Local-agent contribution ledger#

The exact-context accounting implementation on 2026-10-13 used direct Codex work as explicitly required by the owner. No local delegation or inference occurred. The new counter and tests are Codex-authored; genuine tokenizer/handler parity is model-free evidence, not a local-agent implementation contribution. Existing local-first policy remains unchanged outside this task.

The initial entries below use the existing direct Bridge with GPT-OSS 20B, thinking low, against base cc88d4d (preserved local v0.5 baseline). Only one inference ran at a time. The coordinating source was a separately hashed frozen snapshot, imported by explicit path under isolated Python. Agents returned proposals; supervising Codex applied and tested corrections in the separate development worktree.

Private raw request/result/usage evidence remains outside the public source tree under the local companion-release development evidence directory. This ledger contains no personal paths or session identifiers. Model/digest and actual token counters are retained in that evidence, not inferred from characters.

The correction harness initially refused to start a second owned server when an unrelated Ollama application appeared. No process was killed. After observing its inventory and empty active-model list, the correction used that shared server without terminating or unloading it. This was a runtime-admission failure, not a rejected model answer.

These are genuine code, test and documentation contributions, with substantial supervising correction. No invented local-development percentage or general coding-quality claim is made. The local model's high confidence did not guarantee valid output. New deterministic controller tests and overall architecture remain supervising Codex work.

Genuine candidate worker run#

A subsequent GPT-OSS 20B job used the frozen candidate runtime at base 8b02eba. Its launcher/controller exited promptly; the independent worker completed, exited, and atomically saved result + pending review event. Actual reported usage: 853 input tokens, 129 output tokens, 92.35 seconds bridge duration, 40.24 seconds model load and 12.30 seconds provider generation. These durations have different scopes; they are not invented additive timing categories.

The answer claimed any(d.get(k) ...) accepts falsy values, which is false, and recommended defaulting absent normalized remote metadata to false. Supervising review rejected it: the parser always supplies remote, and silently permitting missing safety metadata is not the contract. The falsy/missing-key helper regressions pass. The outcome is recorded in isolated usage telemetry.

The callback packet is prepared, but automatic approval review blocked sending repository-derived content/results and local paths to authenticated Codex without specific user approval. No real Codex session/callback was launched. The job is locally succeeded; its run is waiting for review, not finalized or accepted.

Feature-preservation follow-up: execution machinery disclosed#

The initial code/test/documentation proposals and remote-helper correction were direct Bridge.analyze calls. They did not exercise WorkflowEngine planning, role stages, shared workflow budgets or checkpoints. The initial independent candidate worker was also a direct Bridge job. Those contributions must not be described as multi-stage workflow development.

This follow-up used the actual frozen pre-companion v0.5 Service.workflows engine for an explicit review_change design review. It ran repository-explorer (repository-discovery, with no configured repository search; context-summary) and code-reviewer (code-review) sequentially, carried the first handoffs into the second stage, skipped non-applicable risk stages, and stopped at the explicit supervisor decision. Both model calls used GPT-OSS 20B, thinking low. Actual usage: 1,861 input / 256 output tokens, two attempts, 33,968 reserved tokens, 132.293 seconds charged to the workflow. Reservations are not actual token usage.

The review proposed tests for persisted configuration/policy identity, handoff rehydration and checkpoint enforcement. These overlap the supervisor's planned coverage and were incorporated/confirmed in the integration regressions. Its suggestion to restart "mid-step" was corrected: completed handoffs survive a restart; a crash before handoff persistence cannot resume generation. No executable patch came from this review. Adapter implementation and exhaustive role tests remain supervising Codex work. Outcome: corrected, useful design/test review.

A separately frozen candidate worker then exercised a real durable workflow task in delegation.mode=direct. This is the workflow engine's single-stage path, with shared budget/profile/checkpoint tracking, not a raw Bridge call. GPT-OSS reported 633 input / 64 output tokens, 62.31 seconds bridge duration. Its warnings were hypothetical risks rather than demonstrated defects; they were treated as test considerations, not accepted bug findings. It stopped before supervisor acceptance; no authenticated callback was sent.

The same independent worker subsequently used existing Service.second_opinion(consensus=True), not WorkflowEngine, for fixed sequential GPT-OSS coding and exact qwen3:14b-q4_K_M reasoning opinions on the preservation contract. Actual usage was GPT-OSS 508 input / 60 output, 53.52 seconds; Qwen 363 input / 165 output, 78.88 seconds. Results disagreed: GPT-OSS supported the contract; Qwen was uncertain and requested clearer definitions of exact overrides, resource gates and disagreement ownership. Some uncertainty repeated information already supplied. The supervisor clarified the user guide and preserved the disagreement; no model vote accepted a checkpoint. Outcomes: GPT-OSS accepted as a bounded opinion; Qwen corrected as documentation feedback. This is transport and independent-opinion evidence, not broad Qwen coding-quality certification.

All these calls were sequential. Their task/result/digest/provider usage evidence is retained in the private preservation-contribution and preservation-live directories. No model download, deletion, authenticated callback or publication occurred. See preservation matrix and workflow integration for the resulting behavior.

Focused recovery repair#

Two bounded GPT-OSS 20B contributions were used, sequentially, without a model benchmark. A direct existing Bridge call reviewed the proposed reconciliation protocol. Its suggestions supported payload-equality, partial-step and copy-before-ack tests; some uncertainties repeated guards already specified, and physical database corruption is outside the tested process-crash model. Outcome: corrected design/test feedback. Codex implemented and independently tested the repair.

The repaired installed wheel then ran a real review_change task through the companion and existing WorkflowEngine's direct path. Exact GPT-OSS routing used 1,595 input and 72 output tokens in one attempt. It reviewed the actual receipt delivery method and competing-recovery test. Its only proposed defect was an incorrect claim about equivalent digest comparisons; the model finding was rejected. Supervising Codex supplied the corrected analysis and validation. The local-completion snapshot retained a pending review and no final approval. Supervisor checkpoint, recorded review and final validation were separate steps.

Private evidence records the exact task, model digest, installed runtime hashes, payload, provider usage, supervisor feedback and state snapshots. Local agents provided bounded opinions, not authority to alter records or approve recovery. No authenticated callback, model download or model deletion was performed.

Selective execution development contribution (2026-10-08)#

The unchanged companion-development Bridge at base 0b817c5 supplied one real adaptive advisory on a bounded regression-test plan. Exact model: qwen3:4b, digest 359d7dd4bcdab3d86b87d73ac27966f4dbb9f5efdfcc75d34a8764a09474fae7, Q4_K_M. It completed in 15.40 seconds and suggested concurrent duplicate admission, cross-lane dependencies and partial-recovery checks. Codex used those as test prompts, wrote/reviewed the implementation and tests, and rejected its generic uncertainties about requirements that were already explicit. No local patch was applied automatically. This was advisory input, not independent acceptance.

A new generic-chat advisory smoke then failed completion validation after 41.82 seconds. The output-budget behavior exposed a mismatch with the proven native Ollama path. The implementation was corrected to reuse Bridge's structured schema, skill hook, exact guard, thinking setting and accounting. A real corrected smoke reported 1,011 input / 145 output tokens, 7.263 seconds model loading, 17.756 seconds provider total and 20.322 seconds controller elapsed. Its factual missing-model finding was checked against selection.py; its unrelated uncertainty was excluded.

The smoke demonstrated finished execution with a distinct awaiting_review state before the supervising disposition. Detailed private evidence is retained under workspace data/selective-local-contribution, data/selective-live-advisory and data/selective-live-native. These runs are not a matched pilot, quality certificate or evidence of cloud-token savings. No hosted routing, authenticated callback, model download/deletion or GPU benchmark was used.

RTX qualification contribution (2026-10-08)#

The final installed candidate ran the same exact qwen3:4b digest through MCP and an independent controller/worker on RTX 3060, 16,384 context, code-review@1.0. It proposed cloud-suffix and remote_host checks. Codex inspected selection.py and existing tests, supplied missing exact-selection coverage and discarded speculative and incomplete findings. Outcome: corrected advisory handoff; no patch integrated. Usage was 1,105 input / 275 output tokens, 6.944 seconds loading / 11.031 seconds provider total. These are smoke observations, not comparable speed or quality claims.

Actual dock data, completion-before-review, owned cancellation and restart evidence is retained under data/selective-egpu-qualification-20261008. See live qualification. No OpenHands/Strata agent contribution, cloud savings or release readiness is claimed.

The subsequent blocker-resolution pass used direct Codex diagnostics and review. It produced no new local-model contribution and did not run the SDK agent. Earlier genuine advisory evidence remains intact. The qualified qwen3:4b tag is now absent; no replacement tag inherits its result.

Installed-model screen contribution (2026-10-09)#

The same fixed candidate at base 0b817c5 used its installed Bridge transport, structured parser and accounting for one identical advisory each from GPT-OSS20, Qwen3.6-35B, Gemma4-26B and Qwen3.8-27B. Exact digests, settings, outputs and supervisor dispositions are retained. Qwen3.8's findings, minimal fixes and proposed tests passed independent fixture review with no correction. GPT found both bugs but needed two editorial correction groups; Qwen3.6 and Gemma exhausted their declared output allowance without a final advisory. Actual synthetic tool exchanges were checked separately.

Codex authored the fixture/oracle/harness, reviewed outputs, investigated the sandbox and prepared documentation. No model patch was integrated and no SDK agent executed. GPT's automatic local compatibility migration changed its current tag after its tested resident exited, so that new binding inherits no result. The report also discloses Ollama's two-blob automatic pruning side effect. This preliminary direct Bridge assistance is not a durable worker, human UI, implementation or matched-pilot qualification, and implies no cloud-token savings.

Bounded 4b-instruct qualification (2026-10-09)#

Codex performed authorized direct bootstrap: command restrictions, native SDK wire conversion, frozen correction baselines and failed-output collection. The local 4b-instruct model made nine real SDK inference calls and produced incorrect fixture edits. Protected tests failed; no suggestion/patch was accepted or rescued by Codex. The useful timestamp task was frozen and its failing baseline verified, but was not dispatched past the failed qualification gate. See the bounded report. This failed attempt does not establish general model quality, savings or release readiness.

Stronger-model bounded qualification, 2026-10-09#

The stronger-model report records genuine LAB-to-OpenHands-to-Ollama delegation: exact qwen3.8:27b generated the UTC timestamp change, ran the protected test, and received independent Codex acceptance without candidate repair. The patch remains isolated and uncommitted. 35B's first fixture failed tool policy; its second generated correct retained bytes but failed normal collection. Codex's context/budget/qualification/evidence bootstrap edits are separate direct work, not local-model contributions. No savings/equivalence claim.

Qwen27 standard model milestone#

Qualified qwen3.8:27b on Ollama 0.40.2 used the retained OpenHands 1.53.0 networkless restricted mailbox. Immutable runtime snapshots kept in-flight jobs independent of source edits. All model calls retained the accepted 65,536 context, 28 GPU layers, draft 4, 1,536 maximum output and thinking-off profile.

Security controls, default/profile bindings, resource guard, accounting, package integration and independent verification are Codex work because they establish or exercise trusted authority. The small Codex-only test comparator is evaluation investment, not a local contribution. No delegated task is counted as accepted until its independent review and protected checks pass.

Qwen27 close-out contribution#

The existing installed controller ran one bounded accounting-documentation advisory with the same accepted Qwen27 profile. Codex reviewed it against the implementation and integrated the paragraph in SELECTIVE-PILOT.md unchanged: accepted without Codex repair, one attempt, 291 input / 130 output tokens; cache usage unknown. The actual receipt again records the qualified digest, runtime flags and positive GPU allocation. It is implementation investment, not a savings measurement. Security/accounting code, acceptance and Git checkpoints remained trusted Codex work. The prior contributions and all failed/retried attempts remain preserved.

Qwen27 publication documentation contribution#

On 2026-10-10, one bounded native advisory at source base 71791b3 drafted the SDK/mailbox transport sequence and trust explanation from selected source context. Actual provider usage: 621 input / 244 output tokens, cached input unknown; one inference attempt, no local retry. Loaded digest, context and GPU/runtime flags matched the qualified profile. Two oversized context submissions were rejected before job creation/inference; context was narrowed without raising limits.

Accepted after Codex repair: the local draft omitted tool observations and made mailbox file access look like a direct relay-to-host call. Codex checked runner, sandbox and selective execution source, corrected those details, added the tool loop and independent verification, and integrated the result in the workflow guide. Trusted review recorded the original candidate hash and correction rationale, then finalized the run. This was advisory documentation work, not a fresh OpenHands runtime test. Codex performed architectural verification, accounting, publication safety and GitHub operations; those authority-sensitive tasks were not delegated. All usage is documentation/publication investment, not measured operating savings.

Strata upgrade advisory — 2026-10-12#

One bounded test-suggestion job used the unchanged close-out installation, source base a9a4bad, exact Ollama 0.40.2 / Qwen27 digest aaee06c39dcf..., 65,536 context, 1,536 output ceiling, 28 GPU layers and draft 4. The existing LAB durable worker held its serial reservation in a separate data directory. Usage was 258 input / 318 output, cached input unavailable; one inference, no retry. The original model answer is retained privately.

Accepted after Codex correction: identity-change, unmatched-result, image and uncertain-disconnect suggestions matched the source and regression cases. The selector suggestion incorrectly implied every explicit Strata selection needs a prefix; contracts can use separate backend/model fields. Codex corrected that distinction, recorded hash-bound review coverage and finalized the run. No model-authored code patch was integrated. Architecture, identity/security implementation and final review were direct Codex work because those decisions are supervisor-owned. This advisory is not Strata or OpenHands qualification.

Exact run/job/receipt IDs are retained in private advisory-dispatch.json and advisory-review.json, correlated by task key strata-test-suggestions. Model total 220.592 s, including 57.899 s loading; durable worker interval 225.162 s. Samples observed minimum available host RAM 7,723,597,824 bytes and peak total GPU use 10,277 MiB, within the unchanged ceilings. Reservation released; cached and cancelled usage are not invented. This is implementation overhead, with no savings comparison.

Approved Strata preparation — 2026-10-13#

The isolated LAB runtime was the reviewed e5c6337 wheel; the active a9a4bad Ollama installation was preserved. Codex handled pinned installation, resource ownership, artifact identity, failure diagnosis and independent acceptance, which require supervisor authority. The intended bounded LAB/Strata advisory was blocked before dispatch: main expert packing hit the unchanged RAM guard. No local inference, OpenHands implementation or model-authored change occurred in this follow-up. No alternate backend was used to turn that failure into a pass. The current report records exact assets, measurements, cleanup and the scoped retry proposal. Earlier Ollama advisory/implementation contributions remain historical evidence.

Direct-GGUF amendment — 2026-10-13#

The approved single preparation and independent artifact validation passed from 7cbaa11 with the unchanged reviewed e5c6337 runtime. Codex owned source/security, resource guards, identity checks and final review. The pinned Strata server's internal replay/restart policy blocked startup under the zero-retry requirement, so the intended stage-B advisory and stage-C implementation were not dispatched. No model-authored changes or Codex repairs to a model candidate exist in this amendment. Documentation/evidence updates are direct Codex work; another backend was not substituted. See the outcome.

Patched Strata advisory — 2026-10-13#

Codex directly owned serving/security changes, resource admission and independent acceptance. After A passed, one real MCP/durable LAB advisory used the exact Strata 0.1.43 engine with LAB-specific serving patch. It suggested a test proving that a failed response cannot dispatch a partial tool call. Codex checked the suggestion against the adapter and passing error/partial-tool tests and accepted it without correction. It is useful bounded test advice, not model-authored code. The already written patch and tests remain Codex-authored; no OpenHands candidate or Codex repair to a local candidate exists.

The job/attempt, binding, sole added server request and reported counters are correlated in private evidence: 122 input / 34 output / 0 cached tokens, 66.107 s including MCP submission, 56.801 s worker/admission interval. These intervals overlap. The later RAM stop blocked C and further live safety checks; no alternate provider, automatic restart or cloud inference was substituted. See the measured outcome. No token-savings claim is made.

Final serving revision with mapped-page release — 2026-10-13#

Fresh real MCP → durable LAB advisory used final patch 3a0ca878… and the single isolated STRATA_FILE_RELEASE=1 mitigation. The same bounded test suggestion was independently accepted without correction: 122 input / 34 output / 0 cached tokens, 71.508 s including MCP submission, 61.780 s worker interval and 43.8 s provider generation. These intervals overlap. Exact binding, request increment, usage and candidate hash are in the trial evidence. Codex owned source/binary/security review, resource admission, monitoring and acceptance; no OpenHands implementation or model-authored code was produced. C lacked memory headroom; competing inference subsequently triggered the GPU guard. Deliberate live failure/cancellation checks remained unrun. No retry, alternate provider, cloud fallback or token-savings claim was made. See the current outcome.

Readiness-only audit — 2026-10-13#

The user explicitly required direct Codex work to preserve observed resource conditions. No local delegation, model request or startup occurred. Codex traced provider/store metadata, reconstructed the SDK envelope offline and reviewed the qualification-only observation gate/tests. Prior final-patch A/B acceptance remains separate; no new local-agent contribution, token saving or implementation success is claimed. See the single readiness plan.