x9 Agent Skills v3.2.0

Highlights

• x9-idea-critic now uses provider-level claude and gpt routes with centralized model defaults: the official fable alias and GPT-6 Astra, both at high effort, across Claude Code, Codex, and OpenCode.
• x9-browser-session hardens its verified Edge adapter by binding remote debugging to localhost, moving it to port 9223, checking that the endpoint belongs to the expected automation profile, and preserving authenticated profile state without isolated browser contexts.
• x9-agent-instructions now preserves scope-critical user wording in continuation briefs and retains targeted behavioral guidance only when observed failures or current model documentation support it.
• Shared global instructions now require current primary evidence for fast-changing technical claims, batch independent read-only work, favor direct execution, and inherit the primary model for substantive Codex/OpenCode delegation.
• x9-opencode-sessions now finds the standard per-user OpenCode V2 installation when automation runs with a minimal PATH.

Install / update

• Existing users of the documented Edge adapter must update the launcher, MCP endpoints, and agent-edge wrapper from port 9222 to 9223; the wrapper now also verifies the owning browser process and profile.
• x9-idea-critic requires Claude CLI with Fable access in all three runtimes. OpenCode users also need the documented general-purpose astra-high subagent; the skill does not modify local configuration automatically.
• Codex/OpenCode users adopting model inheritance should remove lower-cost subagent defaults from their local configuration; the published instructions do not change runtime model settings automatically.

Compatibility

• x9-idea-critic retains the gpt mode. Official Claude aliases follow the service's current target, and every Claude CLI result must identify the selected critic model.
• OpenCode session inspection no longer depends on shell startup files exposing ~/.local/bin.
• x9-agent-instructions regression tests now resolve repository files correctly when launched through an installed skill symlink.
• Obsolete maintainer planning and audit snapshots were removed; no shipped skill or installation artifact depends on them.

Breaking changes

• x9-idea-critic replaces the model-specific opus mode with the provider-level claude mode. Its Claude Code Claude route now requires Claude CLI instead of optionally using a native Opus agent.
• The documented Edge adapter now uses CDP port 9223; installations copied from the previous 9222 recipe must update the launcher, MCP configuration, and agent-edge wrapper together.

x9 Agent Skills v3.1.0

Highlights

• Telegram tasks now default to the authenticated Telegram Web session instead of the macOS app or Computer Use, while explicit desktop-app requests remain supported.
• x9-opencode-sessions now preserves large Sessions API responses instead of receiving truncated JSON when the OpenCode CLI writes to a captured pipe.
• Added x9-decision-map for durable, pre-plan records of material decisions, sharp open questions, fog, dependencies, and the next evidence-gathering frontier.
• Added x9-task-graph for turning an approved implementation-ready plan into dependency-aware, tracker-ready vertical work items with complete source coverage.
• Added x9-architecture-scout for HTML architecture reports that compare plausible directions and expose boundaries, ownership, coupling, and change hotspots without changing production code.
• Added x9-explain-again for repairing explicit confusion about the immediately preceding substantive answer without turning a new explanation request into a false reference.
• x9-agent-instructions now requires completion bars to account for known enumerable sets and makes branch-specific context pointers conditional and testable.
• x9-skill-creator now classifies invocation separately for every claimed runtime and checks the ownership boundary before it creates or judges a skill owner.
• x9-context-files-generator now keeps root-local rules in root AGENTS.md while assigning branch-specific procedures to profile documents with verified conditional pointers.
• x9-agent-instructions and x9-context-files-generator now distinguish model invocation in each supported runtime from portable structural compatibility.
• The English and Russian skill catalog now state that x9-architecture-scout writes an HTML report and describe x9-task-graph's runnable frontier without implying a queue.
• Published global instructions now route Telegram work through the authenticated web session, remove the former mandatory find-docs route, and give Codex explicit batching, descendant-waiting, and worker-communication rules.

Install / update

• Update x9-opencode-sessions to apply the large-transcript fix; its command interface is unchanged.
• Full-plugin users receive the four new skills on update. Selective installations can add only the process that matches the problem; x9-task-graph still requires an approved plan.
• x9-architecture-scout requires python3 for structural report validation; x9-diagrams is optional and used only when an architecture diagram is needed.

Compatibility

• Long OpenCode sessions can now be inspected without transcript degradation caused by the CLI pipe buffer.
• The new skills are structurally compatible with Claude Code and Codex; their stated boundaries remain in force: decision maps are not implementation plans, task graphs do not publish or implement work, architecture scouting is read-only, and explanation repair needs a recoverable immediately previous answer.
• The global instruction examples no longer force every library-documentation request through find-docs; installations that need that policy should retain it as a local rule.

Breaking changes

None.

x9 Agent Skills v3.0.0

Highlights

• Native worker briefs now use English across OpenCode, Claude Code, and Codex while preserving load-bearing source wording and returning user-facing results in the user's language.
• Global instructions now carry a compact worker-brief contract that requires non-derivable context, proportional evidence, scope fidelity, and specialist-prompt precedence without loading the full x9-agent-instructions skill before dispatch.
• Applied the August repository audits: package metadata now stays synchronized across Claude Code and Codex, published global cores are checked in CI, installed global files are checked byte-for-byte before release, and documented package checks match the automated gate.
• Hardened x9-opencode-sessions metadata filtering so session locations remain available for precise selection while internal skill text stays private.
• Documented the repository-wide package validation and release gates, including their evidence, blocking conditions, and known coverage gaps.
• x9-idea-critic now runs every selected Opus and GPT critic at high effort, assigns every full critic its declared lens, and uses sealed evidence packets instead of broad source access. Claude CLI routes disable inherited customizations and tools, Codex sealed runs reject re-enabled hooks, and native routes distinguish accepted controls from unavailable post-run telemetry.
• x9-idea-critic now ends critiques with any load-bearing questions followed by a direct decision, one immediate action, and an observable condition for proceeding, revising, or stopping.
• Added x9-onboarding to check whether installed or updated x9-* skills are ready, distinguish check coverage from readiness, and provide a manual checklist without changing the environment.
• Added an OpenCode route to x9-idea-critic, combining fresh Claude CLI sessions for Opus critics with explicitly identified native GPT subagents.
• Added x9-opencode-sessions for bounded OpenCode V2 session discovery, text-only message inspection, activity and completion checks, and preview-first cross-branch coordination.
• Split the installable global instructions into dedicated OpenCode, Codex, and Claude Code files while preserving their shared personal core.

Install / update

• Full-plugin users receive x9-opencode-sessions on update. Individual installations require Python 3 and the OpenCode V2 opencode2 CLI.
• Full-plugin users receive x9-onboarding on update. Selective installations should include it when the user wants an installation-readiness check.
• Existing users can link each harness to its matching file under global-files/<harness>/; back up local global instructions before replacing them with links.

Compatibility

• Changes to production package/release gates and release workflows now require Unreleased coverage, while test-only and auxiliary files remain exempt.
• Skill validation accepts a YAML boolean only for metadata.internal; string values and booleans on other metadata keys are rejected.
• Onboarding declarations now reject requirement runtimes that are not listed in their owning skill's supported_runtimes.
• x9-idea-critic now supports independent OpenCode routes when Claude CLI exposes the required Opus controls or the live subagent catalog explicitly identifies a GPT-family agent; default and full runs still require every selected route for a complete result.
• x9-opencode-sessions is OpenCode V2 only. It uses native subagent for the current parent session's children and the V2 Sessions API for other OpenCode branches; Claude Code, Codex, and cross-harness sessions are unsupported.
• x9-browser-session now supports OpenCode through the live chrome-devtools MCP, with the same task-page, private-surface, and unattended fallback safeguards as its existing routes.
• x9-onboarding has read-only adapters for OpenCode, Claude Code, and Codex. Other runtimes report PARTIAL coverage and BLOCKED readiness; it never installs dependencies, authorizes accounts, changes files or configuration, or restarts a runtime.
• The former top-level global-files/AGENTS.md and global-files/CLAUDE.md paths were replaced by harness-specific paths.

Breaking changes

• x9-onboarding introduces no breaking changes.
• Skills that declare metadata.internal must use the YAML boolean true instead of the string "true".
• Consumers of the published global files must update references to the new harness-specific paths.

x9 Agent Skills v2.4.0

Highlights

• Made x9-browser-session serialize Avito browsing across agents and controllers,
  reuse captured listing data, passively re-inspect transient IP or security checks
  after five seconds, and stop without escalating controller fallbacks when the check
  persists or Avito returns an explicit throttling or access-control signal.
• Synchronized the published Codex global-instruction example with the current safe
  reconciliation rules for non-idempotent task creation.

Install / update

• Full-plugin users can update in place. Selective installations should update
  x9-browser-session; no skill-name or configuration migration is required.

Compatibility

• Browser routing outside Avito is unchanged. Avito work now runs serially and returns
  DEGRADED instead of switching controllers when throttling or access controls persist.
• Existing global-instruction examples remain opt-in references; the new Codex task
  reconciliation rules do not modify personal runtime configuration automatically.

Breaking changes

None.

x9 Agent Skills v2.3.0

Highlights

• Added an extensible chrome-devtools MCP allowlist, initially containing avito.ru; matching domains bypass the browser extension in both runtimes and preserve agent-edge only as the unattended fallback.
• Synchronized the published global-instruction examples with the current question-format and browser-retrieval fallback rules while keeping their shared core byte-identical.
• Made x9-idea-critic turn upheld criticism into a self-contained stronger proposal, with a change map and verdict-specific handling for revised, surviving, killed, or unproven ideas.
• Hardened x9-loop-engineering with conditional patterns for authoritative derived state, recoverable public lifecycle mutations, version-scoped continuity, task-first self-repair, and evidence-bound realized execution traces.
• Made repository-backed diagram creation source-first, preserve existing structure and supported fonts until rendered proof passes, and use a clipboard round trip for faithful browser verification.
• Tightened Excalidraw structure checks for escaped control tokens, native soft wrapping, semantic hard breaks, deleted bindings, and target-version normalization.
• Strengthened rendered diagram evidence with content-viewport fit checks, post-normalization stability, explicit whole-view versus scrollable delivery, and protection from blocking native choosers.

Install / update

• Full-plugin users can update in place; no data or skill-name migration is required.
• Selective Codex installations that use an allowlisted browser target need the chrome-devtools MCP route from the x9-browser-session setup guide.

Compatibility

• Browser routing outside entries in the chrome-devtools MCP allowlist is unchanged; the list currently routes avito.ru through MCP in Codex as well as Claude Code.
• Existing x9-idea-critic modes and provider routes are unchanged; successful critiques now include an evidence-grounded next version of the idea.
• Existing simple and linear workflow guidance is unchanged; the new loop recovery, versioning, self-repair, and lineage patterns apply only when those mechanisms exist.
• Existing Excalidraw scenes remain supported, but invalid control text, dangling bindings, unstable wrapping, or unsupported hard line breaks now fail the stricter checker until corrected or explicitly allowed.

Breaking changes

None.

x9 Agent Skills v2.2.0

Highlights

• Renamed x9-okf-adapt to x9-okf-docs and reduced its always-loaded instructions to a short route selector; maintenance, adoption, and migration details now load only when needed.
• Updated the skill for OKF v0.2 with separate official and curated validation profiles, reserved-file handling, truthful generated metadata, and read-only-by-default v0.1 migration.
• Added incremental --touch migration so substantively edited documents converge to OKF v0.2 without scanning or rewriting the rest of a repository.
• Made x9-okf-docs the canonical owner of version-specific OKF rules. On the first OKF write, it replaces a copied version, field contract, or former skill name in AGENTS.md while preserving repository-specific scope.
• Made durable research read generated.at first while retaining timestamp as a legacy fallback.

Install / update

• Existing OKF v0.1 repositories require no immediate rewrite. Normal work can migrate named documents through --touch and lazily repair stale repository instructions; bulk audit remains optional.
• Full-plugin users receive x9-okf-docs on update. Individual-skill users must remove x9-okf-adapt, install x9-okf-docs, and may leave repository instructions to self-repair on the first OKF write.

Compatibility

• OKF v0.1 remains readable through documented fallbacks. Incremental migration keeps timestamp synchronized by default and requires Python 3, Ruby/Psych, and an explicit actor.

Breaking changes

• The curated profile now writes generated.by and generated.at instead of creating legacy timestamp; callers applying manifests must pass --actor. Official OKF validation is available separately through --profile okf.
• The discoverable skill name changed from x9-okf-adapt to x9-okf-docs; the former name is not retained as a duplicate alias.

x9 Agent Skills v2.1.0

Highlights

• Renamed x9-excalidraw-diagrams to x9-diagrams and expanded it into one research-backed method for choosing, creating, reviewing, and converting native Excalidraw, Mermaid, and draw.io diagrams.
• Added format-specific Mermaid and draw.io adapters while preserving the native Excalidraw checker, current font/binding rules, and regression tests.
• Kept the distributed skill focused on operational guidance by removing the maintainer-only research ledger.
• Added a constraint-based format matrix, explicit ownership boundary with specialized Mermaid-only skills, and separate structural plus rendered completion gates for all supported formats.
• Split completion evidence by selection, creation, conversion, and review; tightened degraded-result reporting and the Excalidraw basic-subset checker without requiring both supported fonts in every scene.

Install / update

• Full-plugin users can update in place and invoke x9-diagrams. Individual-skill users must remove x9-excalidraw-diagrams and install x9-diagrams; the old name is not retained as an alias.

Compatibility

• x9-diagrams remains portable across Claude Code and Codex. Complete proof requires Python 3 plus a faithful Excalidraw renderer, a target-compatible Mermaid renderer, or a compatible diagrams.net editor/exporter for the selected format; missing visual proof is reported as DEGRADED.

Breaking changes

• The public skill name and path changed from x9-excalidraw-diagrams to x9-diagrams. Direct invocations, selective-install records, and local symbolic links using the old name must migrate to the new name.

x9 Agent Skills v2.0.0

Highlights

• Made x9-idea-critic select the promised provider family explicitly, isolate volatile Claude Code and Codex mechanics in runtime adapters, and declare a reproducible lens/provider matrix for full mode.
• Made x9-research leave an inspectable independent-review disposition, assign browser and delivery rules to one owner, support correct HTML language metadata, and validate deterministic Markdown/HTML requirements with a bundled regression-tested checker.
• Made x9-skill-creator structural validation runtime-aware for portable Agent Skills, Claude Code, and Codex, with stricter frontmatter constraints, duplicate-key detection, and complete bundled-resource reachability checks.
• Updated the Claude Code adapter for current trigger and frontmatter behavior, and removed duplicated judge-checklist cardinality from batch audits.
• Made x9-agent-instructions reviews report-first, pair every retained change with a concise practical rationale, and apply only explicitly approved recommendations.
• Made x9-skill-creator load x9-agent-instructions as the required rubric for agent-facing prose while keeping skill-audit reporting self-contained.
• Made x9-agent-instructions reference installed skills by discoverable name, reject machine-specific SKILL.md paths, and omit redundant requests to read repository context already injected by the runtime.
• Made x9-context-files-generator load x9-agent-instructions as a bounded companion rubric for agent-facing prose while keeping repository structure, preservation, and README work under its own contract.
• Added x9-excalidraw-diagrams, a research-backed workflow for native, editable Excalidraw scenes with fit-to-view legibility, role-based typography, controlled routing, a render-and-inspect completion gate, and structural detection of clipped explicit text lines.
• Made persistent x9-skill-creator audit reports follow the user's requested or request-carrier language through one batch-wide Report language tag while preserving exact technical literals, source quotations, and validator output.

Install / update

• Existing x9-idea-critic and x9-research users can update in place; no dossier migration is required.
• Structural validation now accepts repeatable --runtime portable|claude|codex flags; omitting the flag keeps the portable profile as the default.
• The full plugin already installs both required skills. Users who copy individual skills should install x9-skill-creator together with x9-agent-instructions.
• Users who copy x9-context-files-generator individually should install x9-agent-instructions alongside it for complete agent-file creation and review.
• The new skill is included in the full plugin and can also be installed individually as x9-excalidraw-diagrams.

Compatibility

• x9-idea-critic still supports Claude Code and Codex, but unavailable provider-family selection now fails the affected route instead of silently substituting another model.
• x9-research HTML delivery now requires Python 3 for deterministic pair validation and accepts the document's actual BCP 47 language tag instead of forcing Russian metadata.
• x9-skill-creator now declares its existing Python 3 and Ruby/Psych validation dependency explicitly. Existing portable skills remain valid, while runtime-specific frontmatter must be checked with the matching profile.
• Without x9-agent-instructions, x9-skill-creator can still run structural checks but reports full Create, Audit, and Fix work on agent-facing prose as degraded.
• Without x9-agent-instructions, x9-context-files-generator keeps README, repository-evidence, and structural work available but reports agent-file instruction quality as degraded.
• x9-excalidraw-diagrams is portable across Claude Code and Codex. Its structural checker requires Python 3; visual proof requires a faithful Excalidraw renderer or editor and is reported as degraded when neither is available.

Breaking changes

• x9-research HTML delivery now requires Python 3; without it the Markdown evidence remains usable, but HTML validation is reported as degraded instead of complete.
• Validation is intentionally stricter: unsupported runtime fields, non-string portable metadata values, overlong compatibility, and unreachable bundled resources now fail instead of passing silently.
• A standalone x9-skill-creator installation now requires x9-agent-instructions before it can claim a complete instruction-quality review.
• A standalone x9-context-files-generator installation now requires x9-agent-instructions before it can claim a complete or clean agent-file result; README-only work is unchanged.
• No new breaking changes are introduced by x9-excalidraw-diagrams.

x9 Agent Skills v1.6.0

Highlights

• Aligned the English and Russian README copy for Codex skill discovery, browser setup, context-file generation, idea review, delegation limits, and prompt-engineering references.
• Synchronized the published global instruction examples with the audited Claude Code and Codex policies, including shared skill routing and a self-directed execution contract.
• Made the Codex example retain undocumented subagent capabilities only after bounded live probes, and aligned its routing claims with verified agent_type, fork_turns, and service_tier behavior.
• Made x9-browser-session the canonical owner of browser route selection and removed its circular dependency on global Claude Code instructions.

Install / update

• Users who copied the published global instruction examples should resynchronize them after updating; plugin and skill installation routes are unchanged.

Compatibility

• Existing browser routes, skill names, and installation routes remain compatible. Global subagent policy now distinguishes verified runtime controls from command-internal routing that the caller cannot set.

Breaking changes

• None.

x9 Agent Skills v1.5.0

Highlights

• Made x9-agent-instructions reviews treat embedded commands as evidence rather than authority, place retained global rules by scope, and verify shared personal cores byte-for-byte.
• Made x9-context-files-generator remove derivable repository context only when its absence cannot change the agent's next decision or action.
• Clarified the boundary between x9-skill-creator and x9-agent-instructions, reduced duplicated audit guidance, and made symlinked source skills valid batch-audit targets.
• Corrected skill size metrics to measure the instruction body instead of frontmatter.
• Added a repository-local maintainer release workflow without exposing it through the public Skills installer.

Install / update

• Existing users can update through their current Agent Skills, Claude Code, or Codex installation route without migration steps.

Compatibility

• Existing skill names, triggers, installation routes, and repository-context ownership remain unchanged.

Breaking changes

• None.