis
• --json output flag and run artifact export

Testing

• 600 tests (522 core + 78 cloud), all passing