This release adds 3 notable features for engineering teams evaluating rollout.
✓ No known CVEs patched in this version
Topics
+4 more
Summary
AI summaryIntroduces a bounded agent_packet.v1 with gated actions and dashboard separation of token, session, and learning truth.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Feature | Medium |
Sets default token limit to 2,000 with a ceiling of 4,000 tokens for the packet. Sets default token limit to 2,000 with a ceiling of 4,000 tokens for the packet. Source: llm_adapter@2026-07-13 Confidence: high |
— |
| Feature | Medium |
Adds tokenizer‑counted serialized transport JSON for the packet. Adds tokenizer‑counted serialized transport JSON for the packet. Source: llm_adapter@2026-07-13 Confidence: high |
— |
| Feature | Medium |
Implements session reuse for identical task identities, avoiding duplicate `session.started` events. Implements session reuse for identical task identities, avoiding duplicate `session.started` events. Source: llm_adapter@2026-07-13 Confidence: high |
— |
| Feature | Medium |
Introduces a dashboard separating Token truth, Session truth, and Learning truth. Introduces a dashboard separating Token truth, Session truth, and Learning truth. Source: llm_adapter@2026-07-13 Confidence: high |
— |
| Feature | Medium |
Updates CLI commands to the new normal workflow (context, resume, remember, correct, finish, doctor). Updates CLI commands to the new normal workflow (context, resume, remember, correct, finish, doctor). Source: llm_adapter@2026-07-13 Confidence: high |
— |
| Feature | Medium |
Makes `finish` automatically report the latest pending retrieval outcome unless disabled. Makes `finish` automatically report the latest pending retrieval outcome unless disabled. Source: llm_adapter@2026-07-13 Confidence: high |
— |
| Feature | Medium |
Adds `correct` command to record feedback categories (useful/wrong/stale/superseded). Adds `correct` command to record feedback categories (useful/wrong/stale/superseded). Source: llm_adapter@2026-07-13 Confidence: high |
— |
| Feature | Medium |
Introduces Truth / Action workbench that requests a bounded Agent Packet from Synthesis v2 without executing commands. Introduces Truth / Action workbench that requests a bounded Agent Packet from Synthesis v2 without executing commands. Source: llm_adapter@2026-07-13 Confidence: high |
— |
| Feature | Medium |
Documents verification suite: Go tests, Rust crate tests, dashboard test suite, TypeScript check, production build, and contract audits. Documents verification suite: Go tests, Rust crate tests, dashboard test suite, TypeScript check, production build, and contract audits. Source: llm_adapter@2026-07-13 Confidence: high |
— |
| Feature | Medium |
Introduces `agent_packet.v1` with bounded evidence and provenance. Introduces `agent_packet.v1` with bounded evidence and provenance. Source: llm_adapter@2026-07-13 Confidence: low |
— |
| Feature | Low |
Introduces `agent_packet.v1` with bounded evidence, provenance, uncertainty, next actions, continuation state, outcome identity, and an explicit `act|verify|abstain` gate. Introduces `agent_packet.v1` with bounded evidence, provenance, uncertainty, next actions, continuation state, outcome identity, and an explicit `act|verify|abstain` gate. Source: granite4.1:30b@2026-07-13-audit Confidence: low |
— |
Full changelog
ContextLattice v3.16.0 - The Agent Packet
Agents do not need more memory dumped into their prompt. They need the right
truth, compressed without losing its proof, delivered at the moment it can
change the work.
v3.16.0 turns that principle into the default ContextLattice experience.
One packet, not a payload parade
agent_packet.v1 carries bounded evidence, provenance, uncertainty, next
actions, continuation state, outcome identity, and an explicit
act|verify|abstain gate. The normal target is 2,000 tokens with a 4,000-token
product ceiling. Full compiler contracts remain available through --full or
--debug, but they no longer tax every ordinary agent turn.
Serialized transport JSON is tokenizer-counted. Compiled prompt size and wire
size stay separate. If the envelope costs more than the raw-evidence
counterfactual, ContextLattice reports the loss and claims zero savings.
One task, one live session
- Repeated entry with the same task/native/repo/branch/worktree identity reuses
one live session without duplicatingsession.started. - Idle expiry defaults to 12 hours and honors explicit lifecycle TTL/expiry.
- Completed, failed, canceled, and expired state is absorbing.
- List and resume commands default to compact views;
--fullis explicit. - Runtime telemetry separates live, expired, terminal, and total sessions.
Async truth that cannot run backward
Queued, pressure-shed, cooldown, and durable-retry work stays pending or
warming. It is never mislabeled degraded while work remains viable. Source
state and agent steering are monotonic: at most one progress notice, followed by
one absorbing ready or degraded terminal notice.
Configured-but-disabled sources remain visible as configuration truth without
poisoning effective source coverage.
Retrieval with a burden of proof
- Intent alignment, recency, supersession penalties, and normalized cross-source
deduplication shape evidence ranking. - Weak evidence abstains.
- Partial coverage verifies.
- Contradictions can tighten an act gate but never loosen a refusal.
- Verify and abstain packets remove executable mutation commands.
A learning loop agents will actually use
The normal CLI is now:
contextlattice context "current task" --project contextlattice --pretty
contextlattice resume --project contextlattice --pretty
contextlattice remember "checkpoint summary" --project contextlattice --pretty
contextlattice correct "retrieval was stale" --category stale --project contextlattice --pretty
contextlattice finish "verified result" --success --project contextlattice --pretty
contextlattice doctor --pretty
finish automatically reports the latest pending retrieval outcome unless
explicitly disabled. correct records useful/wrong/stale/superseded feedback;
factual mutation remains separate and requires explicit claim fields.
The proof cockpit
The dashboard now separates:
- Token truth: serialized transport, compiled prompt size, net delta, and
measurement coverage. - Session truth: live, total, expired, and idle-TTL state.
- Learning truth: outcome count, first-pass rate, and repair rate.
The new Truth / Action workbench asks Synthesis v2 for a bounded Agent Packet
and renders its gate, evidence, uncertainty, and next steps. Commands are
copy-only. The dashboard does not execute work.
Boundaries preserved
- No new MCP tools.
- No new backend or service.
- No mandatory LLM or cloud dependency.
- No runner subprocess execution moved into the gateway.
- Existing wrappers and full HTTP contracts remain compatible.
- Public policy and generated skills remain non-activating.
Verification
go test -count=3 ./...inservices/gateway-gocargo test -p context_codecincrates- Dashboard test suite, TypeScript check, and production build
- Generated-contract, output-contract, global-installer, open-core boundary,
and public leak audits
The frozen baseline, holdouts, failure evidence, and reproduction commands are
in docs/evals/agent-packet-session-truth-v3.16.0.md.
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About sheawinkler/ContextLattice
Private-by-default memory and context layer for agents with Go/Rust runtime, staged retrieval across fused data backends, and long-horizon context continuity.
Related context
Related tools
Beta — feedback welcome: [email protected]