@maestria/agent-plugins

0.12.0

Minor Changes

• #432 830b22d Thanks @agustinusnathaniel! - Breaking: @adventurer, @architect, and @builder are removed on every host. The main session now works as one main loop: it does recon, design, and implementation itself, loads the new recon and architecture-plan method skills when a task needs them, and dispatches only @reviewer for an independent check of meaningful changes. The builder's implementation judgment now lives in the always-on rules. On OpenCode the orchestrator can read, search, edit, and run the same shell commands the builder could. On Pi, modes no longer block main-session edits, and sonar blocks edits and subagent dispatch. maestria configure lists only reviewer, and stored model settings for the removed roles are ignored.

Patch Changes

• #430 e445db3 Thanks @agustinusnathaniel! - List the diagnose skill in the Codex and portable manifest descriptions.

• #434 09e9090 Thanks @agustinusnathaniel! - Manifest descriptions and host notes now count the seven bundled skills, including recon and architecture-plan, and describe one reviewer profile instead of several specialists.

• #433 111b7d9 Thanks @agustinusnathaniel! - The shared rules and main loop no longer mention the removed role agents: fail-loud repair now points at the diagnose and architecture-plan skills, and when a host does not allow direct work the main loop (and blitz) report that limit instead of delegating to a specialist that no longer exists.

0.11.0

Minor Changes

• #421 b2c63ca Thanks @agustinusnathaniel! - Directives are generalized for any project and pick up new agent-behaviour habits. The global em-dash ban is gone, and specialists name only the skills maestria ships (diagnose, plan, docs-update, create-pull-request), referring to any other skill by capability when one is installed. Commits follow the repository's own convention (conventional commits when it has none), changesets are treated as one kind of release notes where a repository uses them, and the PR skill no longer contradicts itself about committing evidence files.

  Testing guidance keeps its intent with softer absolutes: choose coverage before implementation instead of backfilling tests, and for complex features prefer end-to-end coverage with a reproducible artifact over many unit tests. A bug's reproduction is kept as a regression test unless an existing test already pins the behavior, and a bug that does not reproduce gets a defensive fix only when the user asks for one, labeled unverified.

  The global rules keep what every agent's report needs: check whether something is already done, say how real each finding is, run the checks its stakes call for and state what was checked, check related work when feedback names one instance of a problem, and correct every place an earlier claim reached. Agents also act on an unprovable precondition only when a wrong guess is cheap and the first step tests it, never call a version latest from memory, start independent work instead of idling, and keep unanswered user questions across compaction. User-facing reporting moves to the orchestrator, so each specialist profile carries about 330 fewer words: answer first with Yes, No, or Not yet, end a message that leaves work running with what is watched, how, and until when, show a draft in full when it changed or the final version is asked for, ask one decision at a time, say what happens if a requested hold is skipped, and treat exploratory questions as discussion.

  The orchestrator keeps autonomous commit, push, and PR but stops at the last allowed step when a user or project says "local only", "don't push", or "no PR", and pushes only to the existing remote. It verifies each premise before it goes into a brief and cites the repository and commit, gives delegates a size budget, points briefs at standing rules the delegate can load and inlines the rest, reads a running delegate's latest messages and head before briefing it, records each PR's single owner where every session sees it, hands ownership over only with green work and a handover of head, CI, and open threads, verifies each brief item with a file and line in a scratch worktree, reshapes the design after two rounds of point fixes, reruns an unrelated CI failure once before reporting it, and keeps a held PR as a quiet open draft. The create-pull-request ready check requires a correctness pass over the whole diff and no unresolved review thread, fixes correctness findings by default, gives merge steps unasked, and re-reads the approval state after every push. Reviewers check duplicate, out-of-order, or retried events and partial failure.

• #426 19b306b Thanks @agustinusnathaniel! - Specialist profiles now embed only the parts of the global rules their role acts on, instead of the whole file. rules.md is split into four parts: Every Agent (floors, reporting, scope, authorization stops), Implementation and Review (testing, comments, visual evidence, docs, acceptance), Writing Changes (branches, commits, hooks, background processes, generated sources), and Main Session (outcome, delegation, bounded repair, precedence and shipping). Adventurer and architect carry the first part, reviewer the first two, builder the first three, and the main session (global rules, Cursor rule, AGENTS.md, and commands) all four. Each specialist profile shrinks by about 440 to 1,260 words; rules still embed because subagents on Claude Code and Kimi do not receive the main session's always-on rules.

• #425 aab1cfc Thanks @agustinusnathaniel! - The bundled plan skill is removed; current models and host plan modes plan well, and the architect still writes phased plans directly. Its durable rules move into the directives: for work that spans sessions, the orchestrator keeps a tracker that closes each task with its evidence and a decision log that records settled decisions as they are made and marks superseded entries instead of rewriting them, and the architect gives every compatibility shim in a migration a removal condition or a reason to keep it. Packages now ship five skills (handoff, iteration-limits, create-pull-request, docs-update, diagnose), and npx skills add agustinusnathaniel/maestria no longer lists plan.

Patch Changes

• #429 0a91f09 Thanks @agustinusnathaniel! - Kimi Code now reads its own session context, kimi.system-prompt.md, through systemPromptPath. It has the same policy, routing, and mode guidance as the shared AGENTS.md but only the Kimi host note, so the other hosts' notes no longer count against Kimi's 32 KB limit. The shared AGENTS.md keeps the Devin, Codex, and Hermes notes.

0.10.1

Patch Changes

• #419 b77d6b7 Thanks @agustinusnathaniel! - On GitHub, agents now upload PR visual evidence with gh pr create --attach or gh pr edit --attach instead of committing screenshots to the branch, and every PR in a stack carries its own evidence. UI changes default to a before/after table (Area | Before | After | What changed).

0.10.0

Minor Changes

• #413 14f6662 Thanks @agustinusnathaniel! - Agents now reproduce a bug with the cheapest failing check before patching it, and report the evidence instead of patching when it does not reproduce. They keep that reproduction as a regression test only when it guards a real production failure existing tests miss, and every kept test names that failure and uses inputs production can produce.

  Plans list their PRs before any code (fixes first, refactors last, shared-behavior changes never riding along with a feature) with each PR's pre-merge and post-merge steps, and record settled decisions in the plan's decision log. PR bodies gain an optional ## Merge steps section, a ready check runs before a PR is marked ready, and the title and body are re-read against the new head after every push.

  Agents treat handoffs and delegate reports as leads to verify, check whether an item is already done before calling it pending, rate each finding by how real it is and what to do about it, ask one decision at a time with a concrete scenario and a recommendation, and put a requested draft last. Builders copy the shape of sibling files and challenge every guard against evidence, reviewers flag guards for impossible cases, and the orchestrator numbers brief items, gives each PR or branch one pushing owner, and verifies pushed work before reporting it.

• #406 1f82d9b Thanks @agustinusnathaniel! - Delivery ownership moves from the global rules into the orchestrator: workflow modes, autonomous commit, push, and PR creation, PR slicing, and the PR skill pointer. Commands and always-on routing context still carry the complete policy, while each specialist profile and specialist request drops about 250 words of rules it could not act on. A specialist invoked directly, without a workflow command or the orchestrator, still follows commit and protected-branch rules but no longer pushes or opens PRs without asking. On OpenCode, the plugin keeps sending the delivery rules to OpenCode's own agents such as build and plan, so only maestria's specialist subagents drop them.

• #408 952116f Thanks @agustinusnathaniel! - Specialists go from seven to four: adventurer, architect, builder, and reviewer. The retired diagnose, planner, and writer agents become methods: the new bundled diagnose skill (adventurer runs a read-only diagnosis, builder applies the fix) and plan skill (loaded by architect), while builder takes over documentation work with the docs skills. Every package now ships the diagnose and plan skills. maestria update removes retired Codex profiles; on Pi, delete diagnose.md, planner.md, and writer.md from ~/.pi/agent/agents if you no longer want them listed.

• #412 dad26aa Thanks @agustinusnathaniel! - Agents now keep your work safe and follow a few habits you kept asking for: they never git reset, restore, clean, revert, or force-push existing work unless you asked for that operation; they let repository hooks run instead of bypassing them; and the delivery owner waits for CI after each push and fixes or reports failures before calling delivery complete. They answer your direct questions before continuing, never shrink requested work by human time estimates, re-check their claims when you say "cmiiw", and fix pre-existing failures in the files and checks a change touches instead of stopping at them.

Patch Changes

• #405 f04e56d Thanks @agustinusnathaniel! - Agent rules and specialist prompts are shorter: duplicated rules, step lists that only restated the obvious, generic review checklists, and stale lines are gone, while every behavioral obligation is kept, moved to the skill that already owned it, or retired with a stated reason. Specialist profiles and workflow commands each carry about 160 to 435 fewer words.

0.9.0

Minor Changes

• #396 e62199d Thanks @agustinusnathaniel! - Plans that span sessions keep a tracker with evidence per task, and the orchestrator treats done marks as claims: it audits every done item in scope before resuming when the user reports missing work, another session wrote the tracker, or a done task fails.

Patch Changes

• #398 fd6defa Thanks @agustinusnathaniel! - Replace the plugin logo shown by Cursor and Codex with the new maestria mark: the maker's solid diamond meeting the checker's ring, in the docs site's green and brass.

0.8.0

Minor Changes

• #390 bc0f4f9 Thanks @agustinusnathaniel! - Bundle the create-pull-request and docs-update skills in every maestria package, so the global rules never point an agent at a skill it does not have. @maestria/agent-plugins now ships four skills on every plugin host (Claude Code, Codex, Cursor, Kimi Code, Devin, ZCode, OMP, Hermes), @maestria/pi loads them from its skills/ folder, and the @maestria/opencode plugin registers its bundled skills/ folder through skills.paths. The copies are generated from packages/core/agent-directives/skills/, and npx skills add agustinusnathaniel/maestria now offers all four plugin skills to runtimes maestria does not package. If you installed either skill separately, remove that copy to avoid a duplicate (maestria update removes copies the CLI installed).

• #390 bc0f4f9 Thanks @agustinusnathaniel! - Consolidate Codex, Claude Code, Cursor, Kimi Code, Devin, ZCode, and portable Agent Plugins resources into one declarative plugin package. Native installers retain platform identifiers and managed ownership while fetching the shared archive. Remove the former standalone source packages; native runtime integrations remain separate.

Patch Changes

• #390 bc0f4f9 Thanks @agustinusnathaniel! - Install @maestria/agent-plugins straight from the package root, following the PostHog and Cursor pstack layout: the host manifests sit at the root and select shared agents/, commands/, rules/, and skills/ folders, with no nested installation bundle or duplicate component tree. Each specialist profile and workflow command carries its complete role, routing, and global policy, composed from canonical core directives at sync, so no host depends on preloading role or policy skills. Codex profiles are generated TOMLs with complete instructions. One root AGENTS.md carries the global policy and routing for Devin, Kimi Code (through systemPromptPath), Codex (installed by the CLI), and Hermes, with a short note per host; only OMP and Cursor keep their own rule files because their formats differ.

  The documentation now has one Installation page and one Usage page with a section per host, including separate CLI and desktop setup for Codex and Claude Code. The former per-host page URLs redirect to those sections.

  Codex profile updates keep model settings when a prompt contains TOML-like text, and sync dry runs that hit a processing error now fail preflight before anything is written.

• #390 bc0f4f9 Thanks @agustinusnathaniel! - Close behavior gaps behind frequent user corrections. Every brief and handoff now restates the original goal and acceptance criteria (pointing to an owning OpenSpec or Spec Kit spec when one exists), and completion claims are checked against them rather than the latest sub-step; reviewers flag requirement drift. Agents read user-supplied references and the original requirement in full, ask before rewriting content the user authored, meet an explicit request for comprehensive work instead of trimming to the smallest route, fix pre-existing issues in scope only when asked (otherwise report them), avoid collateral edits and deletions, take already-authorized routine next steps (commit, push, PR) without asking again, and explain results in plain language. The optional spec-contract skill is retired; its intent now lives in the always-on rules and the generated handoff skill.

• #390 bc0f4f9 Thanks @agustinusnathaniel! - Generate the handoff and iteration-limits utility skills from their global rules sections (Delegation and Context, Bounded Repair and Fail-Loud Behavior) instead of separate hand-written text, so the skills always match the policy. The global rules now state the stop conditions only the skill carried before: define a verifiable termination condition before looping, treat only [fix] blockers as repair work, and run final verification then stop instead of restarting a full review for a small fix. Pi's handoff skill description no longer claims a seven-field contract.

• #390 bc0f4f9 Thanks @agustinusnathaniel! - Ship one profile per specialist in agents/<role>.md for every Markdown host, with each host's read-only control in shared frontmatter, and one set of workflow commands in commands/. Per-host agent and command copies, the dispatcher-only orchestrator profile, and Devin's mode skill exports are removed; Devin routes mode keywords through its always-on policy. Codex profiles are generated beside the Markdown as agents/maestria-<role>.toml.

  Fix Kimi read-only profiles, which denied WriteFile and StrReplaceFile (names current Kimi does not use) instead of Write and Edit. ZCode research and review profiles now deny Write and Edit as well. The architect is now read-only on every host, matching OpenCode, Pi, and Codex: it returns ADR content for @writer to persist. Devin no longer auto-loads the OMP-specific root rule beside its own always-on AGENTS.md. Remove an exclamation-mark-before-backtick sequence from the global rules that ZCode and Claude Code commands would run as inline shell, drop the unused name key from command frontmatter, and select ZCode agents and commands by directory, the only form ZCode reads. Profiles drop model: inherit, which matches the hosts' default. maestria update cursor carries a configured model from an earlier Cursor-only profile copy onto the shared profile.

  Release the CLI and plugin together: the updated CLI reads Codex profiles from agents/maestria-<role>.toml and configures Cursor models in agents/<role>.md, so run maestria update cursor before maestria configure cursor on an existing install.

• #390 bc0f4f9 Thanks @agustinusnathaniel! - Trim duplicated directive text without changing behavior: each role relies on the single human-facing output rule in the global rules instead of restating it, the orchestrator drops its copy of the mode table and the delivery and visual-evidence paragraphs the global rules already carry, the reviewer's unused lens catalog becomes one scope rule, and non-negotiable markers stay on safety and verification rules only. The docs-update skill pointer now fires before finishing any behavior, interface, configuration, or workflow change, and the visual-evidence rule names rendered documentation sites and CLI output and points to the create-pull-request capture procedure for every role.