execution-efficiency-i0-cost-baseline.md

I0 completion: execution-efficiency cost baseline

Status: complete for the initial cost-ledger increment. This does not qualify I1–I6 or claim that the entire execution-efficiency design is complete.

Implemented contract

• Each admitted attempt records its host role, turn, retry attempt, reason, normalized terminal outcome, observed/unknown usage and start/first-event/first-public-text/total elapsed time. Preparation, rendered projections, admission and confirmed vendor cache usage are distinct.
• Codex and OpenCode renderers report actual part bytes and Unicode scalar counts. Disjoint encoded spans are attributed to policy, schemas, exact-source blocks, conversation/mixed bodies, derived state, live control and framing/encoding. All seven classes occur in this real campaign; byte/character partitions sum exactly to their rendered parts. Attribution uses roles, block types and fixed host markers, not a claim to understand mixed text.
• Measurements reuse encoded buffers and borrowed JSON spans. They do not serialize the whole request again or retain prompt bodies. Prefix hashes remain local; common-prefix bytes are a lower bound and never a provider cache-hit claim. Vendor-owned framing remains outside these named adapter parts; absent adapter measurements are explicitly unavailable.
• Native OpenCode correction projections have a separate reason/outcome and preserve observed native usage before logical-request combination, including a later failure. Native counters must not be added to already combined logical counters. Unobserved native usage stays unknown.
• Context records include assembly/cache/index lookup work, environment refresh work, checkpoint commits and artifact store/read/write/verification/finalization/reuse counters. Store counts describe their documented boundary, not operating-system physical disk traffic. Managed fingerprint work also has the separately retained real Git capture benchmark described in the implementation ledger. No full-disk-I/O total or unseen synchronization count is invented.
• Host tool execution time includes top-level child work/waiting and must not be added again to child timings. First useful action is the first changed file or completed declared external operation; chat-only cases have no such action. Peak process-tree RSS is sampled, not exact.
• Diagnostics are content-free and non-authoritative. Disabling them removes their sidecars and optional rendering/state work while preserving canonical views, tools and accounting.

Current-build real-operation qualification

The retained cost-ledger-replay.py drives actual ProductRunner, current Codex inference, host file tools, local memory, real cancellation and native failure. Its audit-usage companion uses the production canonical C5 decoder on durable provider envelopes, independently of the new efficiency JSONL. No mocked provider or synthesized success is used.

All seven cases passed. The two paired chat/review tasks use identical source bytes, task, workspace path, model and effort per pair, with fresh private run state. Configuration order is reversed for review. Both configurations have byte-identical initial canonical provider views, exact independent answers and unchanged files. Diagnostics-off creates no efficiency sidecar. Reported cache acceptance is measured separately; no cold vendor-cache assumption is made.

Case	Canonical responses	Input tokens	Cached input	Output tokens	Elapsed seconds
chat / diagnostics off	3	47072	15872	359	23.679
chat / diagnostics on	3	47117	7936	438	22.011
review / diagnostics on	2	31382	0	205	12.523
review / diagnostics off	2	31378	0	247	15.078
unavailable-model	1	unknown	unknown	unknown	2.444
cancelled	1	unknown	unknown	unknown	4.061
managed	7	146178	0	1205	68.745

Seventeen successful responses report 303,127 input, 23,808 cached input and 2,454 output tokens in total. The failed and cancelled responses have unknown usage, not zero cost. Fourteen admitted attempts have diagnostics; the other five successful responses intentionally run with diagnostics disabled. Independent per-response high-water sums match cumulative accounting. Optional cache-creation, reasoning, tool, total-token and cost fields remain unknown when absent. Retries are recorded as zero where none occurred; no malformed provider response is fabricated.

The managed artifact case reaches Accepted through real writing, deterministic CSV validation and independent review. It preserves exact requested bytes, the project manifest and original Git HEAD/index. The first actual workspace write is observed at 22.530 seconds; acceptance is observed at 68.745 seconds. Its seven provider attempts consume 67.706 seconds, and four measured top-level tools consume 0.034 seconds. These are boundary-specific observations, not a universal speedup claim. Its sampled process-tree RSS peak is 220,084 KiB. Sampling every 50 ms can miss short-lived children. A formatting check overlapped initial campaign startup; latency rows are observations, not a controlled estimate of instrumentation overhead.

Completion evidence

• cargo build --locked --jobs 2 -p peritus-daemon -p peritus-cli -p peritus-tui passes.
• Affected all-target/all-feature strict Clippy passes, including both provider adapters, the context/agent owners, workspace runtime, product runner and daemon.
• cargo fmt --all -- --check and git diff --check pass.
• Real output, source-preservation, canonical-view equality, telemetry partition/privacy, failure/cancellation and independent usage-reconciliation oracles pass.
• All seven attributed component classes have nonzero actual rendered contributions. No fixture nonce, source body or absolute workspace path occurs in the efficiency logs.

Evidence root: /home/doll/.cache/peritus-efficiency-dif0njrm/candidate/. The final campaign is i0-completion-ledger/result.json; raw cases, private traces and exact inputs remain alongside it. Build, Clippy and formatting records are i0-final-build-r2.log, i0-final-clippy-r4.log and i0-completion-fmt.log. i0-completion-source-identity.json and i0-completion-binaries.json record exact identities. The ordinary consumer SHA-256 is 9ec51be1e413bf7ff66cff2699f7dc4d90141c0aebf4745bc1dfde4f8e547c6c. The source inventory is a hash record, not a claim that every file was audited.

Coverage and independent rollback

This initial ledger is qualified on Linux with actual Codex chat, explicit review, a managed accepted edit, model-unavailable failure and in-flight cancellation. Live OpenCode correction/ layout qualification, other providers/platforms, module/interactive journeys and broad performance campaigns remain explicitly outside this evidence and belong to the affected later increments. Their unavailable measurements are not converted to passing samples.

Set PERITUS_EFFICIENCY_LOG=0 or off to disable diagnostics. This does not select a different provider, alter permissions or prompt contents, change checkpoint boundaries or erase history. No optimization was enabled by this qualification, and no repository commit/push was made.