The Terminus Kit

A lightweight, four-rule context-preservation kernel (CT1-CT4) for this single-unit pipeline — deliberately thinner than the god kit's full multi-worker context-thrift doctrine, because terminus has no fleet of workers to conserve orchestrator context across. Kept because a task zip, a long log, or a large repo under review can still burn a window if read carelessly. Covers delegating bulk reads, filtering pasted/secondary context at the source, keeping decision state reconstructible from disk (reviewing/, outputs/, scratch/, state/), and preferring cheap checklist/vocabulary greps over a deep doctrine re-read.

Terminus's own build/decision-failing invariants (G-*) — a path with no canonical counterpart in the god kit at all (which keeps guardrail *authoring* doctrine in a differently-named file instead). Seven guardrails, each meant to be self-audited before ./done runs rather than treated as an advisory note: no git commands inside the terminus workspace (it is intentionally not a repo), controlled-vocabulary-only labels, mandatory real Docker build-and-run verification, a required effective rubric (platform context or a bundled root rubric.md), category-masking blocked before any quality finding, mechanical corrections exceptions kept narrow, and decision commands run only on the operator's explicit instruction.

The blocking-question schema walked before any review or correction decision is finalized: which pipeline this unit is, which dated doctrine sources currently govern it and whether any conflict, whether the declared category matches the task's actual primary activity, whether real Docker verification has actually run and matches the declared internet mode, whether an effective rubric exists, whether a claimed exception is on the enumerated list, and whether every checklist item carries an explicit evidence-backed disposition. Scoped entirely to one review/correction unit — unlike the god kit's intake, there is no fan-out, worker-pool, or spec-vs-idea-stage branching here.

Terminus-native reusable review/correction patterns (KB-*): the ./rev import pipeline, the two required verification legs (static read plus Docker runtime), the separate correction-vs-review pipelines and their distinct completion semantics, the category-masking detection pattern, grandfathering's own-evidence requirement, the decision command as the terminal mechanical act once judgment is made, and infra-failure-vs-task-defect routing. An explicit provenance note states the file was staged from the god kit's doctrine on scoping derived instances but carries none of that source kit's own ids, names, or internal mechanisms.

The CANONICAL scar log for THIS derived instance only — its own closed id-space, TL1 through TL12 ("Terminus Lesson"), carrying zero L-* god-kit ids and zero EL-* Entabeni ids. Every entry was distilled as a DRAFT first cut from terminus's own AGENTS.md/CLAUDE.md doctrine on 2026-07-20, not from a lived multi-session incident history — expect reword/split/merge/renumber before it's treated as load-tested. Covers the closed category vocabulary, Docker-verified ground truth over narrative, declared-vs-demonstrated environment flags, policy-before-quality gating and category masking, named/narrow corrections exceptions, dated-doctrine conflict resolution, operator-only decision commands, infra-vs-artifact failure routing, exhaustive checklist disposition, primary-vs-secondary evidence, operational-vs-quality separation, and the fixed decision-artifact template.

The single-unit lifecycle for this instance: import, static review, runtime (Docker) verification, decision drafting, then the operator's own decision. Explicitly notes there is no fan-out/wave phase here — unlike the fleet-orchestration lifecycle this instance was derived from, that whole phase family was dropped as inapplicable to a human-gated, one-unit-at-a-time review domain.

The canonical ordered thinking model for this derived instance, ids P1..P12 (lower id wins on conflict) — twelve principles scoped entirely to terminus's own single-unit review/correction domain: empirical grounding in a real Docker run, the closed controlled vocabulary, policy-before-quality ordering, named/narrow exceptions, newest-dated-source-wins conflict resolution, operator-only decision commands executed rather than narrated, failure-class routing, exhaustive checklist disposition, primary-over-secondary evidence, the operational-vs-quality split, the fixed decision-artifact shape, and delegated reading. Each principle cites its originating TL-* scar; carries zero god-kit ids.

The /begin skill for THIS derived instance — starts one terminus review or correction unit (an ordinary task review, an oracle-failure correction, an instruction-rewrite correction, or a correction-review) the terminus way. Explicitly a single-unit, human-gated pipeline: no fleet, no fan-out, no wave dispatch. It loads AND activates the kernel canon (principles, lessons, guardrails) into a named guard shortlist, walks intake to pin the pipeline and the governing dated doctrine, runs the unit's own real import entrypoint (./rev, or ./start-cor-task / ./start-cor-review), confirms the Docker verification apparatus is actually live, walks the two-leg (static + runtime) verification and the full checklist, then hands off via ./alert and runs ./done only on the operator's explicit instruction.

The /continue skill for THIS derived instance — assesses ONE terminus review/correction already in progress and proposes the next concrete step, reconstructing state empirically from the workspace's own directories (reviewing/, outputs/, scratch/, state/, corrections/) rather than from memory or a pasted summary. It buckets the unit (NOT-STARTED, STATIC-ONLY, VERIFIED-INCOMPLETE-CHECKLIST, DECISION-DRAFTED-AWAITING-OPERATOR, or STALE), asks the operator only what the disk cannot answer, and executes the evident next step rather than offering a menu. It re-arms no fleet watcher because there is no fleet — the only standing 'watch' is the operator's own ./alert.

The /dogfood-it skill for THIS derived instance — executes the evidence re-walk itself, right before ./alert, instead of only writing the plan. It rebuilds the task's real Docker image and re-runs the oracle plus verifier in the declared internet mode (a static-only re-read of the draft is explicitly NOT dogfood here), removing the images/containers afterward, then confirms every marked checklist item's disposition against direct evidence (PASS / FAIL / BLOCKED). A mismatch corrects the draft decision directly rather than becoming a separate side note, and anything only the operator can truly judge is named OPERATOR-JUDGMENT-REMAINING. It front-runs the operator's own review; it never replaces ./done.

The /dogfood skill for THIS derived instance — produces the evidence-walk PLAN before a terminus decision is handed to the operator: a concrete, self-contained re-walk of the actual verification evidence behind the drafted decision, grounded in the specific unit (its declared category, internet mode, rubric source, and marked checklist items), never generic. The plan names the exact docker build / oracle / verifier commands to reproduce the runtime evidence in the declared internet mode, one walk step per checklist item that rests on runtime evidence, any targeted edge probes, a one-line pass bar, and a report-back rule that folds any miss back into the draft before ./alert. It is the bridge between 'checklist complete' and 'the evidence was actually walked.'

The human-readable companion to kernel/lessons.yaml — the same TL1-TL12 scars in prose, explicitly marked as a DRAFT first cut distilled from terminus's own AGENTS.md/CLAUDE.md doctrine on 2026-07-20 rather than a lived incident history, append-only from here as real cases surprise the reviewer.

The human-readable companion to kernel/principles.yaml — the same P1-P12 ordered thinking model in prose, each principle naming the TL-* scar it rests on; explicitly written as the follower of the yaml, not the other way around, per this instance's own precedence rule (lower id wins on conflict).

The orientation document for this derived, scoped instance — a stripped subset of a larger proprietary orchestration kit, carrying only what terminus's single-unit, human-gated, non-git task-review/correction domain needs, entirely in its own TL-* id space. Lays out what was kept (all 7 kernel files rewritten for terminus's domain, plus 4 command skills) and what was deliberately scoped out (fan-out/wave phases, the fleet-standup templates/tools, multi-repo profiles, admin-tiered guardrail governance, small-window mode) and why, closing with an explicit honesty note that the seed TL-scars are a draft distilled from AGENTS.md/CLAUDE.md rather than a lived operational history.