Skip to content
trimwire

Roadmap

trimwire is released and feature-complete. This is not a release schedule. The items below are evidence-gated ideas; each ships only if real telemetry or a benchmark shows it earns its complexity.

Shipped

Section titled “Shipped”
  • Eight deterministic cache-safe pruning strategies (default and gentle profiles)
  • Stable-prefix re-pruning ([reprune]) for cache stability
  • Opt-in summarizer: local ollama engine + cloud API engine (multi-provider, fallback cascade)
  • Opt-in anonymous telemetry (trimwire share stats / share benchmark)
  • Community collector + dashboard (collector pending deploy)
  • Offline benchmark harness with quality corpus (5 slices, harm gate, FCS metric)
  • trimwire sweep for on-disk transcript maintenance
  • trimwire preview / trimwire recall / trimwire dashboard
  • Socket-activated always-up service (systemd / launchd)

Possible (evidence-gated, not committed)

Section titled “Possible (evidence-gated, not committed)”
  • Richer deterministic elision markers. Today a stub records the size elided. A deterministic extractive marker (keeping first/last lines, or error|warn|fail lines) would provide a content breadcrumb with zero network or model dependency.
  • Telemetry-driven threshold tuning. Current thresholds are conservative defaults. Real trimwire share stats data could justify per-profile tuning, but only once there is real-traffic evidence of the cost-vs-headroom trade-off.
  • Benchmark additions. A post-compaction corpus (a [summary] turn + recent tail) and a profile-knob sensitivity sweep (savings vs. cache-stability across keep_recent_turns / bloat threshold) to show the frontier rather than three fixed profile points.

Beyond Claude Code

Section titled “Beyond Claude Code”

The largest opportunity: prune for other agent harnesses too — aider, opencode, cline, Codex, and more. The pruning value is universal, but no transparent, deterministic proxy exists for them yet.

Two things make this tractable:

  • Some already work today. Harnesses that let you point their Claude/Anthropic provider at a custom URL (opencode, cline) get full pruning right now — they speak the same format trimwire already understands.
  • The rest need a small adapter layer. trimwire’s pruning runs at one internal seam, so adding a harness means translating its request format in and out around that seam — the pruning logic itself stays the same. The first target would be OpenAI-Chat-Completions harnesses (e.g. aider, via a single OPENAI_API_BASE env var); the goal is a trimwire install <harness> flow.

This is discovery-complete but not committed. The full engineering plan (adapter design, phasing, the aider spike) is tracked internally in docs/MULTI-HARNESS-PLAN.md.

Non-goals

Section titled “Non-goals”
  • Gateway-side LLM summarization on your Anthropic OAuth token. Prohibited by ToS (Anthropic actively enforced this in 2026; the subscription token is licensed only for ordinary Claude Code use). Also: it adds latency on the request path, spends tokens you didn’t ask to spend, and breaks bit-exact reproducibility. The opt-in summarizer ([summarizer]) threads all four objections: it uses a model you choose on your own key, runs in the background, and caches the result. See SUMMARIZER.md.