Instructions for quiet, continuous Intent semantic maintenance with the itt CLI. Ships the Skill only: no hooks, and no project is enabled by installing it.

GitHub stars
316

Install

/plugin install intent-cli@intent

Run it inside Claude Code after adding the marketplace with /plugin marketplace add dozybot001/Intent.

Contents

Discovered from source

Skills 1

Intent CLI

Preserve enough verified meaning for another agent to continue, not a conversation log. Continuous maintenance is the normal mode; every turn needs assessment, not a new Snap.

Project scope

  • For first-time setup, missing itt, or optional IntHub authentication, read onboarding.md. On first installation, ask once whether to enable real-time recording for this project, together with any needed installation approval; an explicit enable request already answers this. Install, verify, and resume the original workflow. Local recording needs no account.
  • Resolve the canonical semantic root; Git is optional. A trusted hook context for this root and turn supplies the opening snapshot. Otherwise check itt maintenance status. Read execution.md before executing commands.
  • itt init enables maintenance and installs project-local Codex hooks for a new history. Existing histories need itt maintenance on once. In Claude Code, add --host claude-code to init, maintenance on, and maintenance status. itt maintenance off disables only this project and removes Intent's Claude Code entries. Skill or plugin installation does not enable other projects.
  • When asked to enable real-time recording, finish project configuration, then tell the user to review/trust this project's two hooks: UserPromptSubmit (opening context) and Stop (closure check). Codex: Settings → Hooks or /hooks; if absent, check project-directory trust first and follow codex-hooks.md. Claude Code: trust the folder and confirm both in /hooks, restarting the session only if they are missing, per claude-code-hooks.md. Do not repeat installation; configuration alone is not proof of execution.
  • When development work starts in a project with no .intent/, run itt maintenance status once. If it returns offer: true, ask the user once, in one line, whether to turn on real-time recording here, and carry on with the task either way. On yes, run itt init and give the trust step above. On no, run itt maintenance decline, which stops the question for this project. Without offer: true do not ask. Skip this for read-only questions or when the user said to leave Intent out.
  • A warning from itt status or a notices entry from itt inspect saying the hooks have not run means nothing is being recorded by itself. Tell the user the step it names, once per session, and until then assess and record each turn yourself.
  • IntHub is optional. When requested and authentication is missing, have the user run itt auth login in their own terminal and approve the code in the browser; never request a token in chat. Read onboarding.md for the full path.
  • A project from Intent 7 answers writes with REPOSITORY_NOT_MIGRATED. Run itt migrate once when the work needs a write: it keeps every ID and backs up first. Reads work without it.
  • Disabled/uninitialized projects do not get automatic semantic writes. Explicit one-off recording remains available; initialize only when requested setup, recording, or restoration needs it. Recovery starts read-only.
  • User read-only, skip-this-turn, and disable instructions take precedence. Never automatically sync, log in, publish, or change Git remotes. Explicit push requests use sync.md; pull, fetch and clone requests use pull.md. Follow the requested sequence; planned later uploads are not current authorization.

Quiet work loop

  1. Read once. Reuse current-turn hook context; otherwise run itt inspect in the enabled root. For missing facts, run itt inspect --intent ID --history 3 on the selected Intent only. Treat truncated context as incomplete. Code rediscovery and user explanations are not Intent recovery evidence.
  2. Preserve meaningful progress. Reuse the matching goal; record verified milestones while fresh, particularly before long/risky phases. Only the coordinating agent writes. Do not record each tool call, intermediate edit, or plan as a completed fact.
  3. Close before responding. Changed open goals need accurate continuation checkpoints. With a hook token, submit one receipt below; without a token, assess closure without inventing one or claiming enforcement.
itt maintenance close recorded --turn TOKEN [--objects ID ...]
itt maintenance close no-op --turn TOKEN --reason REASON
itt maintenance close failed --turn TOKEN --reason REASON

recorded verifies actually changed objects (IDs may be derived automatically). no-op is normal when there is no continuation-critical change or the user skips recording: give a brief reason, not a filler Snap. failed acknowledges unavailable or unverified recording. Receipts are local turn metadata, not semantic objects. A Stop retry finishes closure only; it never authorizes repeating sync or task side effects.

Routine success and no-op stay silent. Briefly disclose unresolved/partial recording with successful IDs. Diagnose recoverable errors and continue after verification; do not blindly retry creates or suggested fixes. An unresolved history error pauses affected history operations, not unrelated authorized work.

Semantic boundaries

  • Intent: one coherent objective with its own outcome/lifecycle. Split independently resumable goals; reuse matching active/suspended goals when actually continuing them. No quota; do not split by query, session, file, command, commit, or implementation layer.
  • Snap: one append-only verified milestone, conclusion, correction, or checkpoint within one Intent. Split independently verifiable/supersedable conclusions; combine evidence for the same conclusion. Cross-Intent work needs separate Snaps. Skip routine logs; correct with a later Snap.
  • Decision: a rare rule binding future Intents on different problems. Explicit durable user rules are confirmed; inferred implementation choices stay local or omitted. Essential clarification is at most one short batch per workflow, not an interruption per candidate.
  • Relation: only between Intents, and only when it changes how the work resumes. A goal found inside the current one is --part-of it (the parent cannot finish without it) or --derived-from it (independent); one that takes over a goal's aim --succeeds it, or end the old goal with --into NEW; --depends-on is a real prerequisite, --conflicts-with two goals that cannot both be achieved as stated. Never relate for mere similarity. A Decision that came out of one Intent's work takes --from ID; project-wide rules stay sourceless.

Intent what names the goal, why its motivation. Snap fields preserve the verified change and reasoning. A changed open or paused Intent's latest Snap must independently answer Verified / Boundary / Next / Blocker / Constraints with --verified, --boundary, --next, --blocker (none when absent) and --constraints, not labels in what/why; inspect returns them as each entry's checkpoint (older Snaps keep them as text). An unchanged accurate checkpoint needs no rewrite.

A query ending does not end the Intent. Mark done only after verified resolution and recorded completion/deferred boundaries; cancel deliberate abandonment with a reason; suspend a real pause after its checkpoint.

Explicit recording / recovery

Record only work verified in current context, not “everything since last time”; zero writes is valid. Briefly report a one-off recording result, unlike quiet ongoing maintenance.

For recovery, inspect before old chat/code, then state goal/reason, boundary, next/blocker, and Decisions from Intent alone. Mark gaps honestly; three recent Snaps still insufficient means report the gap, not unlimited history. Merely viewing a suspended goal does not activate it.

Sponsors

Details

Marketplace
dozybot001/Intent
Category
Community
License
MIT
Keywords
intent, semantic-history, agent, continuation