v0.2.5

Added

• Stable cockpit line — Promoted the public @martinloop/mcp@0.2.5 surface with run triage, compact proof resources, and richer discovery guidance.
• Run-store hardening — Discovery and review flows now degrade cleanly when run-store entries or explicit runsDir paths are missing or unreadable.

Changed

• Release proof lane — Synced root README, MCP release docs, package metadata, smoke scripts, and trusted-publishing workflows to the shipped 0.2.5 contract.

mcp-v0.2.5

@martinloop/mcp v0.2.5

@martinloop/mcp@0.2.5 is the stable cockpit line for the public MartinLoop MCP package. It follows the 0.1.4 operator foundation and 0.2.0 cockpit expansion releases, then extends that surface with run triage, degraded-store hardening, and release-proof polish while keeping martin_run as the single execution entrypoint.

What Shipped

Execution visibility

• added martin_list_runs
• added martin_get_run
• added martin_get_attempt
• added martin_get_verification_results
• added martin_run_dossier
• upgraded martin_run to return richer inspection, verification, and artifact rollups

Context and discovery

• added static resources for server health, recent runs, MCP usage, and publish readiness
• added resource templates for runs, attempts, and verification
• added prompts for governed kickoff, failed-run debugging, and publish-readiness review

Hardening and polish

• cached repeated doctor and run-store inspection work per process
• tightened error taxonomy and selector validation
• added martin_triage_runs
• added martin://runs/triage
• added martin_triage_run_store
• run-store scans now skip unreadable loop records and continue with warnings instead of taking down the whole inspection lane
• kept martin_run, martin_inspect, martin_status, martin_doctor, and martin_preflight backward-compatible
• aligned docs, manifests, and release checks to the full shipped surface

Release Verification Gates

• pnpm --filter @martinloop/mcp lint
• pnpm --filter @martinloop/mcp test
• pnpm --filter @martinloop/mcp build
• pnpm --filter @martinloop/mcp smoke:pack
• pnpm --filter @martinloop/mcp smoke:published:pack
• pnpm --filter @martinloop/mcp verify:release
• pnpm --filter @martinloop/mcp smoke:published

Compatibility Statement

• existing five-tool callers remain supported
• all newly added MCP surfaces are read-only except martin_run
• verification evidence is derived only from persisted Martin run data
• no new write-capable MCP tools are introduced in 0.2.5
• 0.1.4, 0.2.0, and 0.2.5 are the public MCP deliveries in the current release lineage

v0.2.4

Added

• Agent prompt pack — Added the public martin_start, martin_preflight, martin_triage, martin_resume, martin_prove, and martin_release_check prompts plus compatibility aliases.
• Prompt-pack docs — Added guidance for when to use prompt-led kickoff, proof review, and release-check flows in agent hosts.

Changelog

v0.2.3

Added

• Context Diet release surface — Promote npx martin-loop dossier --latest as the compact context packet for handoffs: summary, proof card, budget status, verifier evidence, rollback or artifact evidence, and the next safe action.

Changed

• Refresh the public README, OSS quickstart, and CLI-facing docs so operators and agents can pass smaller evidence-rich run context without digging through raw logs.
• Keep the standalone @martinloop/mcp package on 0.2.0 for this slice; 0.2.3 is a root martin-loop release only.

v0.2.2

Added

• Persisted-run triage in the public release surface — Promote npx martin-loop triage as the fastest way to rank saved runs by failure category, budget pressure, human escalation, and missing verification evidence.

Changed

• Refresh the public README, quickstart, CLI docs, and release notes so the 0.2.2 package surface clearly documents triage-first operator review and degraded run-store warnings.
• Keep the standalone @martinloop/mcp package on 0.2.0 for this slice; 0.2.2 is a root martin-loop release only.

v0.2.1

Added

• Local MCP install profiles — Add minimal, diagnostic, and full-local profile generation for martin mcp print-config and martin mcp install, with minimal as the default local stdio profile.
• Read-only-by-default host setup — Document the default allow-list so local MCP installs start with setup, preflight, run listing, triage, and dossier review before exposing execution.

Changed

• Refresh the public README, quickstart, CLI docs, and MCP host guidance so 0.2.1 clearly shows how to choose the right local MCP profile for Codex, Claude Code, Gemini, and generic hosts.
• Keep the root martin-loop package on the 0.2.x line while leaving @martinloop/mcp@0.2.0 unchanged for this slice.

v0.2.0

Added

• First-value local CLI path — Add npx martin-loop doctor for local readiness checks and npx martin-loop dossier --latest for receipt-style follow-up after a run.
• No-spend demo proof flow — Document the fastest local path from install to proof: doctor, demo workspace, stubbed run, and dossier review.

Changed

• Promote the root martin-loop package to the 0.2.x line while keeping the 0.1.8 Red-Blue Testing and public git-surface guard as baseline features, not new 0.2.0 launch claims.
• Refresh the packed README and quickstart so the published npm surface shows the doctor, no-spend demo, and dossier --latest flow directly.

v0.1.8

What's Changed

• feat(core): add Red-Blue Testing adversarial probe suite by @GobiShanthan
• ci: remove npm auth setup from publish workflows by @Keesan12 in #57
• Clarify README positioning for agent proof workflows by @Keesan12 in #59
• Remove blog export artifacts from the public OSS package by @Keesan12 in #58

🔴🔵 Red-Blue Testing

Adversarial probe suite that runs before a patch is accepted. Six deterministic probes detect patch-level cheating: assertion deletion (T01), silent export reverts (T02), manifest scope creep (T03), context directory poisoning (T07), budget self-reporting (T10), and grounding evasion pragmas (T11).

Three risk tiers:

• baseline — 6-probe sweep, no model call
• high_risk — paranoid 12-probe scan, no model call
• release_critical — paranoid scan + optional Haiku model call for deeper inspection

A single block-severity finding rejects the patch. warn findings are recorded but do not block.

Exported from martin-loop SDK: runRedPhase, shouldAcceptPatch, buildRedFindings, resolveRedBudgetPolicy

Full Changelog: v0.1.7...v0.1.8

v0.1.7

What's Changed

• feat(mcp): prepare v0.2.0 cockpit release by @Keesan12 in #51
• docs(mcp): clean public v0.2.0 release notes by @Keesan12 in #52
• docs: add public release guardrails by @Keesan12 in #53
• docs: sync root readme for v0.1.7 by @Keesan12 in #54
• ci: use trusted publishing for root release by @Keesan12 in #56

Full Changelog: v0.1.6...v0.1.7

mcp-v0.2.0

@martinloop/mcp v0.2.0

0.2.0 turns the Martin Loop MCP server from a governed execution lane into a small local cockpit for reviewing governed agent runs.

0.1.4 introduced the safe operator foundation: check the environment, preflight a contract, run a governed coding task, and inspect saved results. 0.2.0 keeps that contract intact and adds read-only review surfaces so MCP hosts can show what happened after a run: recent runs, one-run dossiers, individual attempts, verifier results, and guided review prompts.

The important safety boundary is unchanged: martin_run is still the only tool that can execute work. The new 0.2.0 additions are read-only inspection and review helpers.

What Changed From 0.1.4

Area	0.1.4	0.2.0
Environment checks	martin_doctor	unchanged
Run preflight	martin_preflight	unchanged
Governed execution	martin_run	unchanged write boundary
Basic run inspection	martin_inspect, martin_status	unchanged
Review cockpit	not included	run lists, dossiers, attempts, verifier results
MCP discovery	tools	tools, resources, resource templates, prompts

Tools

Existing tools:

• martin_doctor
• martin_preflight
• martin_run
• martin_inspect
• martin_status

New read-only cockpit tools:

• martin_list_runs lists recent governed run summaries from the local run store.
• martin_get_run returns a run dossier by loopId or latest.
• martin_get_attempt returns one attempt record by loopId and attemptIndex.
• martin_get_verification_results extracts verifier completion events.
• martin_run_dossier assembles summary, task, budget, attempts, and verification evidence for review.

Resources

• martin://runs/summary
• martin://runs/latest

Resource Templates

• martin://runs/{loopId}
• martin://runs/{loopId}/attempts/{attemptIndex}
• martin://runs/{loopId}/verification

Prompts

• martin_review_run
• martin_triage_failures

Upgrade Notes

• Existing 0.1.4 tool callers do not need to change their current calls.
• Hosts can opt into the new cockpit surface by listing tools, resources, resource templates, and prompts through normal MCP discovery.
• The npm package remains @martinloop/mcp.
• The MCP server name remains io.github.Keesan12/martin-loop.

Verification

The published package was verified from npm after release. The published smoke test confirmed:

• package version @martinloop/mcp@0.2.0
• all 10 tools are discoverable
• both resources are discoverable
• all 3 resource templates are discoverable
• both prompts are discoverable
• local run inspection, status, run listing, dossier generation, and governed stub execution work from the installed npm artifact