Repository navigation
Releases
Releases · kirg0/d9c
v1.35.3
Fixed
- Destructive commands always wait for the user's yes: the skills disagreed — the base skill
asked for approval,d9c-lifecycleleft it to the MCP client — and with the client's prompt
switched off (auto mode, an allowed tool) an agent removed containers right after the dry run.
Now every skill with a destructive command states one rule: show the resolved plan and wait for
an explicit yes, even when the request already says "remove", without stopping things first as a
head start.d9c mcpsays the same in the destructive tools' descriptions and, when they are
offered, in its instructions, so clients without the skills get it too. A test checks the rule in
every such skill; a new eval case (remove-staging-asks-first) passed 3 of 3 runs. - Skill evals: the compose order grader ignores dry-run calls.
v1.35.2
Fixed
- Demo scenarios: a scenario's
compose backupwrites a real tar.gz of the project files
(it was a 20-byte placeholder), and compose pull/up/down report the project's own services and
images instead of a canned nginx line — the first eval run showed an agent rightly refusing to
deploy on that data. - Skill evals: the trigger-discipline case no longer counts the tool list in the trace as a call;
the overview rubric does not demand stopped containers for a question about running ones; the
compose case no longer requires awaitwhen the health is already final. The based9cskill
says to read it before the first d9c command or tool call.
v1.35.1
Fixed
- CI lint: the current
errcheckflags discarding the result of an error channel; the two
places that wait for an aborted reader (execover the API and over SSH) now say why the error
is ignored, somake checkpasses with the latest golangci-lint again.
v1.35.0
Added
- Claude Code marketplace: the repository is a plugin marketplace (
.claude-plugin/marketplace.json)
with thed9cplugin inplugins/d9c: its manifest, an optional saved-host option and.mcp.json,
which startsd9c mcp, so installing the plugin wires up the MCP server. - Agent skills: ten skills in
plugins/d9c/skillsin the open Agent Skills format, each
usable through the CLI and the MCP tools: the based9c(install, hosts, output and error
contract, write policy, untrusted data) andd9c-troubleshoot,d9c-logs,d9c-lifecycle,
d9c-compose(backup → pull → up → wait healthy → restore on failure),d9c-images,
d9c-cleanup,d9c-fleet,d9c-security-audit(read-only) andd9c-net-storage. The MCP
server serves all of them as prompts. Tests parse every CLI command a skill names with the real
CLI parser and check every MCP tool, action and argument against the tool schemas. d9c skills install: the skills are embedded in the binary;d9c skills install --target <agent>
writes them into an agent's user-level skills directory (claude→~/.claude/skills,agents/codex
→~/.agents/skills,gemini,cursor,copilot,opencode) or--dir <path>(e.g. a project's
.claude/skills), stamped with the binary version inmetadata.d9c-version;--dry-runreports
created/updated/unchanged without writing, other skills in the directory are left alone.
d9c skills lslists the skills and targets.- Demo scenarios (
-demo-scenario <name>,D9C_DEMO_SCENARIO): demo data with one planted
problem per skill —oom,restart-loop,staging,unhealthy-compose,stale-images,
disk-full,fleet,insecure,isolated-network(plusdefault) — in the TUI, the CLI and
d9c mcp. Container and networkinspectthere has the real Docker layout. A scenario that
exercises writes admits them up to its own tier, on the fake only;-read-onlystill wins.
d9c skills lslists them. - Skill evals: a
claude plugin evalsuite inplugins/d9c/evals(a case per skill plus a
trigger-discipline case) runs the agent against the plugin's real MCP server on those scenarios
(EVAL_D9C_SCENARIO), grading the diagnosis and the tool calls — including that destructive
steps are only planned, never run without the user's approval.go testreplays each skill's
commands on its scenario and validates the suite for free. - CI checks for the plugin: a
pluginjob runsclaude plugin validate --stricton the
marketplace and the plugin;make checkverifies the Agent Skills frontmatter (name = directory,
description with "Use when", only format keys), unknown manifest keys, that every CLI command a
skill names parses and every MCP tool/action/argument exists, and the eval suite. - README "For AI agents" / «Для ИИ-агентов»: the skills, installing the plugin and the skills for
other agents, demo scenarios and evals. - Image commands for agents:
image history <image>(read),image tag <source> <target>,
image push <image>,image build <dir> [-t tag](safe) andimage rm <image> [-f](destructive),
in the CLI and the MCP tools (d9c_resourcesimage_history,d9c_deployimage_build/image_tag/
image_push,d9c_removeimage_rm). Push credentials come only fromD9C_REGISTRY_USERNAME/
D9C_REGISTRY_PASSWORD; the plans name the registry, the containers using an image and a tag that
moves.
v1.34.0
Added
- MCP server
d9c mcp: serves the agent operations over the Model Context Protocol on
stdin/stdout for one session with one host (--host,-H,-context,-demoas in the CLI).
Twelve grouped tools (d9c_containers,d9c_logs,d9c_compose,d9c_system,d9c_remove…)
with anactionargument,fields/limitoutput shaping anddry_runon write tools; results
are the CLI's JSON envelope as text and structured content, failures{"error":{code,message,hint}}
withisError. Tool annotations follow the operation tier (readOnlyHint,destructiveHint,
idempotentHint), and only the tools the write policy admits are listed. The connection is
dialed on first use and redialed after it drops (reads are retried once, writes never). - MCP progress and cancellation: long tool calls (compose up/pull/down/restore,
pull, compose
backup,wait) report each step asnotifications/progresswhen the client passes a progress
token; cancelling a call stops the operation and its engine stream. - MCP prompts and resources: the agent skills shipped in
plugins/d9c/skills(d9c,
d9c-troubleshoot,d9c-logs,d9c-cleanup) are served as prompts with an optionalfocus
argument, from the same text; resourcesd9c://hostsandd9c://containers/{id}/logsreturn
the same envelope as the matching tool.
v1.33.0
Added
- Container lifecycle commands in command-line mode:
start,stop,restart,kill [-s SIG]
andrm [-f]take several container IDs, unique ID prefixes or names, and/or repeatable
--filter key=valuepredicates (name=/image=globs,status=,label=k[=v],network=,
health=; exact matches, never substrings). The result lists one outcome per container; when
only some of them fail, the envelope carries"partial": trueand the exit status is8. - Image and compose deployment commands in command-line mode:
pull <image>,compose up,
compose pull,compose restart,compose down(containers and networks; volumes kept),
compose backup(archive saved to the current directory, absolute path returned) and
compose restore <project> <archive>. Streaming commands run to completion and return the
engine output (the newest 200 lines); a failure is an error that quotes the output leading to
it.compose downandcompose restoreare destructive. - Network, volume and prune commands in command-line mode:
network create <name>
(--driver,--subnet,--gateway) andvolume create <name>(--driver) return the new
object; an existing name or a gateway outside the subnet is a usage error.image prune
(dangling images),volume pruneandsystem prune(stopped containers, unused networks,
dangling images, build cache; volumes kept) are destructive. - Agent write policy: command-line mode is read-only by default.
agent: {allowWrite, maxTier}
in the config — globally and per saved host, the host block overriding key by key — turns write
commands on up to a tier ceiling (read<safe<destructive;safewhen unset).
--allow-writeallows writes for one command unless the config saysallowWrite: false;
read-only mode (-read-only,readOnly:, a host'sread_only) always wins. A refused command
fails withforbiddenbefore connecting, with a hint naming the setting that blocks it;
d9c helplists only the allowed commands andhosts lsreports each host'sagent_tier. - Dry runs and confirmation for write commands:
--dry-runprints the plan of any write
command (targets resolved, their state noted) without changing anything, and is not blocked by
the write policy. A destructive command without-y/--yeschanges nothing and fails with
confirmation_required(exit7), carrying the plan inerror.planfor the user to approve. - Audit log of write commands: each write command, whatever its outcome (done, partial,
planned, refused, failed), appends a JSON line toagent-audit.lognext to the config file
(~/.d9c/agent-audit.logby default) with the time, host, operation, targets, arguments and
result. A write command that cannot open the log fails before connecting. execin command-line mode:d9c exec <container> -- <cmd> [args…]runs a command without a
terminal on every runtime (Docker/Podman API, nerdctl and crictl locally or over SSH) and returns
its exit code, stdout and stderr, each cut at--max-bytes(64 KiB by default) with the full
sizes reported.--timeout(60 s) ends the wait with the new error codetimeout(exit9).
The command is destructive: it needsmaxTier: destructiveand--yes.waitin command-line mode:d9c wait <container> --until running|healthy|exitedpolls the
container until it gets there and returns its state;--timeout(60 s,0= check once) ends
withtimeout(exit9) naming the state it was left in. Waiting forhealthyon a container
without a healthcheck is a usage error at once.- Read commands on several hosts:
--host a,bor--all-hostsruns a read command on those
saved hosts concurrently;datalists each host's result or error, a failing host does not stop
the others, and some failures make the result partial (exit8). With-demoevery host is
served by the demo backend.
Fixed
- Pull reported success on a failed pull (Docker/Podman API): an error the daemon sends
inside the progress stream (unknown manifest, registry failure mid-download) is now reported. d9c stats webwas a usage error: container names and IDs are now accepted as positional
arguments, the same as--container.
v1.32.0
Added
- Command-line mode for scripts and AI agents. A known command as the first argument runs
d9c without the interface:hosts ls,info,ps [-a],inspect,logs
(--tail/--since/--until/--grep),stats,images,networks,volumes,system df,
events(a past window only) andcompose ls/ps/config/logs. All of them are read-only. - Stable output contract. stdout carries one JSON envelope
{"schema":1,"host","runtime","data","truncated"};-o textprints a table,-o ndjsonone
item per line;--fieldsand--limitshape the result, and a cut list says so with a hint.
Failures are JSON on stderr ({"error":{"code","message","hint"}}, daemon errors reworded)
with distinct exit codes: 2 usage, 3 not_found, 4 connection, 5 forbidden, 6 auth_required. -host <name>picks a saved host by name, in the command-line mode and in the interface.- Safe by default for unattended use. No prompts: a password SSH host takes the password from
D9C_SSH_PASSWORDor fails withauth_required; an SSH host key not inknown_hostsis never
accepted;inspectandcompose configmask env values named like*PASS*,*TOKEN*,
*SECRET*,*KEY*(--no-redactshows them).
v1.31.2
Fixed
- Stale
-helptexts for file flags.-plugins-filenow names its real default
~/.d9c/d9c-plugins.yamlinstead of "next to the binary", and-hosts-fileno longer claims the
plugins file as its default: it points at the legacyd9c-hosts.jsonnext to the binary, which
is only read once to migrate into the config.
v1.31.1
v1.31.0
Changed
- Config files moved to the per-user directory.
d9c-config.yamlandd9c-plugins.yamlnow
live in~/.d9c/(%USERPROFILE%\.d9c\on Windows) instead of next to the binary, so they
survive package-manager upgrades (Scoop and Homebrew keep each version in its own directory)
and work with read-only install locations. The directory is created on first save. Files left
next to the binary by older versions are copied over once on startup (the originals are left in
place);-config/-plugins-filestill override the location.