Skip to content

AI capacity and weekly delivery

Status: collector and Home card live, 2026-09-16. Delegation policy adopted 2026-09-21; SOP document remains draft pending a full-cycle walkthrough. Owner: Administration.

Angel sets business priorities and commercial commitments. The orchestrator maintains a small ready queue, assigns bounded work, reviews diffs and evidence, and reports outcomes. Existing memory, notes-sync, sandbox and SOP rules continue to apply.

Where to look

Need Source of truth
What do I owe? Home and Kanboard; SOP-017 governs tasks
Procedures SOPs
Decisions and plans from anywhere Notebook backed by ~/Notes Git
Hardware design and revision EmbeddedEra product Git repositories, product.yml, source files; engineering portal is a generated view
AI allowance Home usage card and ai-capacity status --json; timestamps and unavailable/stale states included

Server in plain language

Plex is the shop server at 192.168.5.15. The live home page comes from ~/assistant/home.py and static/home.* through assistant.service on loopback port 8200, with Cloudflare Access providing verified identity. Kanboard holds the full task board. Home shows links and waiting actions; the notebook task export is only a snapshot. The assistant Git repository currently has no remote: its code history is local.

The engineering portal is a separate generated MkDocs site: local tools discover product repositories and publish Cloudflare Pages. The notebook and SOP sites also publish to Pages. Changing a product's Git files, changing the live home application and publishing a notebook page are three distinct deployment paths. The server map records operational details; this note is orientation, not another configuration registry.

Verified task snapshot

Read directly from Kanboard on 2026-09-16: 153 open tasks. The tasks.md export was dated September 14 (135 open), and the task-board plan's opening status still described an unbuilt service. Use live tasks when deciding work; a plan's heading is not deployment evidence.

Priority Existing task / deliverable Human decision or AI work
1 #23 payment rails and payment terms Owner completes account/identity decisions; AI prepares configuration checklist and invoice changes
1 #80 CPA meeting; #89 terms and warranty Assigned owners settle commercial/legal decisions; AI organizes questions and draft documents
2 One fixed-scope offer from first-money Owner selects buyer, scope and price; AI prepares one reviewable offer page and proposal
2 Existing robot customer: test service or repeat-build proposal Owner chooses the conversation; AI assembles evidence and a draft, does not send it
3 #148 engineering portal and workflow standardization Finish current tested rollout; unknown manufacturing data remains a release blocker
3 Capacity collector and home card (#235) Current requested tool; native provider interfaces, no inference just to poll usage
Decision #87 platform-work budget Proposed weekly allocation below; owner can adjust
Later #31 affiliates, #101 lifecycle watch, #172 channel adapter Pull only when they support a chosen offer or existing customer need

The product-business proposal misidentifies #47 as a new storefront. Live #47 is an existing eBay redirect task; do not repurpose it silently. Proposed prices and market assertions in older notes are not verified sales evidence.

Proposed weekly allocation

Start with 60% of planned work on offers, paid delivery and sales preparation; 25% on quality/reliability needed to deliver; 15% on internal systems. These are work-time allocations, not token quotas. Finish the current workflow and capacity deliverables before taking another large internal platform feature.

Use the adopted adaptive parallel policy and ready queue in SOP-025; the earlier one-worker/three-ready-task limits are superseded. A ready task has an owner, inputs, deliverable, verification, dependencies and a bounded brief. If a task needs a human decision, present that decision and continue a different ready task.

Daily: choose one revenue outcome and one unblocker. At handoff: report verified changes, live URLs/commits, unfinished work, allowance readings and the next owner decision. Weekly: count offers published, conversations actually held, quotes sent by an authorized person, paid orders/deposits, and delivery completed. Token utilization alone is not a business result.

Existing background usage

Checked local cron/timers and scripts on 2026-09-16: training-sync runs three times daily using Sonnet; notes-weekly runs Sunday evening using Opus with an 80-turn cap; task-verify runs every five minutes and invokes a one-turn Haiku judgement when machine checks are insufficient. These jobs already compete for capacity and must be included in headroom planning. They were not disabled or retargeted during this change. Model calls must be distinguished from deterministic exports/builds: a five-minute timer does not necessarily mean an inference every five minutes.

Before adding an autonomous work loop, measure these existing jobs over a week and review whether the Opus consolidation needs that model. The new collector is observation and advice, not yet a universal throttle for all those scripts. A usage card alone cannot guarantee avoiding a limit.

Capacity policy

The repeatable procedure is SOP-025 AI Work Allocation. The adopted v3.0 policy targets 90–95% useful weekly consumption, with standing authorization for small scoped jobs toward the remaining allowance near reset. Follow its shorter-window, freshness and completion-margin rules. Collector/dashboard advice may still use older thresholds; changing that code is separate work.

The collector uses Codex's documented account/rateLimits/read, Claude Code's status-line input, and the installed Claude CLI's experimental native get_usage control request. It does not scrape credentials or estimate remaining allowance from token counts. Claude's status line is an observation, not an always-live query; headless sessions may not populate it. Missing/stale readings must remain visible. A refresh does not launch an LLM turn.

References: Codex App Server, Claude status line. Provider capabilities and limits can change; check the actual installed versions and returned fields.

Installed tool

Private repository, deployed collector v0.2.0 commit 47ed6f4; Home integration commit 496dc9c in the local assistant repository, with a reproducible patch stored in the capacity repo. Permanent checkout: ~/ai-capacity. Administration sees the card at Home.

/home/angel/.local/bin/ai-capacity status
/home/angel/.local/bin/ai-capacity refresh
/home/angel/.local/bin/ai-capacity claude-usage
/home/angel/.local/bin/ai-capacity status --json

Codex refresh runs every five minutes using a systemd user timer and no model turn. Refresh on Home checks both providers. Native Claude collection uses get_usage without a prompt, tools or MCP servers, with per-process hooks disabled and no session persistence; normal Claude settings remain intact. This interface is experimental and can return cached data. Checked and observed times remain separate; repeated identical native observations do not renew freshness. The existing status-line collector remains a fallback.

Fable appears separately with its own percentage and reset time. It shares Claude's overall weekly allowance, so its percentage is not additional budget. Model-specific restrictions do not automatically bar other models with shared headroom. The latest deployment read reported Claude session 5%, weekly 58%, Fable 94%; these are historical snapshots, not current promises.

Consume one Codex reset credit requires an explicit admin confirmation. The server validates role, trusted browser origin/header/JSON, a fresh positive credit count for new requests, and a persistent UUID. Uncertain outcomes never expire into permission for another key. Reloading can recover the same pending request or a completed outcome; a known completed key returns its saved result without another provider call. Sandbox always refuses. There is no reset CLI command and no automatic redemption.

Verification: 151 collector tests, 35 control tests, six existing endpoint tests, isolated ASGI checks (401/403/405/422, origin/sandbox/confirmation guards, cached outcome replay), and fully intercepted 375-pixel browser tests (Refresh, cancel, same-key retry, lost-reply reload, stale disable, missing Fable, unauthorized hiding). Served JavaScript matched source after deployment; assistant service healthy; API rejects unauthenticated reads. Provider-mutation tests now run with OS network denial, empty provider homes, blocked real CLI names and explicit fixture paths. CI run.

One Sonnet 5 worker implemented bounded changes, with Codex review, corrections and deployment. CLI-reported invocation estimates were $0.6622, $0.0925, $2.8527, $5.2698 and $3.7711; recorded as reported, not asserted additive or subscription charges. The first invocation also reported Haiku 4.5 usage. Worker execution tools were removed after the incident below; subsequent worker access was file editing only, with tests run by the orchestrator in the network-blocked runner.

Weekly rhythm and burn list (proposed 2026-10-19)

Goal: use both weekly allowances (Claude Max $100, Codex plan $100) usefully every week. SOP-025 governs: 90-95% by the reset, small checkpointed jobs in the last 24 hours, no invented work, never the API key or bought credits. The limit is usually ready, approved work and review time, not tokens. Docs facts: codex-docs. Codex Pro tiers show no 5-hour limit, only a weekly allowance (confirm on /status).

When (day 0 = each pool's reset) Do
Day 0 Read both pools (ai-capacity status, ai-budget status). Queue 6-10 ready tasks (SOP-025 step 2): at least two bounded Codex implementation jobs and two Claude scoping/review jobs, plus any approved burn-list items.
Days 1-4 Spend at pace. Codex implements in its own git worktree (-s workspace-write), Claude scopes, reviews and releases. Each morning ai-budget status; a pool more than 10 points ahead hands work to the other.
Day 5 Review day: integrate one target at a time, cross-review (codex review --base on Claude's diffs; Claude reviews Codex's patches), update the outcome ledger.
Last 24 hours Run burn-list items with a fresh reading, a checkpoint and a stop switch. Stop when useful work or review time runs out, not when a number is hit.
Reset Add the scorecard line to the Log: % used per pool, accepted outcomes, rework, and the cause of anything unused (no ready work, blocked decision, provider limit, host limit, schedule gap, review backlog).

Burn list (jobs that are cheap to approve once and always worth something). A backlog entry is not authorization: each item needs one owner OK before it runs unattended; until then it is a proposal.

Job Pool Mode Owner OK
Cross-review of the week's merged diffs Codex (codex review --base) / Claude read-only pending
Regression tests for ~/assistant and the site (check-seo, doc-check) Codex worktree, patch only, no commit pending
sop-audit findings turned into draft SOP revisions Claude Sonnet read-only drafts; SOPs are walked on real jobs pending
Spanish read of /es/tools/* pages (site note Open) Claude Sonnet drafts the owner reads pending
kicad-check / kicad-overlap-check sweep of open boards, one-page summary Codex Luna read-only pending
Datasheet extraction into hardware/ for BOM parts without a note Codex Luna read-only plus drafts pending
Vault hygiene: memory-check, vault-lint, duplicate-note merges Claude Sonnet patch via PR approved 2026-10-19 (owner ran it interactively; direct commits via notes-sync)
Container image and exposure review (digest-pin proposals) Claude read-only pending

Open

  • ~/.local/bin/job-fixer (line ~229) escalates to Opus but records no api-credits outcome, so its $/pass is unknown: wire the outcome call.

  • ~/assistant/daily_review.py resolves sonnet to claude-sonnet-5 because the Agent SDK uses its bundled CLI 2.1.259: pass cli_path='/home/angel/.local/bin/claude' in ClaudeAgentOptions (sandbox branch first), then confirm the ledger shows claude-sonnet-5-5.

  • Walk adopted SOP-025 v3.0 through a real weekly cycle before completing its document-control transition from draft; policy authorization was given 2026-09-21.

  • Confirm the chosen offer/customer and approve a platform-work allocation through task #87.
  • Resume preserved hardware workflow work: CI review corrections, title templates, portal fixes, rollout and release-gate proof; implementation/STATUS.md in the engineering-portal workspace records the checkpoint.
  • 2026-10-19: Codex models per the app picker: 6.1 Sol, 6 Astra, 6 Sol, 5.6 Sol, and under Other models 6 Luna, 5.6 Terra, 5.6 Luna (Terra and 5.6 Luna are the older line). Route by the current picks (6.1 Sol complex, 6 Luna bulk, 6 Astra hardest); confirm the CLI ids with the codex model list on Plex and update the ledger.
  • 2026-10-19: Approve or strike each burn-list job above, then record the first weekly scorecard line at the next reset.

Worker outcome ledger

Use this small evidence ledger for future routing. It records outcomes, not raw prompts or transcripts; cost is included only when the CLI reports it and remains an estimate.

Date Model Bounded job type Outcome after Codex review Routing lesson
2026-09-22 claude-sonnet-5; CLI also reported claude-haiku-4-5-20251001 Existing daily research routine, 20-minute outer limit Completed in 399.6 seconds; $1.1710 CLI list-price estimate. Draft needed factual corrections despite citing a current primary source; primary corrected and deployed review-only preview. Wrapper also allowed an unnecessary nested Agent word-count request Research execution success is not factual acceptance. Enforce tool inventory and prevent recursive delegation in the legacy wrapper; use deterministic counts. Details in content-pipeline
2026-09-18 Haiku Repository inventory/report Did not finish or write the requested report; stopped rather than retried repeatedly Use for short inventories only, with a small explicit stop condition; do not depend on it for a multi-file handoff
2026-09-18 Sonnet KiCad audit/implementation attempt Reached its turn limit and produced no usable diff Keep implementation briefs narrower; prefer a read-only design audit before granting a write task
2026-09-18 Sonnet 5 Task #1 cache-to-symbol metadata audit Stopped after 6m11s with no requested report; despite an explicit no-delegation instruction it spawned three subagents. CLI estimate: $1.1096 Do not use Plan mode for a small file-backed report; deny agent spawning structurally and require the answer on stdout or use a tightly bounded direct read
2026-09-19 Sonnet (Claude Code worker agent) Tester API server (≈820 lines + 30 tests), detailed brief + two mid-run amendments Finished and committed; all its tests passed. A live run on the sandbox ERP then found 4 real bugs its fake ERP couldn't (template description, duplicate posts on retry, notes overwrite, sync race), fixed by the orchestrator Good fit for a well-specified service. Always add a live sandbox-ERP run: a fake ERP only proves what the brief already knew
2026-09-19 Codex gpt-6-astra (low effort) orchestrating two Claude Sonnet sub-runs Tester station client (≈870 lines + ≈1,470 test lines, cross-platform) Hit codex-worker's 1500 s timeout with no final message, but the work was complete: the Sonnet writer, a Sonnet reviewer with 10 findings, and the fixes all applied, 391 tests. The live test against the sandbox server passed first time. Sub-runs ≈ $5.33 list-price estimate on the Claude allowance Codex + Claude sub-agents is Angel's standing preference (2026-09-18) and worked well here. Set CODEX_TIMEOUT ≥ 3600 for multi-part briefs, and have the brief ask for a commit before any review loop so a timeout can't strand uncommitted work

Log

  • 2026-10-22: Angel asked to double the limits: autoCompactWindow 200k → 400k (settings.json + env), ai-budget CTX_LIMIT 120k → 240k; SOP-025 v3.6. Watch the next weekly ccusage line for the cost of the larger re-reads (10-19 baseline: Sonnet avg context 118k).
  • 2026-10-19: ccusage re-run vs the 10-08..16 baseline (Claude + Codex list-price): baseline $1,104 with Opus 67% of spend (Claude-only about 80%), 97% cache re-reads; since the 10-16 reset (10-17..19, partial) $214 with Opus 1%, Sonnet 5.5 39%, Codex 55%, Fable 4%, cache re-reads 96%. Transcript-measured average context per call: Opus 351k (63% of calls over 200k) before; Sonnet 5.5 118k, 0% over 200k after. So autoCompactWindow 200k and the Sonnet default took effect; no further action. Daily spend fell from $210-237 (10-13..15) to $90-113 (10-17, 18). Explore agent probe (explore.md override): all 4 calls on claude-haiku-5-5.

  • 2026-09-18: #101 read-only inventory used Claude Haiku 4.5 (claude-haiku-4-5-20251001), CLI list-price estimate $0.16017755, not subscription spend. Finished but design recommendation rejected: it searched library fragments instead of the existing ERP BOM path and incorrectly claimed Kanboard integration was missing. Independent review confirmed _kanboard.py already exists and BOM reporting identifies ERP assemblies. Useful only for identifying cache reuse; next implementation needs a tighter Sonnet brief. Codex weekly observed 90% on this turn; routine work paused at SOP-025 reserve without redeeming credits. Claude shared weekly 66%, session 3%; Fable-specific 94% is separate and not Sonnet eligibility.

  • 2026-09-16: Angel requested a native usage tool in an EmbeddedEra repo and usage visibility on Home, while preserving existing memory rules. Read live Kanboard and inspected Git/server paths. Codex returned a weekly main window plus separate Spark windows; no main five-hour window was supplied. No purchased credits or reset credits were consumed by the read.

  • 2026-09-16: Deployed capacity tool and Home card; task #235 submitted for verification with source and evidence. Preserved existing memory/sync rules, all unrelated assistant tests, and unfinished hardware workflow changes. Last installation observations: Codex main weekly 7%; Claude session 31%, weekly 57%; these are historical snapshots, use the live card for current values.

  • 2026-09-16 23:49 EDT — Development incident on task #236: Sonnet worker accidentally invoked the real Codex reset client while intending a fixture smoke test (FAKE_CODEX_MODE did not redirect the executable). Independently verified reset-credit count fell from 3 to 2 and main weekly usage reset to 0. User informed immediately; no attempt to conceal or reverse the provider record. This was not an owner-requested redemption. Stopped the worker, removed its shell tool access, and established a seccomp network-blocked test runner with empty provider homes and blocked CLI names before further tests. Unnecessary CLI redemption entry point is being removed; the requested dashboard action remains explicit and confirmed. Evidence stays in workspace planning/controls-implementation.json and planning/controls-offline-tests.log. Do not infer that the credit can be refunded; recovery is unknown.

  • 2026-09-17: Deployed Home Refresh, automatic countdowns, Fable meter, and explicit recoverable Codex reset control for task #236. Collector 47ed6f4, assistant 496dc9c. Confirmed two reset credits remain after the disclosed development incident; no further real redemption was used for verification.

  • 2026-09-19: Store implementation observed Claude CLI 2.1.274 with explicit Sonnet 4.6 ignore nonzero MAX_THINKING_TOKENS=4096: one interrupted pass reported 20,038 thinking tokens with no edits. MAX_THINKING_TOKENS=0 produced bounded concrete edits; allow enough output for file tool calls (5,000 was too short for a full plugin). Evidence: /home/angel/store-build-artifacts/fulfillment-final-fix.json and fulfillment-concrete-fix.json. This is an observed CLI/model behavior, not an authorization to change global plugins or exceed usage limits.

  • 2026-09-19: Claude Haiku read-only scout inspected free Tiered Pricing Table7.1.7. Completed; correctly identified fixed tier support and product editor, but mislabeled Woo v3 REST as Store API and suggested unnecessary SDK flags. Primary used source review and native Store API/checkout tests to verify independently; no SDK bypass flags used. CLI-reported list-price estimate $0.3987416 (not subscription charge). Primary implemented small integration/configuration edits; no bulk worker needed.

  • 2026-09-19: Codex gpt-5.6-terra source worker implemented explicit local test checkout guard/MU safety/theme heading cleanup (611f797). Completed source-only and honored no runtime mutation restriction; primary ran native tests. Missed Woo Blocks reading gateway settings directly; primary added narrow settings filter and browser checked. CLI cost unavailable.

  • 2026-09-19 storefront catalog: one source-only gpt-5.6-terra high worker, no recursive delegation. Primary review caught widget-dependent tracking, guest-link dedup and unbounded category queries; worker repaired. Primary fixed PHP test cleanup syntax, restricted recorded views to public catalog products, and independently ran native fixtures and14browserchecks. CLI cost unavailable; no cost estimate claimed.

  • 2026-09-19: Recent-history fix delegated to one gpt-5.6-terra worker, source-only commit733f066. Primary independently ran10nativefixture checks,14browserdiscovery checks and7mixed-checkout assertions; corrected a browser accessible-name selector to use actual product destination. Worker CLI cost unavailable.

  • 2026-09-19 navigation redesign: one gpt-5.6-terra source worker; main supplied reference research, help-page content and acceptance. Review tightened global category query limits and direct mobile Products access; primary independently ran native6checks, browser124+6checks and reviewed screenshots. No CLI cost available.

  • 2026-09-21: Proposed aggressive useful weekly utilization with two independent workers, a conditional review slot, a ready queue and phased scheduled execution in ai-delegation-plan. User clarified Claude plus lower GPT workers. No launcher, concurrency policy or schedule changed. Tool-free Claude Sonnet 5 critique completed (~30 s; CLI list-price estimate $0.0669256): useful isolation/review suggestions; rejected its unverified Kanboard authentication assertion and 90% cap recommendation that did not fit the requested goal. Raw result: ~/delegation-plan-artifacts/sonnet-review.json.

  • 2026-09-21: Expanded ai-delegation-plan at Angel's request to include Fable from first assignment, adaptive three-helper pilot/later larger pool, separate inference/build/review limits, and a real-task routing scorecard. Parallel bounded tool-free planning reviews: Claude claude-fable-5-1 (22.5 s, CLI list-price estimate $0.30848275) and GPT gpt-5.6-luna medium (cost unavailable). Both useful after primary review; rejected Fable's broad duplicated benchmark and simplistic automatic scaling, and kept GPT's model ladder provisional. Evidence: ~/delegation-plan-artifacts/fable-routing-review.json; GPT result in session. No operating rules, launchers or schedules changed.

  • 2026-09-21: Angel adopted the delegation rules for all chats. Updated global Claude/Codex routing, role profiles, shared memory and SOP-025 v3.0; superseded one-worker/150-line/mandatory-scout rules and old weekly stop thresholds. Policy is effective; full-cycle validation, launcher hardening and unattended scheduling remain separate work.

  • 2026-09-21: Lower GPT gpt-5.6-luna independently reviewed adopted rule edits (cost unavailable). Primary confirmed memory reconciliation had already removed the stale wording seen during review; clarified Fable/shared limits, added an explicit six-helper policy ceiling subject to lower runtime limits, and required effective launch-time permission checks. Global route/mirror checks and Git whitespace check passed.

  • 2026-09-21: Installed bounded background AI draft/review runner in ~/ai-work-runner, tied to explicitly authorized jobs under existing Kanboard tasks. User timer enabled at :15/:45 hourly; no automatic source edits, task closure or messaging. Twenty-four offline tests and native Landlock read/write/descendant probe passed. Claude Fable resolved claude-fable-5-1 and requested GPT gpt-5.6-luna passed tool-free capability probes; Fable CLI list-price estimate $0.07137275, GPT cost/resolved model not exposed. Core implementation and independent backend audit used gpt-5.6-terra; Kanboard inventory used gpt-5.6-luna (costs unavailable). Primary corrected output validation, narrowed filesystem access, pinned CLI hashes and added subscription/cleanup checks after review. Status: ai-background-status. General-purpose autonomous source changes remain outside this runner.

  • 2026-09-21: End-to-end systemd run for job 07777112-074d-490d-83ba-aaed468e2100 succeeded with requested gpt-5.6-luna; primary reviewed the output, fixed canonical project binding, empty/model result validation and accepted-artifact hashing, and recorded acceptance. Final 27 network/credential-isolated checks passed. A second empty-queue service run succeeded without a model launch. Timer enabled; source commits d390273 / 53c8344; first service peak memory 204.7 MB, no swap. Notebook status/results published through existing sync.

  • 2026-09-22 social pairing/artwork: constrained tool-free Claude CLI Fable design and verification passes resolved claude-fable-5-1, medium effort, no MCP/tools or recursive agents. CLI list-price estimates $0.80770275, $0.56207825, $0.45607825 (not subscription charges). Primary supplied full relevant code to resolve two diff-only false positives; worker withdrew them. Primary accepted the useful stale-retry/pre-transaction findings, implemented locally, and independently ran 158 offline backend checks plus eight browser scenarios. One foreground helper at a time; background runner had no active jobs. Start native reading 7% five-hour / 2% weekly / 0% Fable weekly at 16:45:48 UTC; Post-work native observation at 17:03:28 UTC reported 12% five-hour / 2% weekly / 2% Fable weekly, resetting at September 22 21:20 UTC and September 28 17:00 UTC respectively; the native collector may cache readings. Exact snapshot retained privately in social-pair-fix/capacity-final.json. No paid API fallback used.

  • 2026-09-22: Owner-requested unified Cow Remote review used exact claude-opus-5-5, high effort, subscription CLI, no tools/MCP/fallback, 389 seconds. Returned useful findings; primary reproduced three and confirmed blocking EEPROM code. Corrected model misclassification of inherited Play behaviors as new regressions and ruled out its tentative RF-power mismatch. Output exceeded requested 1800-word target. CLI list-price estimate $2.9286222, not extra subscription spending. Source snapshot unchanged. Native usage refresh remained cached/stale, so no fresh allowance claim. Report: cow-remote-unified-review.

  • 2026-09-22: Owner-requested focused incomplete-recording follow-up used exact claude-opus-5-5, high effort, subscription CLI, tool-free/no MCP/fallback, 98 seconds. CLI list-price estimate $0.3617662, not an extra subscription charge. Native observation was recent: 17% five-hour, 3% shared weekly, 2% Fable; no active background jobs. Useful explanation and storage layout math; primary added mandatory invalidation readback and corrected overly absolute CRC/torn-write claims and unmeasured timing. No firmware changes. See cow-remote-unified-review.

  • 2026-09-22: Cow application integration planning used tool-free claude-fable-5-1 medium,65seconds, CLI list-price estimate$1.35630275. Useful source mapping; primary corrected error token assumptions, silent-command blocking reads, capability/state reset and release versioning recommendations. Plan only; source unchanged. See cow-flasher-configurator-mode-plan.

  • 2026-09-22 Cow delivery: constrained tool-free claude-fable-5-1 draft at high effort ended by signal143 with no returned patch; no result/cost available, primary implemented instead. Two smaller medium-effort Fable5.1 reviews completed: firmware132seconds/$1.21567275 and Configurator133seconds/$1.34380275 CLI list-price estimates, not subscription charges. Primary reproduced/fixed useful load-cancel/indicator/test-coverage and diagnostic polling findings; rejected excerpt-induced storage-deadlock speculation and mode/RESET semantics speculation after checking full firmware. Release + host fault injection/scheduler tests + real-browser/mock-USB and catalog checks passed, all three feature branches pushed. Post-batch Claude snapshot was stale; no fresh consumption attribution claimed. No paid API fallback or hardware flash. Evidence ~/engineering/review-artifacts/cow-delivery/.

  • 2026-09-22 feed-only/seven-day buffer: subscription CLI design and independent code review used tool-free claude-fable-5-1, medium effort; list-price estimates $0.65552275 and $0.74005825, not subscription charges. Primary rejected destructive cleanup suggestions, retained legacy history, and verified 166 offline backend checks plus nine browser scenarios; later focused buffer/wrapper checks passed nine cases including guessed-tag rejection. Research is limited to WebSearch/WebFetch, structured output, no shell/files/MCP/recursive agents. First Sonnet draft was rejected for missing inline sources and unsupported engineering assertions ($0.4886542 estimate); changed route/brief rather than repeating it. Two Fable research drafts returned ($1.82931425/$2.100823 estimates; internal WebFetch also reports claude-haiku-4-5-20251001). Primary removed speculative market claims, unsafe footprint assumptions, invented timing and company promises, and corrected tags; user editorial approval remains required. A third bounded Fable draft is running. Native Claude usage at 2026-09-23 02:20:57 UTC: 25% five-hour (reset02:29:59UTC),5%sharedweekly/5%Fableweekly (reset2026-09-28 16:59:59UTC), sourceclaude-cli-usage,age8seconds atread. No paid overflow or API fallback.

  • 2026-09-22 buffer batch completion: third Fable research returned claude-fable-5-1 plus internal WebFetch claude-haiku-4-5-20251001, medium effort, list-price estimate$1.69926725 (not subscription charges). Primary revised the absolute firmware-recovery headline and removed universal safety/bootloader assertions and unsupported company-process claims; final1202-word draft uses four primary documentation links. All three outputs required editorial corrections: useful research scaffolds, not publication-ready without review. Supply7 verified; user approval remains pending. Future research remains bounded and approval-gated.

  • Buffer post-batch native observation at2026-09-23 02:29:34UTC (claude-cli-usage,32secondsold when read):29%five-hour,5%sharedweekly,6%Fableweekly. Session reset02:29:59UTC was passing at read time; this is a recorded pre-reset observation, not a claim of current post-reset headroom. Full snapshot privatefeed-only-buffer/capacity-final.json.

  • 2026-09-23 daily-queue correction: tool-free subscription claude-fable-5-1, medium effort, design review$0.63766275 and code review$0.87248825 CLI list-price estimates (not subscription charges), no MCP/tools/recursive delegation. Primary reproduced and fixed held/missed re-review, shared DB path, skipped release hooks, mutation guards and two-way date reservation; rejected review's assumed human-only approval rule because owner explicitly delegated routine approvals. Also rejected counting only already-published stages as future readiness: prepared inputs/execution acceptance and later delivery evidence are distinct. Final160 offline backend cases plus four375/1440browser cases pass. Runtime daily release remains disabled; no paid fallback.

  • Daily-queue post-review capacity observation: source claude-cli-usage, observed 2026-09-23T10:45:47.409230+00:00, age 423seconds at capture; shared five-hour 9.0%, shared weekly 6.0%. Snapshot retained privately in daily-content-queue/capacity-final.json; native readings may be cached and shared usage is not attributed solely to this task.

  • 2026-09-28 cricket flash-and-test tool (cricket_firmware_008/tools/flash_test.py, about 270 lines): one bounded job to a worker agent with the model set explicitly to Sonnet; the exact returned model version is not shown by the Agent tool. About 71k subagent tokens, about 1.5 minutes; no CLI cost estimate available. Accepted after primary corrections: the polling loop would have restarted the self-test (avrdude resets the chip on every disconnect), make without clean could flash a stale build, a leftover temp file, and a "dirty" flag counting untracked files. Logic tests and dryrun smoke test passed; not run against real hardware. Reorganisation and review stayed with the primary.

  • 2026-10-04: Codex (codex exec, workspace-write in ~/assistant-sandbox) did the phone-layout pass in 3 rounds (gate 130 → 10 → 2 failures fed back); Claude ran gates (Codex sandbox can't run docker; its local pytest stalls — Claude runs pytest) and released v2026.10.7. Gate now checks inner sideways scrollers, clipped content, stat-tile alignment, menu scroll+lock, settings panes, full marketing fixture (431 checks).

  • 2026-10-14 · Sonnet (worker, background) · rusefi-sti-2004: extend the Hellen package build to 6 copper layers (patch applied/reverted by hellen-create-board) · medium · done first try, verified by the worker (6 copper layers in the package, drill hits 1015+1014=2029 match, patch reverts byte-identical incl. on failure) and spot-checked by the primary (commit, apply/revert lines) · ~17 min, no correction except one guard added afterwards (revision.txt BOARD_LAYERS mismatch).

  • 2026-10-16 · Fable ×2 (Freerouting code map; KiCad 11 master study), Sonnet ×2 (KRT code map + trial on a rusefi copy; landscape survey), all read-only background agents, models set explicitly · kicad-autorouter research · hard · all first-pass accepted; primary verified the release assets and ran the native CLI (found it crashes, contradicting the survey's 'no Java needed') · about 170k/156k/146k/55k subagent tokens.
  • 2026-10-16: Hit the Claude 5-hour session limit. ccusage 10-08..16: $946 list-price equivalent, 97% cache re-reads from multi-day Opus 1M chats (four desktop chats opened 10-13 ran 2–3 days each); Codex nearly idle; API credits $47.76/$100. Set model: sonnet and autoCompactWindow: 200k in ~/.claude/settings.json (backup settings.json.bak-20261016) and wrote SOP-025 v3.3 "Three pools" routing. Codex probe OK.
  • 2026-10-16: Made the routing self-enforcing every cycle: ~/.local/bin/ai-budget (tests test_ai_budget.py) as SessionStart + UserPromptSubmit hooks in ~/.claude/settings.json (backup .bak-20261016b), CLAUDE_CODE_AUTO_COMPACT_WINDOW=200000, ai-budget watch after each ai-capacity refresh, ai-budget-weekly.timer Mon 14:10. First weekly reading (10-09..16): $923, Opus 78%, re-reads 97%, avg context 325k/call — the baseline the next Mondays are compared to. SOP-025 v3.4.

  • 2026-10-16: Headless claude -p jobs loaded every MCP server's tool schemas up front (craft apps etc.): 269k tokens before the first word, on every run. quote-drafts (26 runs/weekday) used ~1.3M tokens a run. New ~/.local/bin/claude-mcp-only SERVER... prints --disallowedTools for all other servers (measured 269k -> 49k, Gmail kept); jobs with no MCP use --strict-mcp-config (-> 18k): notes-daily, notes-weekly, note-to-social, job-fixer, reading-weekly. quote-drafts keeps Gmail, training-sync keeps Garmin + Calendar. Backups *.bak.20261016. New interactive chats since the hooks: Sonnet, ~110k average context, max ~130k (old Opus chats 200–480k average). Gotcha: --strict-mcp-config also drops claude.ai connectors (Gmail), and ENABLE_TOOL_SEARCH does not defer tools in -p.

  • 2026-10-18: KiCad execution-fix batch: four gpt-6.1-sol medium CLI helpers under enforced read-only sandbox (two patch-only implementers, two source reviewers); no recursive delegation or paid API use. Accepted with primary corrections: patch framing/native test initialization, derived-zone identity and checker false alarms. Reviews caught dangling symlink, escaped-name and compact-item-removal defects. Primary compiled, ran negative controls and validated native/real cases; outcomes in kicad-autorouter. Fresh shared Codex observation 3% weekly from codex-app-server (batch attribution unknown); Claude weekly 85%, session 13%. CLI cost not reported.

  • 2026-10-18: KiCad ground-repair/breakout batch:11explicit gpt-6.1-sol medium read-only CLI helpers, no recursive delegation or paid API use. Patch drafts and source audits accepted with primary corrections (native mask expansion, ground-calibration strictness, generated teardrop identities/layer aliases, exact endpoint guards and transient width probes). Independent reviews plus primary real reruns caught false completion from excluding teardrops and a via-session null-node crash; final54Python and60native feature controls plus7mechanisms/self-checks pass. Real closure still incomplete (USB4/CAN6/seven3opens); primary owns that outcome. Fresh codex-app-server observation at1792335031.6027598:8% weekly, reset1792927016, shared attribution unknown; CLI cost estimate unavailable. Claude capacity was not used; no reset credit redeemed. Worker briefs/results remain private in /tmp, durable proof in eeroute closure evidence and kicad-autorouter.

  • 2026-10-18: Native KiCad workflow batch:3explicit gpt-6.1-sol medium CLI helpers, enforced read-only, no delegation/Docker/build access. Two source audits (native session boundary; GND site search) accepted after primary source/report checks; narrow GUI tool review caught proof-bearing assertions under Python -O and unbounded cleanup. Primary fixed both and reran actual GUI7checks with -O;54Python and11ground controls PASS. New real replay still3opens with0new findings/regressions; no product-completion overclaim. Source audit jobs were broad (~80k/85k tokens) versus narrow review~13k; tighten file/tool-call scope for subsequent audits. CLI cost unavailable, subscription attribution unknown. Fresh codex-app-server observed1792336833.4975939:9%weekly, reset1792927016; no Claude/API/reset-credit use. Artifacts/commits in kicad-autorouter.

  • 2026-10-18: VectorCraft PCB-planning prototype:2explicit gpt-6.1-sol medium read-only CLI helpers (bounded command audit and independent checker review, no recursive delegation/builds). Primary verified installed0.13.1/native SVG output and caught physical unit loss; reviewer identified native-artwork proof gap, snapshot race and output preservation gaps. All corrected and checked on real H1a with130anchors/3fiducials/10mmbar plus native-only visual comparison,6Python controls and5negative cases. Source boards untouched. Helpers ~30k/~21k tokens; CLI cost not reported and shared subscription attribution unknown. Fresh codex-app-server1792342348.3551319:10%weekly, reset1792927016; no paid API/reset credits/Claude used. Commits/evidence in kicad-autorouter.

  • 2026-10-19: ai-budget weekly (ccusage, last 7 days): $1247 list-price equivalent, $178/day; Opus 55%, Codex 22%; cache re-reads 97% of tokens; average context 245k per Claude call (23582 calls). Baseline 10-08..16: Opus 80%, re-reads 97%, ~480k in long chats.
  • 2026-10-26: ai-budget weekly (ccusage, last 7 days): $506 list-price equivalent, $72/day; Opus 52%, Codex 8%; cache re-reads 95% of tokens; average context 203k per Claude call (20628 calls). Baseline 10-08..16: Opus 80%, re-reads 97%, ~480k in long chats.
  • 2026-11-02: ai-budget weekly (ccusage, last 7 days): $588 list-price equivalent, $84/day; Opus 68%, Codex 3%; cache re-reads 96% of tokens; average context 148k per Claude call (19927 calls). Baseline 10-08..16: Opus 80%, re-reads 97%, ~480k in long chats.
  • 2026-11-09: ai-budget weekly (ccusage, last 7 days): $252 list-price equivalent, $36/day; Opus 56%, Codex 14%; cache re-reads 96% of tokens; average context 209k per Claude call (15484 calls). Baseline 10-08..16: Opus 80%, re-reads 97%, ~480k in long chats.