CLI reference#

The CLI is the primary ContextLattice interface. It keeps scope, evidence, corrections, and completion explicit while hiding lower-level adapter plumbing during ordinary work.

Normal task lifecycle#

contextlattice context "<task>" --project <project> --pretty
contextlattice resume --project <project> --pretty
contextlattice remember "<checkpoint>" --project <project> --pretty
contextlattice correct "<note>" --category useful
contextlattice finish "<verified result>" --project <project> --success --pretty

Use one stable project name throughout a task. The CLI automatically associates the latest pending retrieval outcome when finish is called.

Context#

contextlattice context "prepare the next release" --project contextlattice --pretty

This is the default pre-planning read. It returns a bounded agent_packet.v1 with the objective, ranked evidence, uncertainty, risks, and next action. Use raw or file output only when a programmatic consumer needs it.

Resume#

contextlattice resume --project contextlattice --pretty

Resume reconstructs the latest bounded objective state. It is preferable to replaying an entire transcript into a new session.

Remember#

contextlattice remember "local gates passed at commit abc123" --project contextlattice --pretty

Persist decisions, verified results, blockers, and exact next actions. Do not use checkpoints as a dumping ground for logs or speculative narration.

Correct#

contextlattice correct "the referenced branch was superseded" --category superseded

Available correction categories include useful, wrong, stale, and superseded. A correction should explain the evidence mismatch clearly enough that a later retrieval can avoid repeating it.

Finish#

contextlattice finish "release built and smoke-tested" --project contextlattice --success --pretty

Choose --success, --repair, or --failure to match the verified outcome. Completion is an evidence record, not a confidence score.

Readiness and adoption#

contextlattice doctor --pretty
contextlattice_adopt status --pretty
contextlattice_adopt profiles
contextlattice_adopt proof --agents codex --skip-provider-smoke --pretty

Use doctor when runtime readiness is unclear. Use adoption commands when a new machine, account, or repository needs profile-aware integration.

Agent integration#

contextlattice_adopt integrate \
  --repo . \
  --agents codex,claude-code,opencode,hermes-agent,hermes-ultra,omp,mercury-agent,pi,droid \
  --check \
  --pretty

Remove --check to write or update managed instruction blocks. Existing user-authored content outside those blocks is preserved.

Capability discovery#

contextlattice_skills_index search "browser automation" --pretty
contextlattice_skills_index discover "browser automation" --pretty
contextlattice_skills_index stage owner/repo@skill --pretty
contextlattice_skills_index refresh --due --pretty
contextlattice_skills_index promote owner/repo@skill --yes --pretty
contextlattice_agent_adapter profiles

search scans configured active roots without loading every skill body into startup context. discover normalizes results from Vercel's npx skills find without installing them. stage clones a bounded GitHub-backed source into quarantine, records its commit and digest, and scans for secrets and hazardous instructions. refresh --due updates registered quarantine candidates only; it never changes active skills. promote is the only activation path and requires --yes. Review findings additionally require --accept-review; collisions require --replace and preserve the previous directory as a recoverable backup.

The default refresh interval is 24 hours (CONTEXTLATTICE_SKILLS_REFRESH_INTERVAL_HOURS). Invoke refresh --due from an operator-owned scheduler when periodic checks are desired; ContextLattice does not create a background job implicitly.

Skill efficacy review#

contextlattice_agent_tools skill-evolution usage-record --payload-file usage-stage.json
contextlattice_agent_tools skill-evolution efficacy-review --payload-file review.json

A usage receipt advances through searched, selected, invoked, and verified_outcome. Search affects discoverability only. Efficacy requires the same project, agent, and session in both the Utility Ledger and a matching session outcome event. Reviews can retain, abstain, or create an inactive bounded-note, revision, or retirement candidate. They never edit or activate an installed skill, and third-party improvements remain local-overlay or upstream-PR candidates.

Notes and revisions require three verified baseline uses, three disjoint exact-matched holdouts, positive utility lift, no material regression, an exact current skill digest, novelty, and fixed size limits. See the full receipt and review contract.

Advanced packets and traces#

contextlattice_synthesis_pack "release risk" --project contextlattice --pretty
contextlattice_agent_trace --session-id <session-id>
contextlattice_agent_trace --session-id <session-id> --preset proof
contextlattice_agent_trace --session-id <session-id> --preset export

Use synthesis packs when a complex decision needs findings, topic gravity, graph bridges, constraints, next actions, and open questions over the same bounded evidence. Use traces to inspect the run-shaping trail; missing links remain visible rather than inferred.

Exit discipline#

Interactive terminals default to readable output; pipes and redirected output stay compact for automation. Override with CONTEXTLATTICE_CLI_OUTPUT=pretty|compact, --pretty, or --raw. Commands should still be judged by their exit status and concrete fields. Keep secrets out of pasted examples and logs.