ContextLattice v3.9.0

Release date: 2026-07-01

v3.9.0 turns the Context Pack Quality Ledger from a passive endpoint into an agent-operable outcome loop. Agents can now carry the latest context-pack quality sample through their session and post compact outcome rows with first-pass, repair, retry, follow-up-token, and provider-usage counters.

Highlights

• Adds contextlattice_agent_adapter outcome for compact context-pack outcome reporting.
• Lets contextlattice_agent_adapter complete post outcome telemetry when runtimes know first-pass, repair, retry, or provider-token facts.
• Persists the latest context-pack quality sample id in bounded local agent session state so completion hooks do not need to remember the endpoint.
• Adds outcome-report hints to context-pack outputs without storing prompts, completions, source text, or secrets.
• Extends /telemetry/context-pack-quality/outcome with provider prompt, completion, and total token counters.
• Keeps modeled inference avoidance separate from observed provider usage in telemetry and the dashboard cockpit.
• Updates public agent guidance so runtimes report known outcome facts and do not guess missing fields.

Measurement Contract

v3.9.0 keeps three claims separate:

• Exact prompt-token savings: tokenizer-aware context-pack token delta.
• Modeled inference avoidance: confidence-banded counterfactual until outcome rows calibrate it.
• Observed provider usage: provider-reported token counters posted by the agent runtime when available.

Outcome rows are compact counters. They do not carry raw user prompts, model completions, retrieved source text, or local file paths.

Agent Contract

Recommended completion path:

contextlattice_agent_adapter complete \
  --agent codex \
  --project contextlattice \
  --session-id "$SESSION_ID" \
  --summary "completed" \
  --first-pass-success true \
  --repair-required false \
  --retry-count 0

When a runtime has provider usage, it can add:

--provider-prompt-tokens 12000 --provider-completion-tokens 900 --provider-total-tokens 12900

If the runtime does not know an outcome field, it should omit that field instead of inventing it.

Verification

• Full Go gateway test suite passed.
• Go-native agent tools tests passed.
• Dashboard metric tests passed.
• Dashboard production build passed.
• Python adapter and context-pack scripts compiled.
• Diff whitespace check passed.