v0.28.1

Stability hotfix for the semantic/rerank optional dependency.

Fixed

• A retrieve call with semantic/rerank no longer kills the MCP server when the optional @huggingface/transformers package is missing: the rejection previously escaped the stdio serve loop unhandled, terminating the process mid-call so agents saw an infinite spinner instead of an error. Retrieve failures now return an MCP isError tool result the agent can read and react to, and the serve loop is hardened so no handler rejection can tear down the server.
• A project-local npm install @huggingface/transformers now actually enables semantic/rerank: npx-launched and globally installed servers resolve the optional package from the project root (derived from the graph path) in addition to madar's own installation.
• Failed semantic model loads no longer poison the pipeline cache: a rejected load is evicted, so installing the package and retrying succeeds without restarting the server.

Added

• Semantic/rerank capability gating in the retrieve tool schema: tools/list omits the semantic, semantic_model, rerank, and rerank_model fields when the optional package is not resolvable, so agents never request a capability the machine lacks.
• madar doctor now reports semantic/rerank availability with the exact enable command, without affecting overall health status.
• Semantic model loads are bounded by a timeout (MADAR_MODEL_LOAD_TIMEOUT_MS, default 120s) so a stalled first-use model download cannot block the serial stdio request loop indefinitely.

Full notes: CHANGELOG 0.28.1

v0.28.0

What's Changed

• docs: refresh README landing page by @mohanagy in #506
• docs(contributor): contributors readme action update by @github-actions[bot] in #507
• feat: expand benchmark suite targets by @mohanagy in #508
• docs: publish public benchmark receipts by @mohanagy in #509
• chore(deps-dev): bump the test-tooling group with 2 updates by @dependabot[bot] in #510
• chore(deps): bump neo4j-driver from 6.0.1 to 6.1.0 by @dependabot[bot] in #511
• chore(deps-dev): bump @types/node from 25.9.1 to 25.9.2 by @dependabot[bot] in #513
• chore(deps-dev): bump vite from 8.0.14 to 8.0.16 by @dependabot[bot] in #512
• sync by @mohanagy in #516
• fix: enforce strict runtime proof for benchmarks by @mohanagy in #515
• feat: support scoped benchmark repo roots by @mohanagy in #517
• fix: harden partial runtime proof guidance by @mohanagy in #518
• fix: harden explain-runtime proof gates by @mohanagy in #519
• fix: complete proof-backed public explain-runtime full-win rows by @mohanagy in #520
• Merge next into main for 0.28.0 release by @mohanagy in #521
• Merge pull request #521 from mohanagy/next by @mohanagy in #523
• chore: prepare 0.28.0 release by @mohanagy in #522

Full Changelog: v0.27.9...v0.28.0

v0.27.9

Added

• The next-track adoption bundle is now the stable 0.27.9 release: public benchmark suite and language fixtures, one-command madar try, opt-in telemetry, verified agent quickstarts, the design-partner loop, and the proof-first launch checklist now ship on main.

Changed

• Stable now carries the git-backed freshness model across madar pack, madar prompt, madar handoff, madar doctor, and madar status.
• Native-agent compare keeps stricter task-bounded prompts, separate baseline/Madar artifacts, clearer install and trace attribution, and more honest timeout evidence.
• Generated Claude and Cursor MCP configs now launch the installed madar CLI directly, and madar install <platform> works alongside --platform.

Fixed

• Windows Claude installs now use a generated local .claude/madar-user-prompt-submit.cjs script instead of an oversized inline hook.
• Windows native-agent --exec commands now honor the expected cmd.exe contract.
• Compare trace classification no longer mistakes deferred ToolSearch Madar selection for broad exploration.

v0.27.9-next.0

Added

• Adoption proof surfaces now cover the full next-track rollout bundle: Madar now ships a public multi-repo benchmark suite and language fixtures, the one-command madar try first-run flow, verified agent-specific quickstarts and smoke-tested install docs, a design-partner feedback loop with reproducible receipts, and the proof-first launch/distribution checklist. Closes #469, #470, #472, #473, and #474.

Changed

• Public adoption messaging and instrumentation are now aligned across the next track: the README/npm discovery path now reflects the Madar positioning and legacy-name cleanup, prerelease docs stay on blob/next, and the opt-in telemetry funnel records coarse install/generate/pack/prompt/doctor/status/compare adoption stages without collecting source text or paths. Closes #467, #468, and #471.

Fixed

• Graph freshness and stale-context guarantees are now scoped and git-backed: madar pack, madar prompt, madar handoff, madar doctor, and madar status now report whole-repo vs selected-context freshness using the graph build revision plus current git diff as the source of truth on git workspaces, with privacy-safe governance receipts and tighter stale-context behavior for follow-up flows. Closes #477 and #479.

v0.27.8

Added

• README reference content now has dedicated docs for Pack Schema v1, adaptive context packs, execution slices, MCP tools, installers, strict profiles, MCP Registry metadata, command reference, and discovery rules.

Changed

• README is now a shorter npm-facing landing page focused on onboarding, agent choices, core surfaces, evidence boundaries, and documentation links.
• Review feedback clarified that CLI-only commands can use the generated graph without agent installation, corrected the SPI version reference, and loosened brittle docs assertions.

npm publishing is handled separately.

v0.27.7

What's Changed

• [Roadmap] Publish and validate MCP Registry metadata by @mohanagy in #438
• Improve first-run onboarding docs by @mohanagy in #439
• Tighten release hygiene verification by @mohanagy in #440
• Add doctor lint for agent instruction drift by @mohanagy in #441
• docs: add install compatibility guide by @mohanagy in #442
• feat: add local proof report command by @mohanagy in #443
• fix: harden MCP security boundaries by @mohanagy in #444
• [Roadmap] Publish implementation-task outcome evidence by @mohanagy in #445
• feat: add share-safe handoff command by @mohanagy in #446
• feat: expand benchmark suite matrix by @mohanagy in #447
• feat: add pack governance receipts by @mohanagy in #448
• feat: add opt-in local telemetry by @mohanagy in #449
• feat: deepen framework retrieval coverage by @mohanagy in #450
• feat: expose session efficiency diagnostics by @mohanagy in #451
• docs: position Madar for review and security workflows by @mohanagy in #452
• docs: sharpen ICP and product promise by @mohanagy in #453
• docs: define team and enterprise offer by @mohanagy in #454
• docs: document share-safe dashboard decision by @mohanagy in #455
• docs: define language expansion decision by @mohanagy in #456
• docs: research distribution channels by @mohanagy in #457
• docs: add design-partner workflow loop drafts by @mohanagy in #458
• docs: add federation flagship receipt by @mohanagy in #459
• Sync main into next by @mohanagy in #460
• Prepare 0.27.7-next.0 release by @mohanagy in #461
• Fix compare validation safety followups by @mohanagy in #462
• Prepare 0.27.7-next.1 release by @mohanagy in #463
• Next by @mohanagy in #464
• Prepare 0.27.7 release by @mohanagy in #465

Full Changelog: v0.27.6...v0.27.7

v0.27.6

What's Changed

• fix: surface runtime primary path nodes in answer-ready packs by @mohanagy in #401
• [P2] serialized_budget.enforced is false on 0.27.4 explain packs; packs ship 2× over their declared budget by @mohanagy in #402
• Release 0.27.5 by @mohanagy in #403
• Fix SPI readiness detection by @mohanagy in #409
• Fix PR 409 review follow-up by @mohanagy in #410
• Fix core profile prompt contract by @mohanagy in #411
• Boost prompt identifier retrieval by @mohanagy in #412
• Add implement outcome scoring to compare by @mohanagy in #413
• Remove legacy URL ingest support by @mohanagy in #414
• chore: release 0.27.6 by @mohanagy in #415

Full Changelog: v0.27.4...v0.27.6

v0.27.4

What's changed

Fixed

• runNativeAgentArmWithTimeout abort/settle race: after aborting a timed-out arm, a 200 ms grace window lets the runner fully settle before timed_out is recorded — preventing a partially-completed result from being misclassified. Truly stuck arms still resolve via the grace timeout.
• assessNativeAgentPromptContract returns not_measured for post-pack broad exploration: broad exploration after a pack call can be justified by missing context or coverage gaps that the trace model doesn't capture, so marking it violated caused false failures.
• suggestBenchmarkGraphScope normalises absolute source_file paths: project root is inferred and stripped before scope extraction, so the reported scope is the project name instead of the first filesystem segment.
• candidateScopes in MCP response evidence handles absolute and Windows paths: scope detection now finds the segment immediately before a generic directory marker (src, test, lib, …) rather than splitting on the first /.

---

Full changelog: https://github.com/mohanagy/madar/blob/main/CHANGELOG.md#0274---2026-05-29

v0.27.3

Added

• Explain packs now default to compact answer-ready output: madar pack, MCP context_pack, and compare prompt packs can return bounded answer-ready JSON by default while preserving full diagnostics behind verbose mode. Closes #376.
• Runtime-generation fixtures cover the full report path: pack-quality regression coverage now includes the runtime-generation explain/report flow so spine selection is tested against the real orchestrator, storage, and quality-gate handoff. Closes #377.

Changed

• Installed-agent guidance is stricter about Madar-first exploration: generated Claude/Codex/Copilot/Cursor/Gemini instructions now tell agents to inspect evidence.pack_confidence, recommended_first_read, and evidence.agent_directive, and to try one focused Madar follow-up before broad raw search. Closes #378.
• Native-agent benchmark outcomes now gate full wins on regressions: compare reports include benchmark_outcome checks for routing/tool/latency, total tokens, fresh-token usage, provider cost, and turns, so faster routing with worse cost or fresh-token behavior is marked as a partial win instead of a full win. Closes #379.

NPM publish is intentionally left for the maintainer.

v0.27.2

Changed

• Native-agent compare now explains missing verbose traces explicitly: report.json records whether Claude verbose trace data was available, and CLI/docs clarify that --verbose is required for MCP-call attribution while provider usage can still come from --output-format json. Closes #368.
• Native-agent traces distinguish pre-Madar broad exploration: verbose compare reports now classify broad tools before the first Madar MCP call as madar_invoked_after_broad_exploration. Closes #369.
• Native-agent benchmark summaries separate routing wins from token proof: valid attributed runs now carry claim_assessment so fewer tools/faster latency can be reported separately from provider-token reduction, and fresh-token regressions keep token-reduction claims marked not_proven. Closes #370.

npm publish is intentionally left for @mohanagy.