Skip to content

Releases: JhonHawk/tricell-hive

Hive 0.6.0

Choose a tag to compare

@JhonHawk JhonHawk released this 26 Oct 00:41
ecb394d

Security

  • The documentation site and the starlight-docs-site templates move to
    Astro 7.4.0, Starlight 0.42.6 and Biome 2.5.15, and enforce a 7-day release
    cooldown. The site also overrides patched undici, source-map-js and zod and
    drops an unused dependency that brought extract-zip, so its pnpm audit
    goes from 11 high advisories to none.

Added

  • Stage skills flow-verify, flow-deploy and flow-security, with an
    "Entry and exit" contract in every stage skill. flow-verify checks a built
    candidate at a level the project settings and the request set, and a check
    skipped at your request counts as accepted without verification; flow-security runs the
    security review only when you ask; flow-deploy is a separate stage after
    integration that promotes through the declared environments and checks what
    is served. The README, llms.txt and the documentation site describe them
    (#169).
  • Open topics: the agent keeps a list of the topics you add while work is open
    or closing, tells you in one line how it classifies each, closes none
    without a destination, and names them in its reports and route or close
    questions (#160).
  • Test guidance: tdd.md now sets which kind of test to write and a filter
    before writing one (the test that already owns the behavior, low-value
    patterns, no test-only entry points in production code), recommends
    property-based tests where an independent oracle exists, and verification
    compares the old and the new path on the same inputs when a change replaces
    or migrates something (#157).
  • Options of a question carry their estimated size when one starts work on
    another ticket, adds a component or dependency, or is about ten times
    another, and plans that edit agent guidance get a guidance-consistency
    reviewer. This repository declares Delivery: automatic and
    Review: /code-review (#158).
  • Codex wait floor: an interactive hive install or hive update with the
    Codex host offers, selected by default while it is not applied and Codex is
    0.164 or later, to write min_wait_timeout_ms = 300000 and
    default_wait_timeout_ms = 600000 under [features.multi_agent_v2] in
    Codex's config.toml, so a parent agent waiting for a delegated child does
    not wake every 60 seconds. Hive saves config.toml.hive-backup (0600) before
    the first write and its journal keeps only fingerprints and the lines it
    added; hive plan remove for the codex host removes only the unchanged lines
    Hive wrote, then the backup, after a clean revert; a file Hive cannot edit safely becomes a manual
    step; hive doctor adds a row for it. This replaces the rule
    that Hive edits no host configuration file itself. The minimum applies to
    Codex's v2 multi-agent runtime (v1 unverified), and an older Hive binary that
    rewrites Hive's state loses the record, so plan remove would leave the
    lines. A default-selected capability raises the minor version.
  • Check a dependency before installing: before adding, updating or installing
    a package, including a one-off npx run, the agent confirms the package,
    picks a version the release cooldown accepts, looks up advisories for the
    exact version without sending private package names to third parties, scans
    the lockfile, and asks before installing when a critical or high advisory
    has no safe path.
  • The security-advisory skill checks, read-only, which repositories a
    published advisory or package compromise affects, with evidence for each
    one, and leaves the fixes to each project's own session.
  • The viewer combines walkthroughs with the same project, host and session ID
    into one sidebar entry, with chronological ticket segments, internal jumps
    and independent copy actions. The session header collects explicit PR or
    commit delivery links while keeping each segment's association.
  • CodeGraph is an optional, manual capability: the installer offers it for
    Claude Code, Codex, Cursor and OpenCode with its official source, and
    hive doctor shows whether codegraph is on PATH without running it. The
    global guidance says when to use a CodeGraph index, when to run
    codegraph sync -q so its relations are current, and which results to
    confirm with text search (#104).
  • Plan review adds a User experience area for plans that create or change a
    screen, dialog or form where the end user enters or chooses data. Its
    reviewer walks the end user's task looking for values the system already
    holds or entries repeated per record, and UI planning records where each
    value a view asks for comes from.
  • Before a production deployment of code or infrastructure, the build records
    its rollback triggers: each signal to watch with its threshold and window,
    and the recovery action it calls for. After the deployment, the health check
    compares the signals against them and reports a trigger that fires.

Changed

  • Workflow rules no longer contradict each other across plan, build, verify and
    close: the dedicated review runs once on the local candidate before the
    validation request or the push (a reviewer that only works on a pull request
    runs right after it), and under automatic a review is always offered and
    never recommended away; automatic stays automatic for UI changes and
    reports the first rendered view without asking; interactive stops before
    the push even when verify is skipped; a later completion point is defined
    once and authorizes both ticket moves; a question that depends on another's
    answer waits for a later turn; the findings question offers resolving them in
    this work; unattended work starts on silence after the route answer when
    delivery and review are declared; branch and worktree deletion wording is
    safe; one delegation criterion; role briefs match role contracts. The site,
    README, llms.txt and CONTRIBUTING describe the result (#180).

  • flow-build ends at the built candidate and hands it to flow-verify; the
    independent UI children, the full suite, the dedicated code review and the
    walkthrough delivery move there, and promotion and deployment checks move to
    flow-deploy. flow-walkthrough/references/ui-gate.md and
    ui-review-criteria.md move to flow-verify/references/.

  • The security review (hive-review-security) runs only when you ask for it.
    A declared Review value no longer authorizes it, and the accepted-candidate
    gate, the role description and the security-advisory scope assessment no
    longer launch it. When a change crosses a security boundary, the agent names
    the boundaries and offers the review in one line, without a question that
    waits and without stopping the work: in the delivery statement, also when
    Delivery and Review are declared, or in the next message once a boundary
    appears. A requested review runs before validation, or before the push or
    merge when there is no validation stop (#167).

  • Verification at checkpoints: a plan with numbered criteria groups its tasks
    into checkpoints chosen by risk, and flow-build verifies the tasks of a
    checkpoint together instead of after each task. A task that depends on a
    task of its own checkpoint starts without waiting for a verdict, and tasks
    that share a file never run at the same time. Each task is committed when it
    is written, with fixes in additional commits that name it, and the work is
    published for merge only when every task is verified or accepted by you. A
    plan without checkpoints is still verified task by task (#168).

  • The UI review keeps the UI library's design system as its baseline. The
    reviewer first reads the library configuration, theme and installed
    components, the project UI guide and any declared accessibility requirement.
    A library default or a recorded project decision is not a finding in the
    checks that measure that value, such as contrast, even below a WCAG
    threshold, unless the project declares an accessibility
    requirement or the ticket asks for one. Non-text contrast follows WCAG 2.2
    SC 1.4.11 precisely, a fix that customizes the theme or an installed
    component needs your decision or the project's, and a global finding is
    reported once and recorded in the project UI guide so it does not reappear
    for every view (#173).

  • Incidental findings at a close: a P3 finding or a wording detail is no
    longer offered as a ticket, related P2 findings are proposed as one group, a
    UI finding is proposed only after it was seen rendered, and the close
    question does not recommend a ticket opened from the work's own findings. UX
    reviews report earlier problems only on the screens the change touches, and a
    deployed review whose code matches the local one starts from the local
    report (#159).

  • A task whose second verification still fails no longer stops to ask you
    when the remaining fix is small, inside the task and obvious: the
    implementer makes that one correction and a third verification checks it;
    you are asked when the fix needs your decision or that verification also
    fails (#156).

  • Fewer questions in the flows: effects beyond the local machine are
    authorized by class (local, non-production, production, critical) instead of
    one at a time. The test data of a non-production environment takes one
    question for the whole set, while each critical effect keeps its own. The
    steps of one objective are asked together and answers already in force are
    reused. Minor findings from plan reviewers are applied and listed in the plan
    summary, follow-up fix pull requests share the ticket's delivery, a condition
    on a grant lasts only for its stated scope, and a research route option now
    says whether it opens a ticket instead of a separate ticket question (#151).

  • Unattended delegation keeps running when you write only to tell the agent
    to continue, acknowledge progress or ask how it is going; any other
    message ends it (#143).

  • A costly check now has a definition: it spends money or ...

Read more

Hive 0.5.0

Choose a tag to compare

@JhonHawk JhonHawk released this 23 Oct 10:06
8c2882b

Added

  • The walkthrough viewer groups compact sessions directly under projects, with
    session counts, full descriptive titles, ticket/date metadata and a route icon.
    Its grouped toolbar shows work status and actions to archive, restore, pin and
    delete a copy. Archived copies are hidden by default and can be shown with the
    archive icon beside Projects. Archiving is independent of pinning, retention
    and manifest status. Projects initially expand only when they contain work in
    progress; manual expansion, sidebar visibility, archive filtering and selection
    survive refresh within the same browser tab and origin.

  • hive view, a local walkthrough viewer the user starts and stops
    (start, stop, status, open, serve, register). Agents register repositories and use
    hive view status --dir <work folder> to confirm closure; the server on 127.0.0.1 copies each repository's
    walkthrough.json and its files into Hive's state directory, indexes project, ticket and session metadata and presents sessions grouped by
    project in a compact sidebar, purges unpinned copies after 14
    days (--ttl), and deletes a copy on request without touching the
    repository. A new manager command raises the minor version of the next
    release.

  • flow-walkthrough renders its page from data: the agent writes
    walkthrough.json as the new references/walkthrough-manifest.md sets and
    runs scripts/render_walkthrough.py, which validates it and writes
    walkthrough.html from a fixed template with a summary, a section index and
    a capture carousel; the page's fixed labels follow the manifest's lang
    (es or en). The agent no longer copies and fills an HTML skeleton.

  • codex-question-card, an optional capability that hive install offers
    with the Codex host, selected by default in interactive installs when the
    installed Codex reports the feature in codex features list (uncheck it or
    type none to skip; non-interactive installs never run it; an older Codex
    gets a manual step that says to update it). After consent it runs codex features enable default_mode_request_user_input, Codex's own configuration writer, so Codex
    shows its question card in Default mode instead of asking in text; the step
    is verified by reading the Codex config.toml, hive doctor shows a
    Codex question card row with the record and the enable, disable and list
    commands, and codex features disable … turns it off. The text installer
    gains a default selection ("Press Enter to accept the default selection, or
    type none to skip."), an offer can show an effect instead of a version
    phrase, and the installer's boundary is stated in the tool-dependencies
    document: Hive edits no host configuration file itself and, with consent,
    runs the host's own command that writes one. A capability selected by
    default raises the minor version of the next release.

  • flow-walkthrough, a skill that owns the human validation of a user
    interface: the first rendered view of a new layout, the correction rounds,
    the captures, a short recorded pass when a flow needs one (agent-browser record, checked with ffprobe) and the handoff as a message and a
    dark-mode HTML page rendered from a JSON manifest (see below). Its reference
    ui-gate.md is the one home of the UI gate: the implementer's rendered
    inspection, the independent children with their briefs and stops, and the
    accepted-candidate gate; ui-review-criteria.md moved beside it. Build's
    verification reference now points to them instead of holding the handoff
    and the gate.

  • UI verification, an optional ## Hive key with full, functional or
    none, selects which independent children run on a built UI change; without
    it, Build asks once per repository and session in the review round.
    hive doctor and the project form accept the key.

  • The documentation site adds Getting started, Skills, and Inspiration pages
    alongside the complete release history from this changelog, with unreleased
    work distinguished from published versions.

  • A code repository can declare Delivery: automatic or Delivery: interactive
    and a Review mechanism (or none) in its ## Hive section; the agent then
    states the mode and the review in one line before the first edit instead of
    asking, and hive doctor and the project form accept the new values.

Changed

  • UI evidence is refreshed for affected views and journeys after any later fix,
    including code-review fixes after approval. A changed visible result needs
    approval again; unaffected evidence remains valid. Adopted after the Facti
    Claude Code session c574e5a1-ec3e-44da-995e-2a9a89ba2e6b of 2026-10-23.

  • Gaps found in two sessions of 2026-10-19 (Grok 01a15706, Codex 01a156af;
    issue #93). A notice that only reports the exit of a process the agent
    stopped on purpose gets no message, also while a question waits. Close stops
    the processes the task's children left running, which the agent records when
    each child returns. The implementer of a UI change loads the changed view in
    a real browser with no test doubles before reporting it. Each UI gate child
    verifies a fixed candidate, named by commit or tree hash and served from its
    own build and port; a fix during the gate makes a new candidate. The first
    rendered view and the final validation offer to open the viewer or approve,
    with corrections written in the answer's free text. When an answer leaves a
    selection to the agent, it lists what goes in and what stays out and asks for
    confirmation before acting.

  • Two costs seen in both Grok cam sessions of 2026-10-20 (01a1579f,
    01a157db; issue #94): after a context compaction the agent rereads only a
    reference whose condition applies to its next action, plus the active
    SKILL.md when its steps are not in the summary, instead of every reference
    read earlier; and it writes to persistent memory only at a user decision, a
    ready plan, a work-item close, a worktree's creation or removal, and a session
    summary when the user ends the session, a cadence that holds over memory
    tools' own save instructions. Close writes its memory records in one pass.

  • The pilot runner (tests/pilot, not installed) accepts a text-only close
    question only on hosts whose headless mode has no question tool (Codex,
    Claude, Pi); on Grok and OpenCode it now requires the native question (#66).

  • The local walkthrough viewer replaces its vanilla shell with React and public
    MIT shadcn/ui and ReUI components in a custom compact project/session sidebar.
    Its versioned bundle runs inside the Go binary without Node, a ReUI license
    or third-party network access. Frontend development uses Node >=24 and pnpm
    12.8.1, with frozen dependencies and read-only bundle freshness checks.

  • Walkthroughs carry optional explicit session.status (in_progress or
    closed); historical manifests without it remain unknown. Badges are
    independent of Git classification. flow-close publishes closure and confirms
    the exact current viewer copy before eligible source cleanup; an unconfirmed
    copy keeps its evidence. See the
    canonical close procedure.
    Update the viewer binary and installed content together, and coordinate a
    restart of an existing older viewer before using the new producers: replacing
    files does not update its running process.

  • flow-walkthrough takes captures by default and records a flow only when it
    validates an interaction or a transition that captures cannot show, or when
    the user asks; it no longer records every flow at the validation handoff.
    The walkthrough skeleton leaves flow-report, and the walkthrough's
    retention unit now includes walkthrough.json.

  • flow-walkthrough opens nothing by itself: the message gives one line with
    the page's path, and the agent runs hive view open only when the user
    chooses that option in the validation question.

  • Inspiration opens one page containing OpenSpec, gentle-ai, and Superpowers
    instead of a collapsible sidebar group.

  • Waits, messages and the first rendered view follow one rule each (the
    analysis of cam session e2c8d39f, 2026-10-18). Every progress or decision
    message opens with a bold context line, stage · ticket · progress, and a
    question card opens with the stage and ticket. A turn that a notification
    starts without changing what you need is that line followed by the change,
    without listing the work still running or repeating an open question. A
    child that returns, or that the host reports as stopped, while it waits on
    work of its own is treated as stopped and checked at once. A task brief
    carries the skills and project documents that the project's rules direct for
    that work. The first rendered view of a layout is delivered as
    walkthrough.html, and the message gives the command to open it before the question. A task
    with a rendered UI effect is committed after its first rendered view and the
    UI review, and the UI review, the full suite and, when the change crosses a
    security boundary, the security review run before you are asked to validate; later fixes go in an additional commit per
    task, never an amend. The Build, Close, Research and Plan pages describe
    this, and the delivery notes record Claude Code's notice for a stopped child.

  • The language rule names URL paths and route segments, including those of
    public pages, among the identifiers written in English: a URL path is an
    identifier, not user-facing text. hive-review-code reports a new
    identifier, file name, URL path or route segment in another language as a
    finding, and hive-verify-task notes it beside its verdicts.

  • The global guidance points to the browser-automation reference before any
    browser is driven, including recordings and demos outside a build.

  • Unattended delegation has a fourth source: the user...

Read more

Hive 0.4.0

Choose a tag to compare

@JhonHawk JhonHawk released this 15 Oct 08:35
c929924

Added

  • When a human reviewer is assigned or requested on a pull request, the agent
    reads each of their comments against the code before changing anything, asks
    in the thread when one is ambiguous, applies them blocking-first and verifies
    each one, disagrees with evidence when the code already does what was asked
    or the request would break declared behavior, and replies in each thread.

  • The completion report lists each implementation decision the agent took
    without asking, when the user could reasonably have chosen otherwise, with
    its reason and how to revert it, in attended and unattended work alike.

  • Pi user-scope installs run pi install npm:pi-subagents without a version
    pin, so npm resolves the latest release; an existing pin is left unchanged.

  • hive install offers agent-browser as an optional capability: in an
    interactive install on macOS arm64 (brew install agent-browser) or Linux
    amd64 with Node 24+ and a writable npm prefix (npm install -g agent-browser), it shows the exact commands, installs the latest version
    after your consent, and runs agent-browser install to download Chrome only
    when no browser is found. Commands run without a shell or sudo, with a time
    limit; a failed step is reported without touching the installed core.
    hive doctor now names hive install for a missing agent-browser or
    browser. Elsewhere the capability stays manual. This adds an installation
    route, so the next release raises the minor version.

  • Grok user scope installs ~/.grok/hooks/hive-read-whole-references.json, a
    PreToolUse hook on read_file that runs the hidden hive grok-hook read-whole subcommand to drop offset and limit when a session reads a
    Markdown file under the shared skills directory (at most 1000 lines and 64 KiB),
    so references are read whole. It fails open and never denies a call. State gains
    the hook record kind, which older managers reject; a plan built from a go run
    or temporary binary that would install it is refused.

  • Optional OpenCode V2 context loader, delivered inactive at user scope as
    hive/context-loader/{index.js,context-loader.json}, outside auto-discovered
    plugins/, with no package.json. Explicit
    native activation supplies canonical references only for hive-review-plan and
    hive-review-harness; it does not classify prompts, change permissions or
    enforce compliance. The manager generates its script snapshots and metadata
    schema 2 from the validated release payload; requests read no reference files
    or metadata. Canonical references retain one editorial home. Refresh requires
    regeneration and plugin reload, not editing installed Markdown. The derived
    files use existing ownership, drift and recovery machinery; their SHA-256
    digests record provenance, not cryptographic authentication. Activate the
    installed hive/context-loader directory, not the source template. Remove the user-owned plugin entry
    before uninstalling or downgrading to content without it.

  • Declared reading-footprint scenarios report unique source bytes and characters,
    including required procedures and selected activity entrypoints, separately
    from actual tokens. New offline packages use manifest v2 and include the loader
    source; the current manager still verifies v1 packages. Older v1-only managers
    cannot verify v2 packages. This optional integration meets the minor-version
    feature criterion; no release number is assigned here.

  • The route question can offer an iterative route when the work alternates
    investigation and implementation toward one measurable objective. Its option
    states the objective, how it is measured, the number of rounds the agent
    proposes, and when it stops. Once you choose it, the agent reports each round
    in one line and keeps going within local editing and verification; it stops
    and asks when the objective is met, the rounds run out, a round brings no
    measurable improvement, a step needs an effect the route does not cover, a
    round needs a plan, the scope changes, or a decision is yours.

Changed

  • Children never receive the parent's conversation history: when the launch
    tool offers to pass it, the agent launches without it, since the brief
    carries what the child needs and inherited history never delivers the skills
    the parent read. Children never launch children of their own, and a child
    that cannot finish alone reports it; the ten implementer roles say so.

  • Before calling a change mechanical, the agent checks whether the code it
    alters also serves behavior the request did not name, such as a shared
    helper, parser or option; if it does, the change alters a contract and is
    not mechanical.

  • A failure you report, such as a bug, an error or a regression, with or
    without evidence, enters through Research unless the same message asks for
    the fix: the agent confirms the cause first, then asks in one round whether
    to track it and how to continue, and Build asks the delivery and review
    questions when it starts.

  • hive-verify-task runs on its configured verifier model where the host sets
    one, as a preference rather than a condition: the agent no longer compares
    models before each launch, asks which model to use, or discounts a verdict on
    the child's own statement of its model. What makes the verdict independent is
    a fresh context and a brief without the implementer's report or claims; the
    verifier never runs on a smaller model or lower effort than the implementer,
    and when it cannot run it is retried once on the implementer's model before
    the agent asks.

  • When an MCP server or other integration the task needs fails to connect or
    authenticate, the agent says so in its next message, names what it uses
    instead and what that source cannot show, and the action that restores it,
    instead of substituting another source silently.

  • A research child cites a live observation by the exact read-only command that
    produced it, so the parent reruns it unchanged instead of rewriting the query, and
    children that share a question get disjoint scopes so none repeats another's reads.

  • The browser tool follows the situation: agent-browser for journeys, UX review,
    screenshots and measurements, Chrome DevTools MCP for audits and performance traces
    or an independent reproduction, Playwright MCP only without agent-browser. A child
    never installs a missing browser tool; it reports the install commands and the user
    authorizes once. The frontend, UI design and task-verification roles now read the
    browser rules and use only the browser tool their brief assigns.

  • git-workflow no longer prescribes a worktree location: agents use the host's own
    worktree feature or the repository's documented place, never a temporary directory,
    make sure an in-repository worktree path is ignored, and record each worktree they
    create in persistent memory so another session or host can resume it; flow-close
    updates that record when it removes the worktree. Agents create a worktree only when
    the current checkout holds others' changes or a non-base branch, another task writes
    in the repository at the same time, or the task needs two checkouts at once; otherwise
    they create the branch in the current checkout.

  • OpenCode V2 role profiles now use OpenCode Go models instead of GitHub Copilot:
    opencode-go/deepseek-v4.1-flash for execution, opencode-go/deepseek-v4-pro
    for reasoning and opencode-go/glm-5.3 for the verifier.

  • The global core retains essential boundaries and explicit read conditions while
    canonical skill references hold reporting, questions, delegation, artifact
    details and project settings. Reporting remains required before every chat
    reply, including ordinary answers without citations. Smaller global text is
    not a claim of lower total context or better model reliability; the declared
    plan/build source union can grow.

  • A research round about work you could start now ends with the route
    question directly: the routes with the flow each one runs, keep
    investigating, or stop at the findings. The agent no longer first offers to
    show recommendations. With a question card that waits for your answer, the
    message keeps the explanation and one summary line per question with the
    recommended option, and the options are in the card; without one, or after a
    dismissal or timeout, the options are written in the message, and a
    dismissed card is asked again in text. After each answer the agent says in
    one line whether it settles the question, opens the next round, or returns
    the work to an earlier flow, and what that invalidates. The size check on the
    global guidance now only reports its size and no longer fails.

  • A ready plan is summarized in the conversation, not only saved to its file:
    the objective, the decisions in force, each task in one line, the checks,
    the delivery, what is still open, and the plan's path. The summary ends with
    the question whether to start Build now or stop with the plan ready.

  • Findings from a review that runs inside work you already authorized are fixed
    as part of that work, without asking again; a review you request on its own
    still does not authorize fixes.

  • Work that will not end in a Git delivery gets the review question once, if
    that choice is still open, before the first edit instead of a separate stop
    when the review is due.

  • After the one follow-up review, a new finding whose fix is small, inside the
    change and touches no contract is fixed with a test, without another review
    round, and the report says it was fixed without review; anything larger comes
    to you as a decision.

  • hive-research returns a bounded reply: the answer, a status with its stop
    reason, each deciding claim with a locator and excerpt, counterevidence and what
    it did not check. flow-research checks those claims by opening their locators,
    gives each research child a s...

Read more

Hive 0.3.3

Choose a tag to compare

@JhonHawk JhonHawk released this 11 Oct 06:00
c988480

Added

  • An urgent production fix has its own route: it enters the base branch and is
    promoted at once through each environment, in order. When a promotion would
    carry work you are holding back, the agent offers a hotfix branch from
    production as a separate decision, only where the project's branch policy
    accepts it, and merges that same branch back instead of cherry-picking. A
    short postmortem is offered at close.

Changed

  • When the work needs its own checkout, the agent creates the Git worktree
    where the repository's guidance places worktrees, otherwise in the sibling
    folder ../<repo>-worktrees/<work>, never inside the repository or in a
    temporary directory. It names each worktree in its report and removes it
    with its branch at close. Worktrees made by a host's own feature keep the
    host's location.

Fixed

  • Environment branches after the base branch, such as qa and production,
    again receive changes only through the promotion path: no direct commit,
    push, cherry-pick, or hotfix, and every deviation is asked per push, even
    under unattended delegation.

Hive 0.3.2

Choose a tag to compare

@JhonHawk JhonHawk released this 09 Oct 06:55
9bbb57e

Added

  • The documentation site includes a new architectural Hive social preview image
    with Open Graph and X image metadata on every page, plus a branded homepage
    sharing title.

Changed

  • A message that waits on the user ends with one fixed pause line, such as
    "Esperando tu decisión:", instead of burying the choice in a report or the
    pending list. Research ends each round of findings by offering to show the
    recommendations, and a plan decision that changes the design is presented at
    once under a pause label.
  • A work item is a ticket or the scope you authorized, not each task of a
    plan: between tasks the agent reports progress and continues. When work
    continues a plan in progress, the closing question's continue option covers
    all remaining tasks and asks once for the permissions they still need, and
    the agent no longer suggests a new session in the middle of that plan.
  • While authorized work remains, the agent sends each progress line together
    with its next action, so the turn does not end on an announcement, and it
    answers a question or notification in one line and keeps working. Each build,
    deployment, or CI run it starts gets a completion signal before its turn ends,
    preferably the host's own monitor or watch tool.
  • When work resumes or continues a plan, for example after a close, the agent
    lists the remaining tasks in the host's task list again.
  • A plan's change record quotes the user's request verbatim, and the
    acceptance criteria of each ticket it names with the date they were read, so
    requirements are not lost to paraphrase on the way to the implementer.
  • An implementer of a plan task reads the task, the quoted request, and the
    acceptance criteria it closes from the change folder; the delegation brief
    points to them instead of paraphrasing them.
  • When the delivery mode makes commits, the agent commits each plan task once
    it is verified, so every commit holds one verified task and its fix rounds.

Fixed

  • Every choice the agent asks for, including decisions, routes, the close
    question, and yes-or-no offers, is written complete in the message: context,
    options with their effect, and the recommended one with its reason. The native
    question card of each host is used as a shortcut and written to stand alone,
    so the decision survives when the card or the message is hidden.

Hive 0.3.1

Choose a tag to compare

@JhonHawk JhonHawk released this 08 Oct 23:20
19c1071

Changed

  • When diagnosing a failure that can still occur, the agent first runs a command
    or observation that shows the exact symptom, and only then forms hypotheses or
    edits code. When it cannot build one, it says what it tried and asks for a
    reproducing environment, a captured artifact, or temporary instrumentation.
    Incident mitigation does not wait for it.

Fixed

  • Work folders under _support/ carry the change id with words that identify
    the work, shortened when long, never the ticket key alone; without a change
    they use the ticket key followed by a short description.

Hive 0.3.0

Choose a tag to compare

@JhonHawk JhonHawk released this 07 Oct 15:06
20d14ea

Changed

  • Unattended delegation also starts when you say you are leaving, such as going
    to sleep, while authorized work is running. It lasts until that work completes
    or you return, decides open questions itself and reports each decision with how
    to revert it. Irreversible or destructive actions stay forbidden.
  • Before handing a pending item to you, the agent checks its premise and whether
    the session can do it itself. Closing a work item also restores local state the
    task changed.
  • A restriction enters project guidance, a plan, or a question option only when a
    source states it or you decided it.
  • Questions asked in text give each option one line; shared effects and
    unavailable routes are stated once after the options.
  • The README and llms.txt link the documentation site, and the site header
    links the GitHub repository.

Hive 0.2.1

Choose a tag to compare

@JhonHawk JhonHawk released this 06 Oct 08:08
86b3c52

Added

  • Documentation site in site/, with one page per flow (research, plan, build,
    close), published at https://hive.tricell.tech/ from this release on.

Changed

  • The starlight-docs-site template pins Astro 7.3.5, Starlight 0.42.4, and
    pnpm 12.8.1, denies the esbuild install script it does not need, and ships a
    biome.json so pnpm lint ignores build output.
  • README.md, llms.txt, and pages of the documentation site are updated in
    the same pull request as any change to what they state.
  • CONTRIBUTING.md defines what the version number tracks and when a release
    raises the minor or the patch version.

Hive 0.2.0

Choose a tag to compare

@JhonHawk JhonHawk released this 06 Oct 05:47
3d882f8

Added

  • One-line installation: curl -fsSL https://hive.tricell.tech/install.sh | sh
    runs get-hive.sh, which downloads the latest release package, verifies its
    checksum, keeps it in ~/.local/share/hive/packages/, and runs its
    install.sh. --dry-run previews the installation without changing any host
    (the package is still downloaded and kept; add --hosts when no terminal is
    available), and --version selects a release (#11).

Changed

  • git-workflow asks pull-request descriptions to state whether a revert fully
    undoes the merge and what the change would affect if it were wrong.

Hive 0.1.0

Choose a tag to compare

@JhonHawk JhonHawk released this 06 Oct 03:48
f62f58f

First public release.

Added

  • Shared guidance layer for deliberate work with coding agents: authorization,
    preservation, evidence, and artifact-placement rules.
  • Activity skills: flow-research, flow-plan, flow-build, and flow-close.
  • Specialist role definitions with bounded tasks, owned paths, and evidence to
    return.
  • Go manager with the commands install, status, doctor, update, plan,
    apply, and recover. It previews destinations, checks ownership and drift,
    preserves user content outside managed resources, and supports recovery.
  • Six host adapters: Claude Code, Codex, Grok Build, Pi, OpenCode V2, and
    Cursor CLI.
  • Downloadable packages on GitHub Releases for macOS Apple Silicon (arm64) and
    Linux arm64 and amd64. Each package includes the executable, content,
    verification inventory, LICENSE, and THIRD_PARTY_NOTICES.md, and installs
    offline with ./install.sh. Installation refuses a package that lacks the
    license files, and a content test keeps the notices in sync with go.mod and
    the Go toolchain.

Removed

  • The bootstrap.sh online installer and the hive bootstrap subcommand.
    Install from a complete package or a source checkout instead. State written by
    the retired subcommand (an installer field) is still read and ignored.