Ceratops Credit Savings Analysis

Ceratops Credit Savings Analysis

Goal

Analyze completed model-call evidence for avoidable credit spend. Use full-analysis for a generic single-thread or closure request, and one named surface only when the user names it. This skill recommends controls but never applies them.

Public Action Routing

Action References

• Run the all-run fixed-surface analysis: references/full-analysis.md
• Analyze deterministic helper contracts: references/helper-contracts.md
• Analyze context and evidence reuse: references/context-evidence.md
• Analyze rework and validation: references/rework-validation.md
• Analyze tool and handoff flow: references/tool-flow.md
• Analyze instructions and reasoning flow: references/instruction-reasoning.md

Shared Evidence And Controller Invariants

• Resolve one exact source or one controller-frozen per-thread source set. The current thread is only the valid CODEX_THREAD_ID; never infer it from recency. Exact-name and recent-thread selection use the versioned source contract. An incremental closure begins strictly after the previous completed closure; active runs and the boundary run are excluded.
• For a single-thread full or standalone analysis, run python scripts/credit-analysis-workflow.py run --request REQUEST.
• On every fresh single-thread or per-thread-batch plan, validate the installed contract and bind its version and hash to controller state. Resume only from state whose immutable artifacts still validate. Require mutation authority false, use a task root under <repo-parent>/tmp/<repo-name>/<thread-name>, and keep retained evidence inside it. Analyze every available retained descendant discovered from source lineage as ordinary source runs; report unavailable references and exclude only descendants created by the current analysis.
• Planning retains complete protected evidence and read-only canonical snapshots. Full analysis treats every completed run as one semantic unit, freezes run order and UTF-8 byte counts, and divides only an oversized run into the minimum ordered input parts that each fit Luna. Calls remain attached to their run context. If seventy Luna attempts cannot cover all prepared evidence, prioritize the largest runs and their immediate successors, admit as many ordered parts as the remaining slots allow, and record the exact unreviewed remainder. Retain input bytes, planned Luna-output bytes, and actual output bytes for every run part. Reserve the controller prompt and output schema inside each proven UTF-8 input envelope.
• The end-to-end run command executes or resumes the frozen plan. Keep plan --request for planning-only inspection and execute --state for direct state-path resume. Treat controller state, evidence and manifest hashes, task identities, candidate membership, prompts, results, and attempt telemetry as authoritative. Execution never recollects the session. Never skip, repeat, reorder, or add a semantic task outside the manifest.
• Before freezing child tasks, resolve and hash the effective global and run-local AGENTS.md chain. If a recorded canonical worktree cwd is gone, use its primary checkout only after exact repository-identity verification and record the substitution. Launch Luna from each verified run cwd and Sol from the verified primary cwd; retain the effective chain hash on every task and attempt and give Sol the hash and any differing run-local rule text. Stop on an unresolved source or rule-evidence omission; never fall back to task temporary root.
• The controller validates gpt-5.6-luna and gpt-5.6-sol at maximum effort from the local Codex catalog. Luna and Sol children retain native rollout state. Every child is approval-free and read-only. The controller owns waiting, timeout, process-tree termination, non-model progress, prompts, evidence, results, and telemetry, and never spends model calls polling children. Accepted calls and attempts retain immutable hashes and resumable attempt records.
• Luna performs high-recall discovery across all five fixed surfaces together. Run up to fifteen Luna children concurrently and admit no more than seventy Luna attempts for one frozen thread tree, including corrective reruns. Launch Luna with a retained native session so a later analysis can collect it as an ordinary descendant thread. Before launch, assign every admitted run part to one of up to six Sol reviewers. Calculate each Luna's output-byte allowance from its reviewer's fixed input and remaining proven capacity, then freeze and prove every assignment at its planned maximum. If a Luna result violates its schema or allowance, rerun that task once with a smaller output allowance; if it still fails, report that run part as unreviewed and continue. Never truncate a result, detach calls from their run context, or create per-surface Luna calls.
• A temporary-control review governs only its described owner/control subclaim and does not veto an independent finding carried by the same candidate. Route every retained Luna candidate exactly once to its preassigned reviewer. Apply the unassessed-call ceiling only to the aggregate routed call set. When that aggregate exceeds the ceiling and both Sol limits permit, review only the unassessed calls with their complete run-part context and replace those classifications; prioritize this recovery over optional direct-evidence review. After the parallel reviewers and any recovery or direct-evidence review finish, run one dependent final Sol to merge their compact judgments, produce the report, and deeply verify the top three deduplicated owner/control findings against exact evidence. Each rejected Sol task receives one automatic corrective retry when the sixteen-attempt ceiling permits. After a non-final task fails validation twice, retain its exact unreviewed candidate, call, and byte inventory, exclude that inventory from final transport, and continue to the final merger. Plan at most eight Sol calls, excluding retries and corrective attempts; allow at most sixteen actual Sol invocations including initial calls, retries, and corrective attempts. The final Sol does not re-adjudicate every candidate or receive the complete source tree.
• Restore canonical identifiers and derive nonsemantic summaries, ordering, surfaces, workstreams, repeated evidence, and savings arithmetic in code. Sol adjudicators merge overlaps and temporary controls, apply recurrence and ROI rules, and classify source calls in grouped form. Persist result-size, duration, visible-token, and reasoning-token telemetry as diagnostics. Run no model bookkeeping calls; stop before execution when the finite plan is malformed or changes admitted run, part, or candidate coverage.
• The planner attempts every completed run. When the seventy-Luna cap prevents complete transport, retain exact partial-coverage records by run and part identity, record count, input bytes, candidate count, and output bytes. Continue with every admitted part and never imply that the unreviewed remainder was semantically reviewed. When eight planned Sol calls cannot preserve all accepted Luna findings, retain the exact candidate and byte inventory of the unreviewed overflow.
• Keep session evidence, accepted surface results, the append-only index, and the final machine result at their controller-retained paths. Do not echo raw session material or caller-local paths unnecessarily.
• Preserve the existing prepare-batch, advance-batch, status-batch, and finalize-batch compatibility interface for recent-thread selection and aggregation. prepare-batch plans one ordinary holistic controller per selected thread. Execute each pending child with execute, then pass its retained final result to advance-batch; never prepare or collect through a parallel child workflow. The batch-summary contract remains a lower-level interface; group similar findings for presentation while preserving each thread's findings and totals.

Common Classification And ROI Rules

• Count spend as avoidable only when available instructions, fresh evidence, stable contracts, direct helper composition, same-pass revision, or a cheap targeted check could have prevented or reduced it. Exclude ordinary model mistakes unless a concise durable producer control would materially reduce recurrence.
• Exclude calls required by active freshness, safety, verification, controlled iteration, or workflow gates. Record conversational tool-protocol overhead as necessary rather than as a helper defect. Surface passes and synthesis make evidence-backed semantic classifications; deterministic code only groups observable evidence, expands selected clusters, and validates the result. Calls with an explicit decision-blocking evidence gap remain unassessed. Calls reviewed by every relevant surface without confirmed waste or a necessary exclusion are reviewed-no-confirmed-waste; this category is neither necessity nor savings.
• Add the credit-specific evidence IDs, implementation status, call counts, recurrence, confidence, implementation cost, and ongoing-complexity fields required by the controller schema and Output Contract. Before proposing a missing control, validate its status against frozen current-source evidence for the relevant instructions, skills, automations, and helper contracts. When a durable safeguard already exists, mark the finding implemented and classify violating behavior as a compliance or runtime gap instead of proposing a duplicate control. Use Minimal only for a local one- or two-line correction with local verification; broader ownership, failure, or verification work is at least Low.
• Treat an overbroad command or tool result contract as tool-flow waste and unnecessarily selected or loaded model context as context-evidence waste. Preserve a supported overlap as secondary evidence without double-counting model calls. Mark a volume-only finding as context-volume, keep all of its call-savings fields at zero, and classify its evidence calls independently.
• Compute net calls saved per affected run as prevented calls minus recurring calls introduced by the fix, and calls saved per similar run as that net multiplied by estimated affected-run frequency. State assumptions, test ROI at the low end of the frequency range, and reject non-positive lifetime value unless correctness or safety independently requires the control.
• Report priced credit only when the controller accepted a valid caller-supplied pricing profile. Never describe token volume as monetary or credit cost without that profile.

Cross-Surface Completion

Completion Gate

• A surface is complete only when Luna has received every admitted run part, each retained candidate has one Sol adjudication, every confirmed finding and plausible risk for that lens remains in the final result, and every capacity omission is explicit. Do not require a semantic dismissal record for every call-surface pair.
• full-analysis is complete only after the frozen manifest accounts for every completed run as reviewed or exactly omitted, proves ordered non-overlapping parts and candidate routing, and records immutable Luna, Sol-reviewer, direct-evidence-reviewer, and final-task identities and hashes. Temporary-control contributions are merged once by owner/control; every retained candidate has one disposition; every confirmed finding remains; every reviewed source call has one primary grouped classification; capacity-omitted calls are excluded from semantic classification; overlaps do not double-count savings; and finalization succeeds idempotently.
• A standalone action is complete only after the selected surface result is accepted and controller finalization succeeds.

Output Contract

• Present findings directly in chat, with artifact paths only on request. Retain every confirmed finding in machine evidence. Before the detailed list, report Confirmed: N; outstanding: M; already addressed: K. Show details for every outstanding finding; include already addressed findings only when requested.
• Give each outstanding finding a plain-language title followed by:
  • Problem: two to four sentences naming the owner, concrete episode, what failed, and why the resulting work was avoidable.
  • Evidence: the affected-call count and concrete command or tool action with its observed result, plus the relevant artifact, answer, and user-correction sequence; show IDs only on request.
  • Fix: the exact durable control, its owner, and how it completes the flow end to end.
  • Verification: the exact behavior test proving every included gap.
  • Complexity: implementation difficulty, one-time implementation cost in model calls, and ongoing maintenance introduced by the fix.
  • Savings: observed avoidable calls and expected savings per similar run, with material assumptions. Distinguish reduced calls from reduced text volume; state when credit or token savings cannot be quantified.
• Do not show status labels, confidence, internal IDs, or helper taxonomy unless requested.
• Keep every finding concise, self-contained, and understandable without follow-up. Explain what happened and why the work was avoidable before using implementation jargon; define each necessary non-obvious term; name the broadest correct implementation scope and concrete next artifact or action; and omit routine operational detail.
• Present each plausible risk separately with Observed: for the concrete sequence, Unknown: for competing explanations, Why not confirmed: for the exact missing fact and why choosing an explanation would be speculation, and How to confirm: for the exact metadata or test. Do not merge risks when that hides a distinct unknown or evidence source, and do not include a risk in confirmed savings. For standalone actions, state that the conclusion is limited to the selected surface and is not a whole-thread reconciliation.
• For full analysis, report necessary, protocol-overhead, avoidable, reviewed-no-confirmed-waste, and unassessed totals separately. Report completed runs, every admitted or omitted part's record count, input bytes, output allowance, actual output bytes, evidence and output totals, and semantic coverage. Never imply that omitted runs, parts, calls, or candidates were reviewed.

Analysis-Only Boundaries

Research Boundaries

• Use frozen local evidence first. Run only a targeted official-source check when a concrete finding depends on current external behavior. Do not perform deep or broad research; when broader research is required, report the exact uncertainty and a concise paste-ready research prompt as the concrete next action.
• Treat intentional full skill-body injection as required runtime context, not avoidable spend. Never recommend changes to reasoning settings or levels.

Boundaries

• Never modify the analyzed prompt, helper, script, skill, instructions, repository, automation, workflow, or tool configuration. Route any later implementation through the owning lifecycle after a separate execution request.
• Collection and synthesis are internal controller phases. Do not expose Luna chunking, consolidation, collect, reconcile, synthesis, apply, or modify as public actions.
• Stop blocked when a selected source cannot be resolved, the completed-run selection is invalid, controller evidence is stale or mismatched, or required semantic evidence is unavailable. Do not substitute visible conversation context for controller evidence.