Packet28 v0.2.62

This release hardens Packet28's always-on hook path and makes the token-reduction claims easier to audit.

Highlights

Fixed compact hook grep correctness for basic-regex alternation patterns such as fn classify\|Mutation.
Preserved actionable file:line regions in compact hook grep summaries, including Bash PostTool captures.
Kept explicit MCP read_regions ranges inline so agents do not need an extra artifact fetch for requested lines.
Compacted fetched search artifacts to path:line:text content instead of verbose nested match JSON.
Reduced MCP first-load overhead by defaulting the tool catalog to the core Packet28 tool set.
Added Build CI and Codex autofix workflow coverage so failed same-repo CI runs can generate verified draft PRs.
Hardened daemon-backed search e2e setup by building packet28d once per test process.

cargo test --workspace --all-targets
Hook benchmark validation: 95.0% eligible weighted mean reduction, live integrity 5/5.
Native MCP benchmark validation: 85.7% eligible mean reduction, artifact fetch 3/3, fetched search artifacts use path:line:text.
Agentic tokenomics: 91.5% slim-context savings, 76.6% with optional artifacts included, 9/9 feature checks.
Codex autofix dry-run prompt generation and script syntax validation passed.

Assets 2

18 May 20:38

Packet28 v0.2.61

This release makes Packet28 safer and easier for agents to use without custom MCP or hook glue.

Reducer-runner cache entries are now guarded by workspace fingerprints, so out-of-band edits do not replay stale cached command results.
Rust reducer fingerprints include relevant .rs, Cargo.toml, and Cargo.lock content for correctness after source changes.
MCP now exposes packet28.agent_status for setup health, active task state, hook config presence, and reducer cache safety.
Setup and README guidance now document the one-command setup path and clarify that manual MCP JSON-RPC wrappers are not required after setup.

The agentic coding tokenomics experiment now proves a post-edit cargo test --lib rerun succeeds without a synthetic edit hook.
Current experiment result: 7,553 estimated context tokens saved, 93.9% savings, and 9/9 Packet28 feature checks passing.
Release gates require the agentic tokenomics manifest entry and focused stale-cache safety tests before publishing.