새로운 소식

미출시

수정됨

• Dev-tooling supply-chain advisories cleared. Bumped the transitive dev-only dependencies in package-lock.json — esbuild 0.28.0 → 0.28.1 (GHSA-g7r4-m6w7-qqqr, GHSA-gv7w-rqvm-qjhr), vite 8.0.14 → 8.0.16 (CVE-2026-53571), vitest 4.1.7 → 4.1.9, tsx 4.22.3 → 4.22.4. npm audit is back to 0 vulnerabilities. No overrides were needed (every parent range already allowed the patched versions) and package.json is unchanged. These packages are test/build tooling and never ship in the published npm package, so end users were not exposed — this is dashboard/hygiene cleanup.

1.24.0 2026-06-16

수정됨

• An unrecognised finding Source:** value is no longer silently downgraded and overwritten. A Source:** token outside the known set (e.g. a security value written by a future mini version) used to be read as the adversarial default and then *permanently rewritten* to adversarial on the next write (resolve / reopen / add) — silent forward-incompat data loss. The original token is now preserved verbatim on disk and shown as-is in mini findings list, while internal logic still treats an unknown source as the default.

변경됨

• CI actions bumped to their current majors. .github/workflows/ci.yml now uses actions/checkout@v5 and actions/setup-node@v5 (was @v4). The old majors ran on Node 20, which GitHub forces to Node 24 from 2026-06-16 (removed 2026-09-16); the bump keeps CI running without the deprecation warnings. The Node test matrix (20, 22) is unchanged.
• **adversarial-project now reminds you to run a separate security pass (backlog item 5).** The correctness red-team used to mention the built-in /security-review only in passing, so the security pass silently never ran unless you remembered it. The prompt now closes with an explicit, reminder-only block — once the phase is done and committed, run mini security in a separate terminal (clean context, its own report) — and the in-prompt "Security is out of scope" note plus the /mini:adversarial-project slash command point at the project's own mini security pass instead of /security-review. It stays a reminder: nothing is auto-run and no security findings are filed into mini findings.

추가됨

• **Close extra findings at the done checkpoint: `mini done --apply --resolve-finding <id> (repeatable).** Until now a phase's done` only closed its single linked --from-finding. You can now also close other open review findings the phase incidentally fixed — each one is recorded on the phase so mini undo reopens it symmetrically. The /mini:done prompt surfaces the open findings as candidates, but closing is opt-in and human-gated: it offers only findings this phase could plausibly have fixed (it hides findings raised *against* the current phase and those already owned by another fix-phase), and closes nothing unless you confirm. Used without --apply, or on a done that doesn't finalize (pending verify, unfinished steps), the flag is now a loud no-op instead of a silent one.
• Close findings by hand: mini findings resolve <id...> / reopen <id...>. Until now a finding was only ever closed automatically, when a whole phase saved with --from-finding reached done; there was no way to close a finding a multi-fix phase addressed or to dismiss a nit. The two new subcommands flip one or more findings (mini findings resolve 160-1 160-2) and are idempotent — an id already in the target state is a benign no-op, not an error, while an unknown or malformed id is reported and makes the call exit non-zero without stopping the rest of the batch. (Wiring finding closing into the do/done lifecycle is tracked as a follow-up.)
• **Record *why* a finding was closed: `mini findings resolve <id...> --reason "<why>". The optional reason is stored as a Reason:**` line on every id in the batch and shown in the entry. It is applied only on the open→resolved flip, so resolving an already-resolved finding never overwrites an earlier reason. Reopening a finding clears the reason (a reopened finding has no closing reason), and passing --reason to reopen is rejected as a usage error rather than silently ignored.
• Security & trust-boundary docs. New [docs/security.md](docs/security.md) documents the agent-trust boundary (finding SEC-1): mini feeds git-shared .mini/ content into the agent's prompts, so a cloned/pulled repo with a poisoned .mini/ is a prompt-injection vector when running mini auto (acceptEdits). It records the mitigations — read an untrusted .mini/ before running, don't run mini auto unattended on un-reviewed repos, and the human done checkpoint as the main safety net. README gained a short ## Security section linking to it, and the warning is cross-linked from docs/README.md and the [mini auto](docs/non-interactive/auto.md) notes.

보안

• Path-traversal guard at the graph write site (defense-in-depth, SEC-4). The full graph rebuild now asserts each collected source path stays inside .mini/graph/ before writing its node, throwing on an absolute or ..-escaping path. Unreachable today (the collector only yields repo-relative git ls-files paths), it blocks a future collector regression from silently writing graph nodes outside the project tree — matching the check the incremental update path already had.

수정됨

• CI no longer fails npm test because of ambient color detection. GitHub Actions sets CI=true, which picocolors treats as "color supported" and enables ANSI escapes even without a TTY. Two status assertions matched literal substrings like 2. Title, but pc.bold inserts escape codes inside the title, so the substring no longer matched — green locally (no CI), red on CI. The vitest config now forces NO_COLOR=1 for the whole test run, making output deterministic regardless of the ambient environment.
• Repeated reviews no longer stack stale findings (backlog item 2). The run report's free text is embedded as # Implementation report into the done/verify/adversarial prompts. If an older report still carried a ## Adversarial findings or ## Verify findings section, the reviewer read those prior verdicts as part of the implementation log and a re-run kept re-stacking them. Those sections are now stripped (exact-title match, up to the next ##-or-higher heading) before embedding; a body that is *only* a stale section falls back to the git-diff path in adversarial.
• Corrupt phase file no longer silently loses its detail (blocker 167-1). The phase-detail loader used to treat an unreadable .mini/phases/phase-NNN.json (truncated, half-written, or left with git merge-conflict markers) exactly like a missing one — it degraded to the bare header summary, dropping goal/steps/notes, and the next save made the loss permanent. Reads now distinguish the two: a missing file stays a benign result, but corrupt JSON or any other read error throws a clear CorruptPhaseError naming the file, so verify/adversarial/load report the real problem instead of a misleading "no phase, run /mini:next".
• adversarial-project dedup step is now permitted. The reviewer prompt tells the agent to run mini findings list first so the same issue isn't filed twice, but mini findings list was missing from the command's allowed tools — in a terminal session that triggered a permission prompt and the dedup silently didn't run. Added Bash(mini findings list:*) to the adversarial-project tool set.
• mini findings add --source project is now accepted (blocker 178-1). The adversarial-project reviewer prompt instructs you to record each finding with --source project, but the CLI's --source choices only allowed adversarial | verify, so the command was rejected before reaching the store — the whole project-range review was unrecordable and reviewers had to mislabel findings as adversarial. The --source choices now derive from FINDING_SOURCES, so the CLI can no longer drift from the data model.
• **Stale-findings stripping no longer truncates reports with code fences (finding 178-2).** The cleanup that removes leftover ## Adversarial findings / ## Verify findings sections before a report is inlined into the done/verify/adversarial prompts scanned headings without noticing ``` ` ``/~~~` code blocks. A report whose impl log showed such a heading *inside* a fenced example was silently cut off from that line onward; conversely a real stale section containing a fenced heading line ended early and leaked its tail. Fenced blocks are now honored — headings inside a fence are treated as literal text.
• mini doctor no longer crashes on a corrupt phase file (finding 178-3). doctor is the tool you run to find out *why* a project is broken, yet its full-state read was the one unguarded read in a function where every sibling read is wrapped — so a single unreadable .mini/phases/phase-NNN.json (truncated or carrying git merge-conflict markers) made CorruptPhaseError abort the whole diagnostic. The read is now caught and rendered as a fail "Phases" check naming the file (the phase-hygiene checks that need full state are suppressed); any non-corrupt error is still re-thrown so genuine bugs stay loud.

1.22.0 2026-06-15

추가됨

• Docs for the security review command. Added reference pages for both [mini security](docs/non-interactive/security.md) / [/mini:security](docs/interactive/security.md) under docs/, covering the range rules (no flags = last completed phase, phase vs. git-ref forms, the phase-mode range-end caveat, genesis fallback), the report-path scheme, the scoped tool set, and the inline-is-not-scoped warning. docs/README.md gained a Security review row in the Review table.
• /mini:security slash command. install-commands now generates the /mini:security command idempotently. Its body forwards the range flags straight to mini context security $ARGUMENTS (no flags = the last completed phase; --from-phase/--to-phase or --from/--to for a range) and carries the same independence warning as the adversarial commands. It adds one extra caveat: the scoped tool set the terminal mini security enforces does not apply to an inline slash run, which uses the current session's permissions — so the slash command is a convenience, not an isolated audit (prefer a terminal mini security run, or /clear first).
• mini security command. Opens a fresh Claude Code session for an independent security review. With no flags it reviews the last completed phase (from its pre-commit SHA to HEAD); --from-phase/--to-phase or --from/--to review a range. It runs report-only with read + read-only-git tools and a single write scoped to Write(.mini/security/**) (no Edit, no findings-store write) — the reviewer writes a durable Markdown report to .mini/security/<range>.md itself. Report paths: phase-<id>.md for the default, range-<A>-<B>.md for a phase range, range-<short>-<short>.md for a git-ref range.
• Docs for the adversarial review commands. Added reference pages for both mini adversarial / /mini:adversarial and mini adversarial-project / /mini:adversarial-project under docs/interactive/ and docs/non-interactive/, including the full range rules (phase vs. git-ref forms, mutual exclusion, genesis fallback). docs/README.md gained a Review section grouping these commands with mini findings.
• mini adversarial-project command. Opens a fresh Claude Code session for an independent red-team review of a range of phases, scoped by phase numbers (--from-phase/--to-phase) or git refs (--from/--to). It runs report-only with read + read-only-git tools and mini findings add (no Edit), mirroring mini adversarial. The first message is built from .mini/project.md, the resolved range and the in-range phase list.
• /mini:adversarial-project slash command. install-commands now generates the /mini:adversarial-project command idempotently. Its body forwards the range flags straight to mini context adversarial-project $ARGUMENTS (--from-phase/--to-phase or --from/--to) and carries the same independence warning as /mini:adversarial (prefer a terminal run or /clear first, since an inline review shares the author's context).
• mini context adversarial-project sub-command. Prints the same red-team prompt as the interactive command to stdout, built from the shared range builder, and takes the same range flags (--from-phase/--to-phase and --from/--to). On an invalid or missing range it writes nothing to stdout, reports the reason on stderr and exits non-zero. This is the prompt source the upcoming /mini:adversarial-project slash command will call.
• project finding source. mini findings add --source project is now accepted and the value survives the findings-store round-trip, alongside adversarial and verify. Groundwork for the upcoming cross-project adversarial-project review; no command emits it yet.

수정됨

• adversarial-project range starting at the first phase. A phase-number range whose start is the project's first phase no longer fails with "no recorded pre-commit SHA". When that phase has nothing committed before it, the range start now falls back to the git empty-tree object, so the review diffs from project genesis. Other phases with a missing pre-commit SHA still report a clear error. (The range *end* for very early phases remains a separate limitation.)

1.21.0 2026-06-14

추가됨

• Verify findings go to the durable findings store. mini verify / /mini:verify now records each finding by calling `mini findings add --source verify into .mini/findings/, instead of appending a ## Verify findings` section into the run report. Like adversarial findings, they survive a corrupt or missing report and a closed phase, and surface later in mini next as candidate fix phases. mini findings add gained a `--source <adversarial| verify> flag (default adversarial`) and each stored finding carries a source tag; mini findings list shows it.

변경됨

• The findings store now holds both review kinds. .mini/findings/ is shared by the adversarial red-team and the verify review; its header reads `# Review findings and mini next` lists open findings from both, tagged by source (id · severity · source · where — title). Older finding files with no Source: line are read as adversarial (backward compatible).
• Completing a fix phase resolves its linked finding. When a phase created with --from-finding closes (mini done), mini now flips that adversarial finding from open to resolved and folds the change into the phase commit; mini undo reopens it so the two stay in sync. Tolerant by design — a missing, already-resolved or hand-edited finding never breaks done/undo. This closes the loop: a finding is born open in the review, links to a fix phase, and clears itself once that phase is done.
• Open adversarial findings surface in mini next / /mini:next. When you propose the next phase, mini now lists the still-open findings from the red-team review (id · severity · where — title) as candidate fix phases, alongside the todo backlog.
• --from-finding <id> links a phase to the finding it fixes. When a phase is born from an open adversarial finding, save it with `mini next --apply --from-finding <id>`: mini records the link on the phase. This does not close the finding (it stays open until the fix is done and verified) — unlike --from-todo, and an unknown id fails the save rather than warning. The link lets mini discuss / mini plan read the finding's full detail (title, severity, location, body) straight from disk in any later session, so the planner works from what the red-team actually found, not from chat memory.
• mini findings — a durable store for adversarial review findings. Findings now live in their own .mini/findings/phase-{id}.md files, versioned with the code (like .mini/decisions/ and .mini/memory/) — no longer buried in the run report of a phase nobody reopens once it closes. Each carries a severity, an open/resolved status and an optional location. mini findings add records one (mini owns the id, format and origin phase, so the reviewer never edits a file), and mini findings list [--all] shows the open findings across all phases. Surfacing them in next/plan/do and a resolve command are planned follow-ups.
• Findings now record the baseline commit they were reviewed against. When mini findings add runs inside a git repo it stamps the finding with the current HEAD SHA (shown shortened as @1a2b3c4 in mini findings list). Because a review runs between do and done while the phase work is still uncommitted, this is the phase's *parent* commit — the code state the review started from — recorded honestly as a baseline, not the reviewed commit. It lets a later consumer judge whether a finding may be stale after the code moved on. Additive: findings recorded before this, or outside git, simply have no SHA.
• mini adversarial — an independent red-team review step. A new command and /mini:adversarial slash command run a reviewer that switches into the role of someone who did *not* write the code and hunts for what breaks it (unhappy path, silent assumptions, premature complexity, gaps in tests), reading the real git diff of the phase. It targets the current phase (or the last closed one), records each finding via mini findings add (and prints a adversarial: pass | findings | blocked status line for the human), and never moves the phase state — closing stays a human decision in done. The terminal command spawns a fresh Claude session (clean context = a genuinely independent reviewer) that is report-only: it gets read + search tools and read-only git (Read/Grep/Glob/LS + scoped git diff/log/show) but no Edit, so it cannot modify the code it reviews — its single write is the scoped mini findings add. The inline slash command is honest that it shares the current session's context and points to the terminal command or /clear for real independence.

1.20.0 2026-06-09

추가됨

• C/C++ mapper for mini map. The project graph now covers .c/.h/.cpp/.hpp/.cc/.hh files: #include imports (local "util/foo.h" and system <vector> forms stay distinguishable), free functions with parameter/return signatures (definitions and prototypes, multi-line declarations, trailing return types, default values), class/struct definitions with their public methods, enum/enum class, typedef (incl. typedef struct {...} Name and function pointers) and using aliases — all with line anchors for targeted reads. Declarations inside namespace { } and extern "C" { } blocks are mapped too, and CMakeLists.txt now counts as a project marker. Same regex trade-offs as the other mappers: preprocessor conditionals are not evaluated and macro-generated declarations are invisible; static free functions count as API only in headers.

변경됨

• Prompt hardening for Fable 5 and newer model generations. Newer models (Opus 4.7/4.8, Fable 5) are tuned to be concise between tool calls and to avoid blocking questions, so they tended to compress mini's overviews into a one-line summary and to save state without waiting for approval. All slash command bodies and session prompts now state the two contracts explicitly, via shared hints in src/prompts/sessionHints.ts: (1) command output shown to the user (status, doctor, changelog, map, audit, model, todo, init, undo, upgrade, import-gsd) must be printed **verbatim in the final message** — the user does not read the Bash tool result; and (2) every question to the user (next, plan, project, decision, done, verify, plus the confirmation steps of init, undo, upgrade, model, import-gsd) ends the turn — no mini ... --apply may run in the same turn as the question. auto is deliberately untouched (autonomous mode wants brevity). Trade-off: well-behaved models now print long outputs in full and approvals cost one extra turn.

1.19.0 2026-06-09

수정됨

• Status line: Fable models get the 1M context window. windowForModel() now recognizes Fable in the model display name, so the usage percentage is computed against 1M instead of 200k. In addition, any model whose id carries the [1m] suffix (e.g. claude-fable-5[1m]) gets the 1M window regardless of its display name — a future-proof signal for new long-context models without a code change.

1.18.0 2026-06-07

추가됨

• /mini:project (and mini project) — shape the project vision. A new plan-before-code step that runs after mini init and **enriches the existing project.md with Approach, Non-goals and Success criteria** (the existing name / target user / constraints are kept). In a session Claude runs a short, deliberately critical four-stage interview (frame & remove assumptions → rough plan with trade-offs → non-goals as rules → final check + success criteria) and saves the result through the mini project --apply contract (stdin), which writes only project.md and never the phase state. project.md stays a one-page steering doc — only the main points are written, not a full spec.
• README: "Walk through your first phase" walkthrough. A collapsible <details> block in Quick start shows the first-five-minutes slash-command path (/mini:init → /mini:next → /mini:plan → /mini:do → /mini:done) as a short You:/Claude: conversation per step — the interactive chat experience, not the CLI transcript. Snippets are explicitly labelled illustrative so the non-deterministic chat output isn't read as a fixed transcript.
• README: "How is this different from Claude Code's plan mode?" section. A new section (above the install steps, right after the demo) preempts the "why not the built-in plan mode?" question with a short comparison table (scope, persistence, memory, git, autonomy). Framed honestly — native plan mode and its in-session todos are ephemeral; mini is the persistent, multi-session layer on disk — rather than as claims about what native plan mode can't do.

변경됨

• **/mini:plan and /mini:discuss reference project.md instead of inlining it.** In a running session these slash commands no longer paste the whole project.md into the prompt every time; they reference it with a read-once instruction ("read it only when you don't have it in context"), matching how /mini:do already worked. Saves tokens per invocation, especially with a richer project.md. Cold paths are unchanged: mini next and the interactive terminal mini plan / mini discuss still inline the full project. The shared wording also covers long sessions (compaction) and a fresh session after a crash.
• README: dropped the hidden CI/billing comment and a redundant disclaimer. The commented-out CI badge with its "GitHub account billing issue" note was removed from the public source; the npm/node/license badges stay. The "illustrative" demo disclaimer now appears once (in the Quick start walkthrough) instead of repeated across sections.
• README: "See it in action" no longer duplicates the demo as text. The inline <details> console transcript of the full cycle was removed; the image-not-loading fallback is now a one-line link to the Quick start walkthrough, which already shows the same flow as text. Also moved the plan-mode comparison section above the install steps (right after the demo) so the "why use this" argument comes first. (Combined −30+ lines.)
• README: "Import from GSD" trimmed and "Workflow tips" moved to docs. The GSD section drops to two sentences (positioning + how to import) keeping the [mini import-gsd](docs/non-interactive/import-gsd.md) link; the four workflow-tips bullets move into a new section in [docs/faq.md](docs/faq.md). This completes moving the reference-style content out of the README. (−7 lines.)
• README: "What gets sent to Claude" + graph sections condensed. The full context budget and the per-call cost-line example move into a new [docs/context.md](docs/context.md) (linked from "Concepts & guides"); README keeps a two-paragraph summary — the token-efficiency pitch (softened to *roughly* 600–1000 tokens) plus a one-line graph blurb linking to [mini map](docs/non-interactive/map.md). The standalone ## Machine-readable project map section was folded into "What gets sent to Claude". (−12 lines.)
• README: auto mode consolidated into one section. The two overlapping auto blocks — the 13-line ### Autonomous /mini:auto flag reference under Commands and the standalone ## Auto mode section — collapse into a single short "Autonomous mode" summary (chains phases, acceptEdits with Bash still asking, human checkpoints at next/done, mini stop to halt) linking to [/mini:auto](docs/interactive/auto.md) and [mini auto](docs/non-interactive/auto.md). The per-flag list moves out of the README — it is already documented on both pages. (−12 lines.)
• README: "FAQ" and "Files in the project" moved into docs/. The full state/layout reference (.mini/ tree, state.json/phases/graph/memory prose) now lives in [docs/files.md](docs/files.md) and the complete FAQ in [docs/faq.md](docs/faq.md), both linked from a new "Concepts & guides" section in [docs/](docs/README.md). README keeps only four new-user FAQ entries (permission prompts, pause & resume, commit-after-phase, API key vs Pro/Max) plus a one-line .mini/ pointer; the version-bump/undo/memory internals are out of the README. (−44 lines.)
• Demo: the cycle GIF now shows the slash-command flow. demo/cycle.sh was rewritten from a CLI walkthrough into the interactive /mini:* dialog (/mini:init → next → plan → do → done) so the GIF matches the primary path in Quick start. It stays offline and deterministic: Claude's replies are scripted/illustrative, but each [ok] … line is real output from the same mini … --apply a slash command calls under the hood. The README GIF caption, alt text and the text-transcript fallback were aligned to the new flow.
• README: top-level Documentation pointer. Added a short Documentation: line near the intro (right after Website:) linking to [docs/](docs/README.md) as the central command reference, so the full two-variant reference is discoverable above the fold. The existing 📖 callout in ## Commands was reworded ("Browse the per-command pages …") so the two pointers have distinct roles instead of being duplicates.
• README: slimmed the ## Auto mode and ## Import from GSD sections. Both collapse to a short pointer each, linking to [mini auto](docs/non-interactive/auto.md) (and [/mini:auto](docs/interactive/auto.md)) and [mini import-gsd](docs/non-interactive/import-gsd.md) (README −32 lines). The auto report contract, the one-session-per-phase rationale and the 3-pass retry were added to docs/non-interactive/auto.md first; the GSD positioning sentence (mini as a lighter-weight alternative) is kept in the README.
• README: slimmed the graph sections into one blurb. The three sections (## Machine-readable project map, ### Incremental update (--file), ### Auto-update after an edit (hook)) collapse to a single short paragraph plus a link to [mini map](docs/non-interactive/map.md) (README −35 lines). The --file incremental nuances and the full PostToolUse hook settings.json snippet were added to docs/non-interactive/map.md first, so the detail moved rather than being lost.
• README: slimmed the ## Models and ## Status line sections. Both are now short blurbs that link to the full docs ([mini model](docs/non-interactive/model.md), [mini install-statusline](docs/non-interactive/install-statusline.md)) instead of re-documenting flags and status-line internals inline (README −48 lines). The status-line gauge/segment anatomy and the cache/refresh mechanics (temp-dir cache, detached background refresh, per-session check + 5h cooldown) were **added to docs/non-interactive/install-statusline.md** first, so the detail moved rather than being lost. The per-call token-cost example moved next to `## What gets sent to Claude`.
• README: slimmed the command sections, single reference in docs/. The full 24-row ## Commands table and the duplicate `## mini commands directly in Claude Code slash catalogue are gone. A concise ## Commands` block now explains the two-variant model (interactive /mini:* vs. terminal mini *) and points to the complete reference in [docs/](docs/README.md), so the command catalogue lives in one place instead of being copied into the README (−57 lines). The ### Autonomous /mini:auto explanation is kept.
• Problem-first README intro. The README now opens with a problem-first pitch (hook + the two ways an unsupervised agent goes off the rails + the *propose → plan → implement → verify* rhythm) instead of a terse technical description. The heading is # mini-orchestrator (matching the npm package and website; the CLI command stays mini), with npm version / Node / license badges. The top region is reordered to intro → *See it in action* → *Quick start* (slash commands as the primary flow, the CLI loop secondary) → *Requirements* → *Installation* (the npx, no-sudo and from-git notes folded into collapsible blocks). The GSD comparison moved out of the product definition into the *Import from GSD* section.

제거됨

• Sponsorship asks. Dropped the README *Support* and *Backers* sections and the .github/FUNDING.yml (and thus the repository's Sponsor button) — premature for a project still gathering its first users.

1.17.0 2026-06-04

추가됨

• Decision records (ADRs) for phases. A phase can now carry a lightweight decision record in .mini/decisions/phase-<n>.md that captures the *why* behind a non-trivial choice (the rejected alternative and the reason) — what the goal and the commit message don't preserve. The file's existence is the single source of truth (no flag in the state, nothing to keep in sync), the format is lean (heading + Decision + Why, max one per phase, no NNNN- numbering). mini status --phase <n> renders it under a Decision: heading and mini status --phase <n> --json carries it in a new decision field. Collecting them automatically, an overview marker, a doctor orphan-check and undo handling are follow-ups.
• Writing decision records. New command mini decision --apply writes a phase's ADR to .mini/decisions/phase-<n>.md from stdin (targets the current phase, so run it before mini done --apply to land it in the phase commit). An empty body or a body without a top-level # heading writes nothing — "no decision" stays the file's absence.
• On-demand /mini:decision. Drafting an ADR is its own command: it drafts a lean record from what actually happened, shows it to you for approval, and writes it — but only on a real crossroads (a weighed-and-rejected alternative), not routine choices. /mini:do and /mini:done no longer carry the full instruction; they only point you to /mini:decision when a phase makes a real decision, which keeps their prompts lean (the done prompt drops ~270 tokens per phase).
• ADR marker in mini status. The phase overview now flags every phase that carries a decision record with a compact ✎ ADR marker after its title, and mini status --json sets hasDecision: true for it. The whole overview costs a single readdir of .mini/decisions/ — no per-phase reads.
• mini doctor orphan-check for decision records. A new "Decisions" check flags phase-<id>.md files in .mini/decisions/ whose phase no longer exists in the state (leftovers after mini undo / migrate --renumber), mirroring the existing stale-run-reports check.

1.16.0 2026-06-03

추가됨

• Shell completion. New console-only command mini completion <bash|zsh> prints a completion script that completes mini's subcommands, each command's option flags (e.g. mini done --+Tab → `--apply --accept-verify --bump --push) and fixed flag values (mini done --bump+Tab → none patch minor major). Enable it with source <(mini completion bash) (or zsh`) in your shell rc; the bash script needs no bash-completion package. The commands, flags and values are derived from the CLI at generation time, so they stay in sync across upgrades.
• mini next --apply --from-todo <n>. When you save a phase that grew out of a [todo](docs/non-interactive/todo.md) backlog item, the new flag ticks that item off in .mini/todo.md automatically, so the backlog never drifts out of sync. The /mini:next prompt now lists open items with their archive number (- [n] …) and tells Claude to pass --from-todo <n>. A bad reference (out of range or already done) only warns — the phase is still saved.
• mini status --phase <n>. Zoom in on a single phase instead of the whole overview: its title, goal, status and duration, every step **with its planning detail**, and the phase's run report (verdict, items pending verification, and the free-text notes). Works with --json too for a machine-readable object. An unknown <n> fails with a clean error.
• mini doctor phase-hygiene checks. The health check now also flags phases stuck in doing with no open work left (pointing at mini done) and stale run reports in .mini/run/ whose phase no longer exists (leftovers after mini undo / migrate --renumber).

변경됨

• --bump now validates its value. The --bump option on mini done / mini auto is defined with an explicit choice set, so an invalid level (e.g. --bump foo) is rejected with a clear "Allowed choices are none, patch, minor, major" message and the choices show up in --help.

1.15.0 2026-06-02

추가됨

• Continuous integration. Added a GitHub Actions workflow (.github/workflows/ci.yml) that runs typecheck, the test suite and build on push to main and on pull requests, across Node 20 and 22, plus a CI status badge in the README.
• README "Backers" section. A thank-you/call-to-action section with a sponsorkit-style <!-- sponsors --> placeholder, ready to list project backers once GitHub Sponsors goes live.
• GitHub Sponsors funding. Added .github/FUNDING.yml (github: czsoftcode) so the repository's Sponsor button points at GitHub Sponsors, plus a short "Support" section in the README.
• GitHub community files. Added contributor onboarding under .github/: issue forms for bug reports and feature requests (ISSUE_TEMPLATE/*.yml) with a config.yml that disables blank issues and links to Discussions/docs/website, a pull-request template with a checklist, and a root CONTRIBUTING.md (dev setup, PR workflow, language policy, links into the README and LICENSE).
• README demo of the workflow cycle. A recorded terminal GIF of the full init → next → plan → do → done loop now sits at the top of the *Quick start* section (with a <details> text transcript as a fallback). It is produced from a real, fully offline run — the new demo/cycle.sh drives every step through the non-interactive --apply flags (no Claude API), and demo/record.sh records it with asciinema and renders demo/cycle.gif via agg. Re-run demo/record.sh to refresh the GIF.

1.14.0 2026-06-02

추가됨

• mini install-statusline command. The opt-in counterpart to the install: enables the mini status line by adding a statusLine block to ~/.claude/settings.json (creating the file if missing, preserving every other key). It never overwrites an existing status line — an already-present mini one is a no-op, a foreign one is left untouched and reported. Supports --dry-run. This rounds out the install/uninstall symmetry (install-commands + install-statusline in, uninstall out) and is what the postinstall hint now points to for turning the status line on.
• Zero-touch trial via npx. The install-commands command is no longer hidden, so npx mini-orchestrator install-commands is a documented way to try mini without a global install: npx runs it one-off, it asks where to put the /mini:* commands (this project vs all projects), and it never writes into ~/.claude/settings.json on its own. New top-level README section "Try it without touching ~/.claude" and reference pages docs/non-interactive/install-commands.md and docs/non-interactive/uninstall.md.
• mini uninstall command. A visible counterpart to install/postinstall that cleans up everything mini wrote outside the project tree: it removes the /mini:* slash commands (user-scope ~/.claude/commands/mini and, when present, the project-scope .claude/commands/mini) and strips **only mini's own** status line from ~/.claude/settings.json — a foreign status line is left intact. Supports --dry-run (preview, changes nothing) and -y/--yes (skip the confirmation); without a TTY and without --yes it aborts rather than act unprompted. Run it before/after npm uninstall -g mini-orchestrator to fully clean up.

변경됨

• Global install no longer silently edits ~/.claude/settings.json. On a non-TTY global install (npm i -g mini-orchestrator) mini still writes the /mini:* slash commands (additive, namespaced), but the status line is now opt-in — it is never wired into your settings.json without a TTY to ask first. The postinstall prints an honest summary of what it created plus a one-line full-removal hint. The interactive install still offers the status line as before.

1.13.0 2026-06-02

변경됨

• mini update is now an alias for mini upgrade. Typing mini update checks npm for a newer mini-orchestrator and installs it (with the same --check / --yes flags) — so a slip between the two similarly named commands does the expected thing, silently. Its previous behavior (syncing the project's generated .mini/ skeleton + slash commands) is no longer reached via mini update; use mini install-commands to refresh the slash commands. The mini doctor out-of-date-commands hint now points to mini install-commands.

추가됨

• /mini:import-gsd slash command. GSD import can now run from inside Claude Code, following the mini pattern (no nested Claude session, no interactive hang). mini import-gsd gained two non-interactive forms: --prompt prints the extraction prompt to stdout, and --apply reads the extraction response from stdin, parses it (preserving phase statuses) and saves the project + phases (--force overwrites an existing project, keeping its model config). The slash command checks .planning/, confirms before overwriting, has the in-session Claude read .planning/ and produce the contract, then pipes it into mini import-gsd --apply. The bare mini import-gsd terminal flow is unchanged (18 generated slash commands now).
• Per-command documentation under docs/. Every user-facing command now has a detailed reference page, split by variant: docs/interactive/ for the /mini:* slash commands and docs/non-interactive/ for the mini * CLI commands. Each page has a description, usage/flags, worked examples with sample output, edge-case notes, and links to its sibling variant and related commands. docs/README.md is the index, grouped like the website (Project setup / Phase loop / Autonomous / State & control). mini stop has a single page under non-interactive/ (console-only, no slash variant). Mirrors the structure of [miniorchestrator.com/en/docs](https://miniorchestrator.com/en/docs); the README will later link to these pages.
• Reference pages for import-gsd. Added docs/interactive/import-gsd.md and docs/non-interactive/import-gsd.md — the one command that was still missing its docs — and listed import-gsd in the docs/README.md index under Project setup. Now every user-facing command has both variant pages.

변경됨

• Project website. The package homepage now points to [miniorchestrator.com](https://miniorchestrator.com) (previously the GitHub README), and the README links to it. repository and bugs still point to GitHub.

1.12.0 2026-06-01

추가됨

• mini status --json. A machine-readable JSON object (project title, the "what" line, configured models, currentPhaseId, the open-idea count, and the phases with their status, startedAt/completedAt, durationMs and steps) for scripts and integrations. Printed to stdout with no decoration; the human overview is unchanged without the flag.
• Phase duration in mini status. Each finished phase now shows how long it took (a compact (took 3m) / (took 2h 5m) suffix), computed from its startedAt/completedAt timestamps; phases without both timestamps show nothing.
• mini doctor command and /mini:doctor slash command. A new mini doctor prints a health-check checklist of the project setup — the state and its schema version (legacy → mini migrate, missing → mini init), project.md and CHANGELOG.md presence, the installed slash-command count vs the expected number (→ mini install-commands / mini update), and mini version freshness from the cache (→ mini upgrade) — each line marked ok/warn/fail with a fix hint. Read-only; the /mini:doctor slash command relays it (17 generated commands now).
• mini changelog <version>. The changelog command takes an optional version argument and prints just that version's section (tolerant of a leading v, e.g. v1.11.0); an unknown version is reported with the list of available versions. Exposed through the /mini:changelog slash command too. (A positional argument rather than --version, which collides with the global version flag.)

1.11.0 2026-06-01

추가됨

• mini changelog command and /mini:changelog slash command. A new mini changelog prints the project's CHANGELOG.md changes: by default the latest released version's section, --unreleased the pending [Unreleased] section, and --all the whole history. A missing changelog is reported gracefully. The read-only /mini:changelog slash command relays the same output in Claude Code (16 generated commands now).
• mini todo maintenance and visibility. The ideas archive gains two housekeeping actions — mini todo edit <n> "<text>" rewrites an item's text in place (keeping its done state) and mini todo clear drops all ticked-off items at once — and mini status now shows the open-idea count (Ideas: N open) in its header when the archive has open items. The /mini:todo argument-hint, slash body and the listing's actions hint list the new actions.
• Feeding the todo archive with Claude's ideas. Two ways to fill the ideas/changes backlog without starting a phase or hunting for ideas elsewhere: (1) when mini next / /mini:next proposes its own ideas (the "leave it to me" path), it now sketches 2-3 candidates, takes one as the phase and offers to stash the rest into the archive via mini todo add; (2) a new `/mini:todo suggest (alias ideas`) action has Claude review the project (project.md, phase history, the machine map) and write a batch of small, concrete ideas straight into .mini/todo.md (skipping duplicates).

변경됨

• mini todo listing now shows the available actions. After the N open / M total summary the listing prints a one-line hint (Actions: list · add "<text>" · done <n> · remove <n>), so the sub-commands — including the explicit list action (the same as a bare mini todo) — are discoverable without consulting the help. The /mini:todo argument-hint and the CLI description now list list as well.

1.10.0 2026-06-01

추가됨

• mini todo command and /mini:todo slash command. A new mini todo keeps an archive of future ideas and changes in .mini/todo.md (a plain, hand-editable markdown checklist): mini todo lists the numbered items (open [ ] / done [x]), add "<text>" appends an open idea, and done <n> / remove <n> act on the listed number. The mini next / /mini:next prompt now surfaces the open items as candidate phase ideas, so ideas collected earlier resurface when it's time to pick the next phase. The matching /mini:todo slash command maps its arguments to the right mini todo call (15 generated commands now).
• mini upgrade command and /mini:upgrade slash command. A new mini upgrade checks npm for a newer mini-orchestrator, reports current → latest, and (after confirming) installs it with npm install -g mini-orchestrator@latest. --check only reports without installing; --yes installs non-interactively. A local dev build (install-local) is detected and left untouched with a hint. The matching /mini:upgrade slash command is non-interactive — it previews with mini upgrade --check, confirms in the chat, then applies with mini upgrade --yes (14 generated commands now), and the install hint lists it.
• Status-line upgrade indicator. When a newer mini version is available on npm, the status line appends a yellow ↑ <version> segment (… 28% · ↑ 1.9.1). It never blocks on the network: it reads a cached reading of the latest published version from the OS temp dir and, when that cache is older than 5 hours, fires a detached background refresh (the hidden mini check-version) to update it for next time. mini upgrade does a fresh, blocking check and also refreshes the cache.

변경됨

• Upgrade check now refreshes on every new Claude Code session. The status-line version refresh previously fired only when its cache was older than 5 hours. It now also fires on each new session (detected via the session_id in the status payload), so you get a fresh check every time you start Claude; the 5-hour TTL then only covers a single long-running session. A short retry cooldown keeps a failing fetch from re-firing on every render.

1.9.0 2026-06-01

추가됨

• /mini:undo and /mini:model slash commands. The interactive terminal commands mini undo and mini model now have non-interactive native slash counterparts, so they no longer block on a TTY prompt in the Claude Code Bash. /mini:undo previews the change with the new mini undo --dry-run (prints what would be reverted, including a possible auto-commit soft-reset, without touching anything), confirms in the chat, then applies it with the new mini undo --yes (skips the Proceed? confirmation). /mini:model leans on the already non-interactive sub-commands (mini model show / <scope> <model> / reset) and gathers a missing scope/model in the chat instead of opening the interactive picker. Both are generated by install-commands / mini update (13 commands now), and the install hint lists them.
• mini verify terminal command. A top-level mini verify command now opens an interactive Claude Code session for the in-depth UI/UX review of the current phase (or, when none is in progress, the last closed one), symmetric to mini discuss and the terminal counterpart of the /mini:verify slash command. The first message is the same prompt mini context verify prints; the session runs with the tools needed to write findings into the report and memory (Read, Edit, Grep, Glob, LS). The verify prompt builder was extracted from context.ts into a shared verifyContext module so both entry points use identical wording.
• --bump and --push switches for /mini:auto. The autonomous slash command now parses --bump <level> and --push from its arguments (alongside the existing --max-phases / --yolo / --verify / --discuss) and forwards them to the final mini done --apply of each phase in the run — so an autonomous run can bump the version and push to the remote just like a manual mini done. As there, --push requires an explicit `--bump patch | minor | major; on its own (or with --bump none`) nothing is pushed. The underlying mini auto CLI already accepted these flags; this exposes them through the slash command body and argument-hint.

1.8.0 2026-06-01

추가됨

• Parallelism guidance for the execution agent. The do and auto prompts now carry a shared instruction (PARALLELISM_HINT, alongside the existing GRAPH_USAGE_HINT) on how to batch tool calls: run fragile/stateful commands on their own — especially anything touching a server (start/stop, kill, pkill, background jobs) — so they don't share a batch with other work; fire independent reads and queries in parallel; and keep dependent steps sequential when one command's output feeds the next. It lives only in the execution prompts (not next/discuss/plan), where there is real work to parallelize.
• npm package metadata for discoverability. package.json now carries keywords (claude, claude-code, anthropic, ai, llm, cli, orchestrator, workflow, agent, project-management, phases, typescript) so the package shows up in npm's full-text search and as clickable tags on its npmjs.com page. Also added author, homepage, repository and bugs, so the package page links to the GitHub repository, README and issue tracker. These take effect on the next publish.

1.7.0 2026-05-31

추가됨

• Version bump now follows the project's language. When done raises the version (--bump patch|minor|major), it writes it to the place that matches the project instead of only package.json. Sources are tried in a fixed priority and the first one carrying a version wins: package.json → Cargo.toml ([package]) → pyproject.toml ([project]/[tool.poetry]) → setup.py → composer.json (only when a version field already exists) → __version__ = "x.y.z" (in a common Python location) → a language-agnostic VERSION file. When no manifest carries a version, VERSION is used and, if it does not exist, created with 0.1.0. The tag (--push) and CHANGELOG stamp read the version from the same source. Writes stay a single-line textual replacement — no JSON/TOML reformatting.

수정됨

• A global install (npm i -g) now sets up Claude Code automatically. Previously the postinstall hook bailed out without a TTY, so a global install installed neither the /mini:* slash commands nor the status line (npm runs lifecycle scripts without a terminal). A global install is now detected (npm_config_global) and, even without a TTY, installs the slash commands into the user scope (~/.claude/commands/mini) and wires the status line into ~/.claude/settings.json (only when none exists — a foreign statusLine is never touched). A local / CI install without a TTY is unchanged: it stays quiet and only prints a hint.

1.6.0 2026-05-31

추가됨

• mini ships its own Claude Code status line. A new mini statusline command reads the status JSON Claude Code pipes on stdin and prints one line: the (shortened) project directory, the model with its version, the context-window size (200k/1M) and a colored gauge + percentage of the **context-window usage** — recovered from the session transcript, since Claude Code does not report token counts to the status line directly. Colors are raw ANSI (dir in bold cyan, the gauge green/yellow/red by fill) so they survive the piped output. On npm install the postinstall hook offers it (asks first) and wires it into ~/.claude/settings.json — but only when no statusLine exists yet; an existing one (yours, GSD's, Claude's) is never overwritten, and without a TTY it is skipped silently. Disable it by removing the statusLine block.
• **The /mini:* slash commands are installed automatically by an npm postinstall hook.** After npm install of mini, the hook offers to install the commands and asks where — the user-level ~/.claude/commands/mini (all projects) or the current project's .claude/commands/mini — defaulting to the scope detected from how Claude Code is installed (a project-local node_modules/.bin/claude suggests the project, otherwise the user level). It is non-interactive-safe: without a TTY (CI, npm ci, piped install) it writes nothing and only prints a hint with the manual command, and any error is downgraded so it never fails the install.
• mini-orchestrator is now released under the MIT License. A root LICENSE file (MIT, © 2026 Stanislav Kremeň) makes GitHub show the "MIT" badge in the repo sidebar, the "license": "MIT" field in package.json makes npmjs.com display it on the package page, and the README's license section links to ./LICENSE. npm always includes LICENSE in the published tarball, so it ships with the package too.
• README documents the local dev install. A note under "From git / for development" describes npm run install-local: it builds and installs mini under ~/.local (a ~/.local/bin/mini symlink + a versioned directory with the package files and production deps), keeps older versions around for rollback, and is verified with mini --version.

변경됨

• **mini install-commands is now a hidden manual fallback, not the primary install path.** Installing the slash commands is normally handled by the postinstall hook (see Added); the command stays available (hidden from --help) for when the hook is skipped (--ignore-scripts, npm ci, CI). It gained --user / --project to pick the location non-interactively and --dry-run for a preview. The generator was extracted into a shared module (src/install/commands.ts) reused by the hook, mini update and this command, so their output can't drift.
• scripts/install-local.sh is now fully English. Its header comments and all runtime output (→ installing into …, → production npm install (runtime deps only), mini X installed., Try: mini --version) were translated from Czech. The stray Czech build line in scripts/copy-assets.mjs (assety zkopírovány) was translated to assets copied as well.
• CLAUDE.md is now fully English. The project instructions are written in English so any developer of the public tool can read the conventions without translating. The file keeps the i18n policy (everything inside the program is English) and the "left untouched" list (identifiers, technical terms, quoted foreign output, parser response contract & status words); the maintainer's personal "communicate in Czech" rule was removed (it lives in the maintainer's global ~/.claude/CLAUDE.md).

1.5.0 2026-05-31

추가됨

• The lifecycle command messages are translated to English. The runtime output of mini next / plan / do / done / auto / discuss (and mini context) — log.* messages, interactive ask() prompts and their choices — is now English, together with the in-code comments/JSDoc. The model-scope labels (SCOPE_LABELS used by mini model) are English too. Command names, flags, the parser response contract (TITLE:/GOAL:/STEP:), state words and paths stay unchanged. (The Claude-facing prompts and the remaining internal modules come in the following phases.)
• The auto-commit subject for a finished phase is now English: Phase {id}: {title} (previously Fáze {id}: {title}), to match the translated README. mini undo is unaffected — it matches the commit via preSha, not the subject; no parser depends on the subject. Existing Czech autoCommit.subject entries in past state.json stay as archival data.
• README.md and CHANGELOG.md translated to English. Both public docs are now fully English, keeping the code blocks, command/flag names, paths, links and the Keep a Changelog structure; the README's internal anchor links were updated to the new English heading slugs. From now on, new CHANGELOG.md entries (phase records via /mini:done) are written in English — the convention is noted in CLAUDE.md.
• The phase memory (.mini/memory/phase-XXX.md) is generated in English. The buildPhaseMemoryMarkdown generator now writes English headings (# Phase, Goal:, ## Steps, ## User's note, ## Auto-commit, ## Discussion) and the step statuses done/doing/todo/skipped. The summary for the next prompt (summarizeMemoryForNext) reads both English and older Czech memory (the existing archive is not broken). phase-XXX.json unchanged.
• The utility command messages are translated to English. The runtime output of `mini status / undo / init / import-gsd / model / stop / map / audit / update / install-commands / migrate / migrate --renumber` is now English (phase labels [done]/[doing]/…, the "Next: …" hints, logs about created/changed files, etc.). mini init and mini import-gsd now create project.md with English headings (## What I'm building / ## Who it's for / ## Main constraints); mini status still reads older Czech project.md too. Command names, flags, /mini:* references and paths unchanged. (Lifecycle commands, memory/reports and the graph mappers come in the following phases.)
• CLI help and UI messages translated to English. mini --help, command and option descriptions, error/validation messages and runtime output (src/cli.ts, src/ui/*) are now English — the tool is aiming at international use. Command names, flags, /mini:* references and paths stay unchanged. package.json description and the glossary docs/i18n-glossary.md (a new CLI/UI terms section) were updated too. The language rule in CLAUDE.md now says: the whole program is English, only the chat and commits stay Czech. (Runtime messages in commands/* and other modules will be translated in the following phases.)
• The internal next/plan/do prompts translated to English. The headless path (mini next/plan/do via the API) now generates the instructions for Claude in English; the response contract (TITLE:/GOAL:/STEP:) stays unchanged. A shared translation glossary docs/i18n-glossary.md was created as a basis for the next phases. (The interactive slash-command path and the shared graph hint will be translated in the following phases.)
• The audit and GSD import prompts translated to English. Audit now generates .mini/codebase.md with English section headings (Overview / Directory structure / Key modules / Technologies); the GSD import has English prose. The machine contracts (NAME:/WHAT:/FOR_WHOM:/CONSTRAINTS:/PHASES: and the status words) stay unchanged.
• The memory-writing and autonomous (auto) run prompts translated to English. The memory prompt and the auto session prompt (and thus the interactive /mini:do / /mini:auto) are now English. The machine contract of the YAML report (phase/verdict/steps/status/verify + their values) stays unchanged.
• Instruction translation completed — all prompts are English. The interactive session prompts (/mini:next, /mini:plan, /mini:done, /mini:verify), the discuss prompt and the shared graph hint were translated too. The discussion notes template and the "Verify findings" section are English and the last-memory summarizer recognizes them (older Czech memory is still caught). The shared glossary is in docs/i18n-glossary.md.
• The /mini:* slash-command files translated to English. The descriptions and bodies of the generated .claude/commands/mini/*.md (source install-commands.ts) are now English including the autonomous auto. Commands, flags and $ARGUMENTS unchanged; the CLI logs stayed Czech.
• /mini:auto --discuss — the flag forces the discuss step in every phase of the run (analogous to --verify). Without it, discuss runs only conditionally for hard phases.
• Verify in the autonomous /mini:auto — the cycle now runs the verify step between do and done for UI/UX phases (Claude judges from the goal/steps/report). A new flag --verify forces it in every phase. The findings are written into the report (and thus into memory) and any problems are fixed within the same phase before closing.

변경됨

• **mini done commits the phase in a single commit — nothing dangles in the worktree after done.** The memory record, the regenerated graph and the final state.json (the move to done) are now created before the commit, so git add -A picks them up into a single phase commit. Previously the commit happened earlier and these artifacts dangled until the next phase. mini undo identifies the phase commit via preSha (HEAD^ === preSha) instead of its own sha — that one is no longer stored in the committed state (it would depend on itself; older phases still have it in state.json, backward compatibility preserved).
• mini verify / /mini:verify is no longer read-only — after the review it writes the findings into the run report (## Verify findings), from where they reach memory through the report too; for an already closed phase it also appends them directly to the memory file. It still does not move the phase state.
• README unified with the actual state of the tool. Missing commands were added (mini stop/migrate/update) along with the versioning/CHANGELOG/tag for done; the .mini/ tree was updated to layout v2 (phases/, graph.json + graph/); the memory description was fixed (the phase-{id}.md file without a timestamp, last-memory.md is a short summary, not a symlink).

수정됨

• After an npm install, mini init creates .mini/.gitignore. The skeleton keeps the gitignore under the npm-safe name gitignore (no dot) — npm publish excludes .gitignore files from the tarball, so on a fresh machine it was missing from the skeleton and mini init/mini update did not create it in the project. It is still written into the project as .gitignore (the rename is handled by assets.ts:FILE_RENAMES).

1.4.0 2026-05-30

추가됨

• mini verify / /mini:verify — an in-depth UI/UX review of the phase by a human. Claude interactively guides you through a visual/UX review (sets the scene, goes through the verify items from the report, adds a broader UX walkthrough and collects findings). It targets the current phase, otherwise the last closed one. It is read-only — it does not move the state, that stays with done.
• mini stop — a cooperative stop of the autonomous /mini:auto. It creates the signal .mini/STOP (from a second terminal); a running /mini:auto reads it at the step boundaries, finishes the current step, writes a report and exits cleanly. mini stop --clear removes the signal. Both variants are idempotent.
• The autonomous /mini:auto — the slash command now completes several phases in a row (next → discuss(conditionally) → plan → do → done → repeat) with --max-phases N (default 1) and --yolo. It stops and asks at the steps that require a human (next, discuss, the items for manual verification in done), and runs quietly for do (it does not retell the edits). It reads the cooperative stop hooks (.mini/STOP) at the step boundaries — the signal is created by mini stop (see above).
• .claude/settings.json with an allowlist (mini:*, build/test, git), so the autonomous run does not bother you with command confirmations.
• mini map --file <path> — an incremental graph update: it remaps just one file (the node .mini/graph/<path>.md + the record in graph.json, atomically via tmp+rename, preserving order) instead of a full rebuild. Can be repeated for several files. A disappeared file removes the node and the record; non-mappable extensions and ignored directories are a no-op; a missing index falls back to a full build. Because the graph nodes are per-file, the result is identical to a full rebuild of the affected file.
• mini map --hook — for autonomous mode: it reads the edited file path from the PostToolUse hook JSON on stdin and remaps it incrementally (no dependency on jq). Without a path it silently no-ops. The README describes a snippet for .claude/settings.json that keeps the graph fresh after every Edit/Write; mini init points to it.
• /mini:init — project initialization directly from Claude Code: the slash command asks in the session for four things (name, what you're building, for whom, constraints), saves the project via the new non-interactive `mini init --apply --name/--what/--for-whom/--constraints [--force]` and, based on the directory content, offers the next steps — for existing code /mini:map and /mini:audit, otherwise /mini:next.
• /mini:audit — a slash command that runs mini audit (an overview of the existing codebase into .mini/codebase.md) directly from Claude Code.

1.3.0 2026-05-30

추가됨

• mini update — brings the non-generated part of the project up to the current mini version: the static .mini/ skeleton (directories + .gitignore) and the slash commands .claude/commands/mini/*.md. Idempotent — creates the missing, overwrites the changed (the skeleton files are mini-owned), leaves the rest unchanged and prints a summary. --dry-run shows a preview without writing. The skeleton lives as a shipped asset (assets/skeleton/ → dist/skeleton) and is the single source of truth: mini init draws from it too and it is easily extended with more static files.
• mini migrate --renumber — renumbers the phases to consecutive integers (1..N by their order in state.json) and unifies the file names in all four directories (phases/, discuss/, run/, memory/) to the canonical phase-XXX. It straightens projects with mixed/legacy numbering (e.g. decimal "fix" ids 1.1…28.1 next to integers). It handles various old name schemes (padded and unpadded, .prev.md, memory with a timestamp). Idempotent; --dry-run shows a preview of the mapping without writing, otherwise it asks for confirmation before the change. It leaves orphans (files with no record in the state) alone with a warning, and stops on a collision of target names so it overwrites nothing.
• The knowledge graph map now supports Ruby (.rb): it extracts the imports (require and require_relative, including the parenthesized form) and top-level declarations — def (with a parameter signature including splat */**, keyword key:, default values, &block and endless methods def x = …) and the types class (kind class) and module (kind module). The default visibility is public; bare private/protected and private def … hide the following members. Visible instance and class methods (def self.x) and attributes (attr_reader/attr_writer/attr_accessor) are attached to the type with line anchors. Comments (# and block =begin/=end) and strings are ignored. The project is also recognized by Gemfile.
• The knowledge graph map now supports Swift (.swift): it extracts the imports (including submodules import Foo.Bar and kinded import struct Foo.Bar) and top-level declarations — func (with a parameter signature, default values, variadics Int..., generics, async/throws and where clauses) and the types class/struct/enum/protocol/ extension/actor. The default visibility is internal; private/fileprivate is omitted (private(set) stays visible). Visible methods are attached to the type with line anchors, static/class func are marked. Comments (including nested block comments), doc comments and strings (including multiline """…""" and raw #"…"#) are correctly ignored. The project is also recognized by Package.swift.
• The knowledge graph map now supports Kotlin (.kt/.kts): it extracts the imports (including wildcard import a.b.* and aliases import a.b.C as D) and top-level declarations — fun (with a parameter signature, default values, vararg, generics and the extension receiver) and the types class/interface/object/enum class/data class/ sealed class|interface/annotation class. The default visibility is public; private/internal is omitted. Visible methods are attached to the type with line anchors. Comments (including nested block comments), KDoc, char literals and strings (including raw """…""") are correctly ignored. The project is also recognized by build.gradle.kts.
• The knowledge graph map now supports C# (.cs): it extracts the usings (using, using static, global using and aliases using Foo = A.B) and top-level types inside a namespace (block and file-scoped) — class/struct/interface/enum/record (including record class/record struct), with public/internal methods (parameter signatures including params, default values and static) attached to the type and line anchors. Comments, XML doc, char literals and strings in all variants (verbatim @"…", interpolated $"…", raw """…""") are correctly ignored. The project is also recognized by *.sln/*.csproj.
• mini done/mini auto: the --bump none option — closes the phase without bumping the version (suitable for partial phases, where the version is raised only at the end of the whole unit).
• The knowledge graph map now supports Java (.java): from the classes it extracts the imports (import, import static and wildcard import a.b.*) and top-level types — class/interface/enum/record/@interface declared as public/protected, with public/protected methods (including parameter signatures, varargs and static) attached to the type and line anchors. Comments, javadoc, strings and text blocks ("""…""") are correctly ignored. The project is also recognized by pom.xml or build.gradle(.kts).
• The knowledge graph map now supports Go (.go): from the package it extracts the imports (single and block import ( … ) including aliases, _ blank and . dot imports) and top-level exports — functions with a signature, struct/interface, type aliases and const/var (also grouped), exported by the capital initial letter. Methods are attached to the receiver type, all with line anchors. The project is also recognized by go.mod.

변경됨

• mini init now creates .mini/ from the same skeleton as mini update (the directories phases/, memory/, discuss/, run/ + .gitignore); project.md and state.json are still generated separately.
• The default behavior of --bump is now none (previously patch): neither mini done nor mini auto bumps the version in package.json by default anymore. To raise it, use --bump patch|minor|major.
• --push now requires an explicit --bump patch|minor|major — a push without a version level (or with none) ends with an error. A push = a release, so it must have a version for the tag.
• Unified phase file names across all .mini/ directories: discuss/, memory/ and run/ now use the same format phase-XXX (3 digits with zero padding) as phases/. The ISO timestamp disappeared from the memory name; a repeated record of the same phase is distinguished by the suffix -2, -3, … instead of the date. Existing files were renamed.

1.2.0 2026-05-30

추가됨

• The knowledge graph map now supports Python (.py/.pyi): from the module it extracts the imports (import, from ... import including relative, aliases, * and multiline) and top-level exports — functions (def/async def) with a signature, classes with public methods and UPPER_CASE/annotated constants, with line anchors. The project is also recognized by pyproject.toml/setup.py; .venv/ and __pycache__/ are ignored.

1.1.0 2026-05-30

추가됨

• /mini:done now creates and maintains CHANGELOG.md in the Keep a Changelog 1.1.0 format: from the report of a finished phase, Claude writes the changes under ## [Unreleased] (the Added / Changed / Fixed sections). On a release with --bump minor/major and --push, the content of ## [Unreleased] is folded into a dated section ## [version] - date (matching the git tag) and a new empty ## [Unreleased] is inserted on top; patches accumulate in Unreleased until the next release.