[loopx-agent]

Current author checkpoint — cfaf2e4

Complete-source operation Host preparation now uses the existing checked dispatch owner for normalized session identity. Codex previously initialized its native process/session before the complete Goal-delivery check in prompt preparation. It now checks complete work and Goal delivery first. The fresh prompt/schema check after startup remains, so a source changed during initialization still refuses before send. The preceding material-result check after the provider also remains. This is a direct Host boundary fix, not a claim that the outer executor previously bypassed admission.

DSH previously read dispatch context and then reread complete-source identity in the same preparation invocation. It now reuses that invocation's checked identity. TypeScript retains the existing normalization owner; an explicit boolean option adds session lineage only for these two callers. Omitted/false options retain the old projection. This metadata does not enter model input or persisted authority. Historical requests, scope/path/lifetime checks, confirmation/consumption and result validation keep their existing owners. No new capability, setting, permission, decision owner or cache is introduced.

Matched component costs use eight actual synthetic Host preparation captures, then replay the exact saved requests through the real frozen preimage. Normalized lineage and standalone prepared prompt/schema are identical. Counting is canonical UTF-8 method/request JSON plus returned JSON after numeric transport metadata:

Host / request	File: old → new	SQLite: old → new	Counted preparation RPCs / byte reduction
DSH / internal complete	85,380 → 42,950 B	85,494 → 43,007 B	2 → 1; 49.6955% / 49.6959%
DSH / complete wire v1	58,910 → 29,715 B	58,966 → 29,743 B	2 → 1; 49.5586% / 49.5591%
Codex / internal complete	42,576 → 42,950 B	42,633 → 43,007 B	1 → 1; −0.8784% / −0.8773%
Codex / complete wire v1	29,341 → 29,715 B	29,369 → 29,743 B	1 → 1; −1.2747% / −1.2735%

Reduction is (old−new)/old. Each Codex case adds 374 B, paying for the earlier complete work/Goal check and normalized metadata. DSH removes the adjacent redundant identity RPC. This component excludes Codex's retained post-start preparation, post-provider checks, other reads, session/SDK work, framing, CPU/latency, model tokens/cache/invoice and settlement. The two Host samples are not combined into production net benefit.

Validation and remaining limits:

• Integrated target main cbc1eeabd740ac557c3a4fb7bf67050ad063b629; immutable product preimage 865970ff3f06966dd7bb0976c80fd8137031434c. The check resolved target main ac74f90c4778381b6fcedd5d76059522591c3f42; its 27 subsequent anti-entropy paths do not overlap these six production paths. This checkpoint does not claim the inherited chat/UI work as new visual delivery.
• Frozen first 28-case oracle: initially 16 failed / 12 passed, including four new-test diagnostic assertions that overlooked DSH's exception cause. Corrected observation on the unchanged product gives 12 failed / 16 passed: eight old preparation call-shape/count observations and four Codex pending-Goal cases that reached initialization. Bad-source, DSH pending-Goal and source-changed-during-startup controls pass. Candidate first 28 cases pass. The final new module passes 36.
• The first actual SDK run was 32 passed / four failed because the new loopback fixture assumed Messages while pinned 0.1.5 uses Chat Completions. A diagnostic failure confirmed this. The existing dual-protocol stream fixture was reused, with eight byte-equivalence checks for the extraction. One subsequent test still failed on an incorrect context-key assertion after native context/result had succeeded; the assertion was corrected to the existing root identity contract. No production change was made for these fixture errors.
• First 11-complete-module run: 979 passed / two failed / 22 skipped. The two capture observers expected the original source during an intentionally changed-source negative attempt; the new early projection correctly forwarded the changed source. The observer now tracks that attempt's expected source, preserving refusal, repeated attempts, original prompt/schema and unchanged-output checks. Final same 11 complete modules: 981 passed / zero failed / 22 skipped (444.048 s). Earlier failures retain their own results.
• Actual loopback SDK qualification uses SDK/runtime-bin 0.1.5rc1, placeholder credentials, and the two existing native-operation/context TypeScript modules compiled unchanged for the test. Four complete-source context/result cases cover File/SQLite × internal/wire; four existing retained-session native cases also pass. Twenty tests requiring an explicit resume-capable runtime and two live-model opt-in tests remain skipped. This does not qualify 0.2.1, new-connection resume, every lifetime, a full plugin install or a real model.
• The full TypeScript owner module passes 58, with no skips; typecheck, changed-Python Ruff, paired staged/committed RFC checks and final whitespace checks pass. The limited diff-scoped semantic advisory finds no supported new production carrier; it is not a semantic proof. The same-domain refactor reuses exact normalization and the existing dual-provider test stream. The 12 final candidate paths pass the public-boundary scan.
• Chat package assets and the actual final Core wheel build. A fresh isolated wheel installation passes four complete modules: 219 passed / zero failed / 20 explicitly skipped (252.479 s). All 1,783 packaged Core Python/TypeScript files match final production source; all 1,178 loaded Core modules originate inside the installation. Only test fixtures come from checkout. No user runtime is installed or changed.
• Exact-head Goal-scoped standard premerge: 7 direct checks passed; 19 selected checks executed, 18 passed / one maintainability-ratchet failure (two existing unreviewed oversized functions: quota._emit_quota_result and todos.machine_section_projection.render_canonical_todo_sections; zero magnitude regressions). The exact-scope quality receipt is missing, so the overall gate fails. No current passing independent review, merge-readiness or merge authority is claimed.

The preceding result-normalizer checkpoint at 849a5ee5d4da4712bb60090be4a6163d27d4b105 retains its separate qualification: malformed complete sources refuse, complete-wire material results retain source evidence obligations, historical/stop behavior and the outer content validator keep their rules. Its frozen oracle was 28 failed / 56 passed, final new module 84 passed, and final source/wheel two-module runs each 176 passed. Eight identical provider results add eight post-provider projection RPCs / 288,686 B; that is an added correctness cost, not compression. The prior session-lookup 339,376 → 232,768 B / 31.4129%, earlier external-Host wire measurement and same-capture body metric keep their own denominators. These costs and benefits are not offset or combined; whole-C1 net benefit stays unknown, not zero.

Whole C1 still needs remaining actual producer/caller/reader and private executor/journal adapter cutover, plus original-quota single-session necessary-information, subsequent-decision/work, accepted-outcome and complete cold/warm/lost-context cost qualification. Dual-session qualification stays separate. That concrete unfinished author work keeps the aggregate proposal Draft; review/CI waiting alone does not. No installed Host, PostgreSQL, real-model adoption or long-horizon benefit is inferred. No live installation, release, default promotion, scientific job or merge was performed.

Prior external Host wire checkpoint — 3e7a96c

External generic CLI stdin previously carried both the complete work_contract and the executable v0 envelope. Complete-source Host request v1 now delivers one admitted source, original capture readers, Turn key and result transport. Native model prompts use the same format. The shared checked-source owner rejects missing/corrupt sources, unsupported versions, mixed carriers and capture drift. Native lineage uses that owner; private plan/result/completion readers retain their existing schema through a thin transport adaptation. The versioned internal executor adapter and historical v0 decoder remain for recovery.

The shipped DSH stdio, Ark and MemT distribution consumers use the shared Core version decoder. Ark permissions/receipt recovery and MemT exact-lifetime gates retain their existing owners. The optional packages explicitly require the matching integrated Core; the version floor alone is insufficient. This explicitly changes the complete-source external wire: v0 external adapters must consume v1 work_contract before accepting new Turns. It does not add configuration or an independent capability. The affected journey is existing Turn execution, refusal, artifact validation and same-Turn recovery. Remaining internal-reader retirement and the original-quota qualification keep whole C1 unfinished.

Validation remains source-bound:

• Core runtime 2a7080749 passes 840 tests / 29 skips across 13 complete Python modules. The skips are 27 unavailable deepseek_harness cases and two explicit live-model release cases; none is a pass or installed-service qualification.
• After integrating pinned target main 9d47a36eec0c05b54edcff8323ee81853814457e, three complete source/capture/journal modules pass 108. Five remaining completion/claim/context/inbox/lease modules pass 82 after the test reads its Todo and verification witness from the complete contract. The earlier two runs each retain 122 passes / one obsolete-wire-access failure; these are separate invocations, not one all-green suite. The final commit changes only that test; product bytes match 009c22915b63e38ff0bbbcc395115c3f8979fc40.
• Exact optional-adapter preimage tests retain 12 Ark failures / 12 controls and 10 MemT failures / 18 controls, demonstrating two hardcoded v0 gates. Source qualification then passes the complete Ark suite (58) and the selected MemT source/lifetime/generation protocol suite (33, with 26 deselected). The MemT result converter already supported the new source; no speculative Core execution-context branch was added.
• Three local wheels build after regenerating the ignored Chat bundle through the existing build. An isolated Python 3.12 environment installs those actual wheels and verifies every loaded Core/provider production module came from them. Real File/SQLite CLI artifact/replay passes 2, full Ark 58, and selected MemT protocol 33 / 26 deselected. Ark uses its actual SDK with local transport/MCP fixtures. The separately licensed MemT native client, model/inference, full memory and natural outcomes remain unqualified; no live installation or service was changed.
• Relevant TypeScript source tests retain their 149 + 15 distinct checks, and current integrated control-plane typecheck passes. Ruff, paired RFC/whitespace and public scans pass for this increment. Historical v0, complete-source corruption/missing/mixed/future-schema refusal, exact original identity and once-only recovery retain their oracles.

Matched synthetic File/SQLite journeys compare immutable 0bbc76a8c72586d0f3970013de0d8ddd9453e4b1 with 2a7080749cd2e24cebe187dced29b63fdd84f75d. A real generic subprocess consumes the complete task tail, writes an independently validated artifact, commits once and replays without another Host or debit.

Backend	Complete source bytes, unchanged	External stdin, before → after	Bytes saved / reduction
File	40,664	66,212 → 42,516	23,696 / 35.7881%
SQLite	40,692	66,268 → 42,544	23,724 / 35.8001%

Same-capture reconstruction predicts exactly the observed wire: remove 25,034 / 25,062 bytes of adapter/JSON field, add 1,338 bytes of original-reader header. This byte calculation is separate from the semantic oracles. Parent RPC counts/bytes and CLI stdout are unchanged in both arms: first execution is 275 / 2,772,083 bytes (File) or 272 / 2,773,234 (SQLite); replay is five RPCs / 213,237 or 213,405 bytes. CLI stdout is 3,443 for execution and 3,989 for replay; Host results are 722 bytes. These dimensions are not summed. CPU/IO, model tokens/fees and natural outcomes remain unmeasured; this is a nonzero wire saving, not whole-workflow net-benefit qualification or a new same-capture selected-body statistic.

The same actual artifact/replay observation on integrated product 009c22915 reproduces every tabled source/stdin/result/RPC/stdout value. It is an integrated readback, separate from the causal 0bbc76a8 → 2a708074 component comparison. The final test-only commit does not change those product blobs.

The standard local premerge run at 009c22915 passes seven direct checks and 18 of 19 selected checks. Its maintainability ratchet fails on two unreviewed oversized-function findings: quota._emit_quota_result and todos.machine_section_projection.render_canonical_todo_sections. These holds are preserved; this increment does not qualify their repair. The final exact-head Goal-scoped quick run passes seven direct and four selected runtime checks, but its overall result remains quality_receipt_missing. A separate unscoped quick run passes its public checks; that result does not replace the Goal policy. No passing current-scope change-quality receipt, independent review, exact-head readiness or merge authority is claimed.

No installed rollout, independent review, merge readiness or merge authority is claimed. Historical size/compatibility failures and the earlier source-bound observations below remain open at their recorded revisions. This PR stays Draft for the concrete remaining writer/reader and original-quota single-session outcome/full-cost work.

Prior journal checkpoint — 0a74529

Journal inspection, publication, settlement lookup and managed retry previously checked the historical executor adapter while ignoring a supplied complete work_contract. Changed source facts could therefore remain classified as a consistent journal. These callers now use the same TypeScript checked-source owner as Host dispatch before permitting recovery or publication. Unsupported/null contracts, missing bindings, changed numeric facts and internally re-signed but divergent adapters refuse. Historical journals without a complete contract retain their existing decoder and settlement identity.

The native journal number parser moves to the existing IO decoder and is shared by publication and readback. Python inspection and managed retry transport the original sparse numeric spellings. This is a source-integrity repair, with additional validation work; it adds no parallel admission, scoring or settlement owner. Existing CLI inspection/recovery is the affected journey. No new configuration or frontend surface is introduced.

This checkpoint integrates target main ec5c9702c79cbc68ab994b18031e9069b0935b59. Its Stride nested authoring example and resumed-Todo continuation are retained; the conflict keeps C1's source-bound authoring reader. The final commit only adapts tests to the current digest-bound native reader and explicitly selects full authority delivery for the historical material-codec case. Its native on/off matrix, full-source equality, original identity, hostile-PATH, refusal and no-new-guard assertions remain.

Validation is versioned and separate:

• The six new File/SQLite cases all fail on immutable pre-repair 75bc0e7617e8e0d74b92b845b861186c7aab6f4b; they pass on the candidate, including journal refusal, numeric source recovery, historical absence and managed retry. Eight invalid-source conditions run on each backend.
• At runtime increment d2c788b805b1759835c7ce937a8074e5f7f748a6, nine full Python modules pass 543 tests after an earlier disk-exhaustion run was retained as incomplete. The separate journal/managed/controller set passes 154.
• After the target integration, seven modules report 354 passes / 8 failures. All eight failures are in the staged-reader module's obsolete response access or implicit presentation mode. Its corrected full module passes 23 tests. These runs are not one combined all-green suite.
• Six relevant TypeScript modules pass 164 tests on the integrated runtime; typecheck, Ruff, staged/committed bilingual checks and whitespace pass. Public scanning covers the current proposal; the existing synthetic Authorization assertion remains a disclosed warning. The final commit changes only the tested reader fixture, so product blobs match 839db145f116b34fa7249438dfe2ad86d49499cc.

A matched synthetic artifact flow compares 75bc0e7 with both d2c788b8 and integrated runtime 839db145. A real generic CLI subprocess reads the complete task tail, writes the independently checked artifact, commits one writeback/spend and replays the same Turn without another Host. Parent-process request/response lines include framing and authentication overhead:

Backend	First execution RPCs / wire bytes, both snapshots	Same-Turn replay RPCs	Replay wire bytes, baseline → repaired
File	275 / 2,773,573	5 → 5	213,299 → 213,375
SQLite	272 / 2,774,724	5 → 5	213,467 → 213,543

CLI stdout remains 3,443 bytes for execution and 3,989 for replay. The replay adds 76 bytes of numeric transport metadata; validation CPU and IO are unmeasured. This is a journal-repair comparison, not the original-quota model/outcome comparison, a new same-capture body-reduction statistic, an invoice estimate or proof of net efficiency. The earlier original-quota reading/cost observations and all frozen size/compatibility failures remain source-bound.

Quick premerge executes seven direct and four selected checks without runtime check failures; the overall gate remains quality_receipt_missing. No passing CQR, independent review, merge readiness or merge authority is claimed. No new wheel, installed App/native-model adoption or default rollout is qualified at this head. Whole-C1 writer/reader retirement, necessary-information and accepted-outcome parity, original-quota cold/warm/lost-context/full-cost comparison remain unfinished, so this PR stays Draft.

Earlier scope and source-bound evidence

Problem and resulting behavior

Quota, native Turn and heartbeat callers composed overlapping executable inputs. This C1 candidate composes one complete source through the existing TypeScript owner before capture, validates its source/effect identity and carries the original obligations to the final host input. Historical journals keep their recorded decoder, bytes, Turn key and once-only settlement.

The preceding runtime increment removes native-child observation metadata’s dependency on the compatibility envelope. Request construction, durable Host attempts and both Codex transports read the original complete interaction context when supplied; historical requests keep their old decoder. This is source convergence: the existing verifier already rejects inconsistent projections, and no prior wrong-child admission or packet-cost improvement is claimed. The preceding DSH increment fixes adaptive tool/session bindings through the same checked selected-work owner.

The preceding runtime increment removes one redundant cross-process version check during complete-contract planning. The existing typed driver_lineage owner still checks version, source integrity, projection and work identity. The same fixed File/SQLite inputs use one fewer RPC and smaller argument/result JSON, with byte-identical final plans, Host requests, prompts and schemas. Historical inputs retain their original version diagnostics. The earlier profile-alias increment retains explicit off/native on and once-only admission/no-spend qualification.

The earlier Host-transfer repair retains supplied non-object work_contract values at request construction and original-capture publication. They now reach the existing TypeScript refusal owner instead of being interpreted as historical absence. Genuine historical omission remains compatible. This adds no validator, field, default or RPC. The ordinary executor's earlier source checks still apply; the evidence does not establish an outer settlement bypass.

Managed Session validation now uses the same complete-source owner. Preparation reuses dispatch's invocation-local checked work identity. After model work, the callback checks the current complete source before running its original independent argv; supplied invalid input cannot become historical absence. Canonical declaration checks before/after argv, actor admission, lease and once-only settlement retain their owners. This adds a post-model complete-source RPC, whose measured component cost is disclosed below.

Current source integration: 99077996a77a02bae1183574e3e23f90627dab47, target-main 107e9e32e8f3df29f63b94524250f4044cc01a80. The same generic Host dispatch check now runs on the original-capture input before automatic-start reservation, publication or new journal writes. Pending-Goal refusal leaves the lane eligible and subsequent valid work can start; historical and saved-result recovery routes remain. Four complete capture, Turn-driver, manager-context and cadence modules pass 357 cases, including 22 strengthened dispatch/recovery cases. 2 actual native exec/CodeMode cases pass with a deterministic loopback provider. Earlier adapter and TypeScript results below retain their own snapshots.

Current qualification limits: the retained full CLI-budget observation is 4 failing groups / 39 passing / 1 environment skip, with all five original holds still open. Exact-head quick premerge executes 7 direct and 4 selected checks without check failures, but its overall result is failed: quality_receipt_missing. Current 132-file scope 9e594a84e97ab18c54bfb776801aaaf83d59b20eedfb437005a0e53fccb47511 prepares as review_required, without a passing current quality receipt or merge readiness. This Draft remains unfinished whole-C1 author work and acceptance. No installed-host, natural-model, scientific-benefit, same-capture body reduction or full net-benefit qualification is inferred.

On the preceding frozen integration 0305630c83170732bee664475b2c8ca0946c4f01, four complete modules pass 346 cases, four actual native routes cover exec, CodeMode, killed hook and unqualified-history recovery, and a separate scripted native consumer follows five initial-selection pages before writing an independently verified synthetic artifact. That selected body is 20,321 B from its 51,227 B complete capture (60.33%); this is a synthetic capture-to-selected-body metric, not new compression attributable to the pin repair, natural model adoption, a whole reading-cost control or net benefit.

Integration retains current main's qualified history versions and private full-wire oracle. Two existing test observers now watch effect_runtime_projection_result, preserving the exact RPC counts, original-input checks and refusal assertions. Before those fixture adaptations, ten observer AttributeErrors and two hidden-path fixture failures are retained. The repaired pre-pin native controls pass both cases; no product gate was weakened. New initial references use the existing explicit reader, while historical recipe/raw cursors retain digest refusal and same-capture recovery. This head has no passing exact-scope change-quality/premerge or merge-readiness claim; previous receipts retain their original scope.

Earlier managed-validation snapshot: a49625e3a7a3f6d87cc4f8dcba3ed556651d521f, with test-only provider isolation aa5da00a41060c4afd670915ba7b16e8d426a45d and the ordinary-fixture correction dd00992dd; final head e3d6283a9fad6e79a5afcd7c45dcac8ee358ad59 adds only the paired ledger. Current managed-validation matrix: eight passed, including actual native File/SQLite Session validation, refusal before argv/state/quota and no model replay on operation reentry. Current expanded shared-route qualification: 384 distinct passing cases across two integrated invocations (68 + 316), with no failures/skips/errors; earlier interrupted validation remains separately pinned. Earlier 18-case Host-transfer, eight-input transport parity and 528-case source qualification remain pinned to their preceding snapshots below. Latest target-main 8a9e116283df2796926ea153241f48c9c8fc3dbb is integrated in 4502a2f25e5410756119dcf3af85a95bf75ece70; 4cdf977c3aa2948bb4cc3ea149bcfd35f399e532 adapts its new explicit fixture caller. Installed-host adoption, complete cost and model/scientific benefit remain separate qualifications.

Draft reflects unfinished author work: the five complete-output budget holds, remaining complete-delivery gaps, supported-reader exit and original-quota single-session acceptance/full-cost comparison remain open. Pending review, CI or a quality receipt is not the reason for Draft.

Goal refusal before start reservation

The preceding check prevented Host effects but followed cadence reservation and initial journal/capture writes. Four real File/SQLite refusal cases retained a waiting interval despite zero Host calls; six valid/historical controls passed. The check now uses the exact input selected by original-capture staging before any such reservation or new publication. The same typed owner, historical decoder, staged replan exception, saved-result finalization and durable-attempt confirmation remain; no new configuration, state, bypass, cache or validation RPC is added.

The strengthened callback/real-argv matrix passes 22 cases and the four complete modules pass 357 (184.12 s). The first broader run's 356 passes and one old root-layout failure are retained; the latter also fails in a clean frozen 26585b7c7 checkout. Its test now reads the scheduler field through the actual immutable capture reader, preserving the original floor and rejected-reduction assertions. Two real-binary native cases pass (32.41 s). Ruff, semantic advisory, DCO, committed bilingual/whitespace, added-line scans and loopx check on all six changed paths pass.

Ten paired conditions run frozen executor functions and the candidate with the same real owners/source input and equal-width isolated namespaces. The measured warmed JSON RPC wire includes request/response lines, framing and error responses; it excludes startup, TCP overhead, CPU, model/invoice and accepted-outcome costs. In the four pending-Goal conditions, non-capture RPCs fall 5→2, capture RPCs 6→3: one admission and two journal writes disappear. Their File/SQLite wire totals are 180,817→85,524 / 181,041→85,636 bytes without capture, and 241,693→144,028 / 242,001→144,224 with capture. No new journal/capture is left, and all four subsequent valid attempts start where their frozen controls wait.

All six valid/historical conditions retain identical RPC counts and measured bytes, and Host inputs match after substituting only the independent capture namespace. The existing full-source dispatch check still costs its RPC; its earlier approximately 42.8 KB component was not removed. These are refusal-path transport and eligibility results, separate from the main same-capture body metric and full C1 benefit. Journal-write consolidation needs its own interruption/recovery qualification and is not included here.

Earlier generic Host dispatch checkpoint

A valid complete-source commitment could previously reach a generic callback or external argv while its Goal narrative read was still pending. Named adapters already checked the existing host_dispatch_context rule. The outer executor now calls that same owner before advancing the durable Host-attempt count, confirming start or launching work. No new state, vocabulary, configuration, decision algorithm or general capability-read obligation is introduced.

The new File/SQLite coverage exercises callback and real argv effects for current, pending-Goal and historical requests. Pending requests have no Host effect; current and historical controls reach the Host. Saved typed results still run their existing validator without starting another attempt, and staged replan material reads keep their existing exception. The corrected frozen baseline has two missing-refusal failures and four passing controls; an earlier invalid source-index fixture was correctly rejected and is not used as defect proof. All 14 new cases, the complete 345-case cohort and two actual native cases pass on the integrated source.

The additional correctness check costs one RPC per new complete-source Host attempt in the fixed component fixture: normalized params JSON plus result JSON totals 42,771 bytes (File) or 42,828 bytes (SQLite). Saved-result recovery retains its original single lineage RPC and byte counts; historical requests without a complete contract add no RPC. The comparison substitutes the frozen pre-repair function with the same real owner and fixture, rather than claiming a whole-branch A/B run. Framing, CPU, model input, invoices, outcome value and full net benefit are unmeasured. This is an explicit cost increase, separate from same-capture body bytes and all previous conditional savings.

The repair reuses the existing shared phase/helper; another wrapper or policy owner is unnecessary. The bilingual ledger and reusable self-repair pattern record the bounded fix and cost. Original writer/reader retirement, cross-host/configuration coverage and original-quota single-session outcome/full-cost acceptance remain open.

Shared owners and actual consumers

• Quota and native Turn share complete composition, numeric identity/edit, source integrity and reference owners before capture. Unknown optional facts remain; unsupported executable semantics reject before dispatch. Original admission, capability, context-delivery and settlement owners retain their authority.

• Default turn plan remains inspection-only. Explicit transaction detail and execution retain the complete source; adding transaction fields cannot turn source-less inspection into executable work.

• Native Codex and structured-host final inputs preserve the complete source and original material readers. Same-capture recovery retains source refusal, selection/reentry and the original Turn's single writeback/debit. Heartbeat routing, safety and complete work guidance share their existing rule owner; historical saved bodies are unchanged.

• Ark/MemT tool identity uses extensions.turn_host.extract_turn_work_identity. Complete input follows the selected-work owner; historical provider input retains raw declared values and strict string/NUL checks. MemT's candidate validator retains the whole original request, including its complete-source fence. Actual CLI input still has no admitted exact MemT lifetime; tests do not invent one.

• Bundled DSH reuses the identity already returned by that dispatch owner within the same invocation. It adds no RPC or cross-attempt cache. Its two old identity walks become one historical raw-value adapter. Fresh-iteration and Goal-instance hashes are preserved; complete adaptive bindings intentionally use the selected primary Todo. The retained private metadata is excluded from final model input.

• Native-child observation metadata uses the original complete interaction context, with the existing local positive-integer decoder and zero additional RPCs. Complete-source verification, launch fences, receipt admission, feature-off isolation and historical raw input are preserved. No protocol field or second authority is added.

This reuses existing CLI/Chat Host entrypoints and adds no capability, setting or user step. Steward and other historical consumers, native-child receipt/journal compatibility, executor v0 transport, optional product journeys and packaged/installed qualification remain open. Historical decoders retire only after their original-version recovery windows reconcile; saved receipts and sources are never rewritten as new-version work.

Earlier matched DSH checkpoint

Frozen ae8ee460a and the candidate consume the same four complete requests from real File/SQLite CLI plans through the actual DSH adapter. The provider is synthetic. Each uses one dispatch RPC; canonical argument/result JSON bytes are unchanged.

Source	Ordinary prompt bytes	Adaptive prompt bytes	Internal authority JSON growth
File	32,377	32,517	114 / 115 B
SQLite	32,425	32,565	114 / 115 B

All four prompts are digest-identical before/after. Ordinary session/tool identities are unchanged; adaptive bindings correct the old Todo mismatch. The last column is a Python-object JSON size proxy, not allocations or wire traffic. The probe changes no work state and spends nothing. This is a caller-correctness result, with no claim of packet compression, latency, model-token, invoice or matched-outcome/full-cost benefit.

Prior costs remain in the bilingual source-convergence ledger: the Ark identity path removed a full-source copy but added 113 canonical result bytes; Markdown removed display padding while still exceeding its original size hold. These are separate checkpoints, not cumulative net savings.

Matched native-child metadata observation

Eight fixed on/off requests extend the same real File/SQLite normal/adaptive CLI fixtures, with synthetic capability context recomposed through the existing owner. Frozen integration 5d9dbb811 and the candidate produce identical metadata, request bytes (40,202–40,778), prompt bytes (30,744–31,030), schema hashes and canonical argument/result sizes. Metadata uses zero RPCs and complete-input preparation one RPC on both sides. This qualifies removal of an obsolete field-layout dependency; latency, allocations, wire framing, model billing and complete Host costs were not measured.

Validation and remaining failures

Check	Result and scope
Current 99077996a reservation boundary	357 passed in four complete modules, including 22 strengthened cases; 2 actual native passed; matched 10 pairs qualify refusal/state changes and unchanged valid-path traffic. Quick 7 direct + 4 selected pass, overall quality_receipt_missing; current 132-file scope is review_required.
Prior 26585b7c7 generic dispatch	345 passed across three complete modules (327.19 s), including 14 new cases; 2 actual native passed (35.36 s). Ruff, semantic advisory (no supported new vocabulary carriers), committed bilingual and whitespace checks pass. Exact-head quick 7 direct + 4 selected pass, overall quality_receipt_missing; no passing current CQR.
Runtime c59777597 required-source integration	143 passed, six complete modules (68.99 s); 2 actual native passed (25.73 s), 124 TypeScript passed, typecheck and changed-test Ruff passed. File/SQLite current-preference lifecycle and actual CLI selected-work/Goal readback are exercised. Native tests use a deterministic localhost Responses fixture and the real Codex binary; scripted usage is not model/invoice evidence.
Frozen test-observation comparison	All five newly encountered failures reproduce on pre-preference C1 60ca5a71e. Original full run: 138 passed / 5 failed; repaired test observations retain unavailable-context signature refusal, complete Goal bytes, native reader provenance, selected Task/privacy constraints and historical decoding. Intermediate 3 passed / 2 failed, then 2 passed, and final full 143-case results are retained.
Retained CLI budget at c59777597	4 failed / 39 passed / 1 skipped, 35.53 s. Actual values: small complete Markdown 24,524>10,000; crowded plans 24,038 / 27,501>22,500; complete alias 24,245>13,000. Existing temporary-namespace depth limit explains the environment skip. Earlier matrix assertions stop before the independent multi-agent case; its historical 25,933>13,000 failure remains. Small differences from preceding observations are not attributed to product compression.
Runtime c59777597 quick premerge	7 direct + 4 selected checks passed. Overall failed, one policy failure: quality_receipt_missing for base f941b5939 / head c59777597, scope b8b0d87a92fac6cacba4e486d15598f1c65b0a5e27445ae0257c565c4c9422f0. No passing receipt, standard-premerge or merge-readiness claim.
Previous source-integration public boundary	Full 131-file C1 diff: 0 errors / 1 non-literal credential warning, inspected as a synthetic refusal string; the four new test/ledger paths scan clean. Committed bilingual and whitespace checks passed. DCO integration/test commits retain the original C1 proposal.
Earlier 94955ee9b integration	369 passed, five complete modules; 2 native / 143 TypeScript passed, typecheck passed. Focused source qualification; whole C1 remains unfinished.
Earlier managed postcondition matrix	8 passed, 20.911 s, no skips/errors. Two frozen bd2ea9c58 invocations cover four missing-refusal failures and four valid/historical controls. Source faults occur at the actual callback before independent argv; the native Session, canonical Task and settlement remain real. Six earlier ambient-provider startup failures are preserved, with fixture-only isolation and its one manual caller corrected.
Earlier integrated shared routes	68 + 316 passed across two invocations on 4cdf977c3 (335.058 / 644.420 s), 384 distinct cases across four complete modules, no failures/skips/errors; conversation-binding TypeScript owner: 45 passed / 0 failed / 0 skipped, 350.35 ms.
Pre-integration expanded native routes	At 9baa678db, five complete native/steward/calendar modules pass 359. The sixth private-conversation module was interrupted after 56 passed / 37 startup failures, 1,890.295 s for the overall incomplete run. All 37 report the provider startup rejection; two same-source environment-unset controls pass (3.064 s), then the exact two cases pass with only the neutral ordinary fixture isolated (4.402 s), keeping the original environment. The fixture change does not relax the production provider check. The nameless interrupt JUnit record is not a passing test, and this is not a full six-module pass.
Same-source native component observation	Two File/SQLite original-Session fixture cases pass, 14.070 s. Frozen Git methods are substituted with identical plan source hashes, dependencies and original argv. The additional baseline argv is read-only fixture validation; this is not a complete baseline-repository or installed-host run.
Previous Host-transfer refusal matrix	18 passed, 1.40 s. Corrected frozen 42b842f93: 10 missing-refusal failures / 8 passing controls. Both transfers, absent/valid/null/false/zero/list/text/empty-object/invalid-integrity cases, native and shared-reader refusal, no DSH runner load. Two initially unfulfilled-Goal controls were correctly rejected and retained before fixing only their complete-source fixture.
Previous Host-transfer source qualification	Initial six-module run: 522 passed / 6 failed, 298.49 s, no skips/errors. Frozen 42b842f93 reproduces all six pre-dispatch Host-context failures. Independent child import shows a different installed Core from the Goal cwd. Test-only source binding/readback plus explicit delivery-off fixes the historical direct-body control: 8 passed, 13.17 s in that complete module. Other five modules' 520 passes remain valid on unchanged product bytes. Original failures retained; no installed Core or user configuration changed.
Earlier 5a759b complete-plan cohort	334 passed / 0 failed / 0 skipped / 0 errors, 202.77 s. Includes actual File/SQLite delivery, failed-attempt recovery, validation and original-Turn settlement/replay. New six-case RPC/refusal matrix: 6 passed, 1.01 s; its frozen-source characterization retained five duplicate-call failures and one passing historical control.
Earlier native-child / Codex CLI / complete-source cohort	197 passed, 149.37 s. Includes real subprocess on/off observation/readback and corrupted-source refusal before another launch or receipt, plus real File/SQLite complete input and recovery. New tests first fail on legacy field access; the first wider run’s two incomplete-Goal fixture rejections are retained. The fixture now delivers its actual file/full revision; the product gate is unchanged.
Earlier 0cd native context / Turn driver	257 passed / 0 failed / 0 skipped / 0 errors, 116.07 s. Same two complete modules passed 257 in 118.43 s before the non-overlapping target-main chat integration; focused alias × off/on matrix: 4 passed, 2.42 s. Before fixture repair: 253 passed / 2 failed, 113.61 s; original failures retained.
Earlier owned operation Host	34 passed / 2 skipped, 24.16 s; the skipped cases require explicit live-model release qualification. No live model was called.
Earlier DSH regressions before repair	4 adaptive failures / 7 passing controls at the frozen implementation.
Earlier DSH and complete-source suites	367 passed, 67.66 s; fresh/reuse, adaptive work, raw historical values and invalid-source/version refusal before provider calls.
Earlier real File/SQLite CLI cases	2 passed, 4.41 s; built-in DSH, complete final input, independent validator, committed writeback and replay with one runner call and one debit.
Earlier optional Ark/MemT suites	44 / 65 passed, 17.60 / 52.00 s, with current Core source and isolated SDK/protocol fixtures. MemT's missing-client collection and five CLI import failures are retained: child imports independently exposed an older installed Core; binding only the test child's Python source path corrected them. No product/user configuration changed.
Earlier d104 quick premerge	6 direct checks and 4 selected checks passed; gate remains quality_stale_receipt (exit 1), with no exact-scope passing quality receipt or merge-readiness claim
Earlier diff, bilingual and public scans	Passed; both RFC mirrors changed together. The two runtime paths' diff-scoped advisory at the pre-integration snapshot found no supported new carriers; unsupported/dynamic forms are not certified. All new author/integration commits carry DCO sign-off. Earlier lint evidence remains pinned to its own snapshot.

The frozen 46bdaddb3 record retains six context-delivery and two driver-profile failures. At d1046499a, the two complete current modules instead produce 253 passes and two profile-alias failures. Actual responses reject native capture without a Turn identity before the tests access scheduler_hint; no scheduler-policy failure is inferred. The fixture now supplies its original Turn and exercises explicit off/native on for both aliases. Native cases follow only the generated original-capture context-read commands through contiguous UTF-8 pages, retaining the original scheduler assertions and verifying one admission, no spend, and unchanged Goal/event-log bytes after recovery. The two complete modules pass 257 cases with no skips or errors, including again on 0cd09b46c (116.07 s). The six older failures retain their original provenance; this recheck does not reassign their individual causes. Earlier 122-case TypeScript and ten-case steward results remain historical, with their environment failures and corrections in the ledger.

The five output holds also remain: small complete Markdown 24,531>10,000, crowded plans 24,044/27,508>22,500, complete alias 24,252>13,000, and the separate public multi-agent alias probe 25,933>13,000. No fixture population, requirement, route, capability, absolute/line/growth limit or historical failure was removed or relaxed.

There was no external model/solver/judge/paid scientific launch, CI status query/manual trigger, live installation, release, default promotion or merge. PostgreSQL and installed-host/model acceptance are not qualified by these source checks.

The 197-test cohort and 34-pass/2-live-opt-in-skip Host cohort were measured at e06e1a5c4f359a0b87e8009afaaf57b0e02f3767. The subsequent target-main integration changes six non-overlapping chat paths; both new complete-host feature-on/off subprocess cases passed again on d1046499ac8752eb2d944650fc29690d38c1b2eb (2 passed, 1.29 s). That earlier d104 integrated head also passed its recorded execution checks; no current passing quality receipt or merge-readiness is claimed.

Matched required-read traffic checkpoint

Frozen target-main c334bbaf196505778149ae4edeefaa013d343259 and C1 d1046499ac8752eb2d944650fc29690d38c1b2eb use identical fixture/command-producer AST and sequentially recreated identical physical fixture paths. The actual source module is verified in each checkout. After each JSON command, execute its declared pending Goal read, unless the original full Goal is already inline. Values below include final stdout newlines; all five rows deliver the complete fixture Goal text and selected Task text.

Public synthetic caller	Main initial + required-read stdout B	C1 initial + required-read stdout B	Difference	Required reads, main → C1
Small ordinary quota JSON	23,619	24,578	+959 / +4.06%	1 → 1
Small compatibility alias JSON	9,446	24,578	+15,132 / +160.19%	1 → 1
Small plan JSON	10,886	11,330	+444 / +4.08%	1 → 0
Crowded plan JSON	23,518	24,044	+526 / +2.24%	1 → 0
Crowded plan with Vision	26,981	27,508	+527 / +1.95%	1 → 0

Crowded C1 plans include the unmarked original Goal/Todo text: the 8,026-byte read is inline, explaining most initial-output growth. That content remains required. The current ordinary-quota comparison retains a positive 959-byte cost; these results establish no traffic saving. Legacy alias and complete input carry different information, so its large increase is not an outcome-matched efficiency comparison. The additional 18-Todo/3-Agent/12-run alias matrix gives 9,603 → 29,571 initial bytes; both sides lack the full Goal on that axis, and it is separate from the historical 25,933-byte probe.

This measures actual emitted body/declared-read stdout for local fixtures. It excludes tool framing, bridge/latency, tokens/cache, model comprehension/decisions, accepted outcomes and human attention; full single-session cost/benefit remains unqualified. Original limits, fixture population and historical failures are unchanged. No product source changed in this measurement checkpoint. The prior native-child metadata and DSH source-convergence results keep their original qualification scopes.

Matched planning validation transport

Eight frozen inputs from actual isolated File/SQLite CLI plans cover complete/historical sources and fresh/resume policies. Against C1 predecessor 0cd09b46c, complete planning drops from two RPCs to one because the typed lineage operation already performs the same version gate. The measured plan → Host request → prompt/schema preparation sequence drops from three to two RPCs.

Complete source, each fresh/resume case	Argument + result JSON before B	After B	Saved B	Saved relative to this sequence
File	92,598	81,554	11,044	11.93%
SQLite	92,838	81,752	11,086	11.94%

All eight final plan/request/prompt/schema files are byte-identical before/after; historical controls retain three RPCs and identical traffic. All four complete model inputs preserve their whole contract and original fixture Goal. Source files stay unchanged, probes spend nothing and no model is called. CSV, SQLite integer totals and Decimal ratios agree.

These are canonical RPC argument/result JSON sizes for this component sequence, excluding RPC framing, prior CLI collection/composition, latency, allocation, actual model usage, downstream work and settlement. Final model input is unchanged. This is a measured removal of one redundant transport/validation call, with no cache, new protocol, setting or authority; it does not establish the original-quota comparison, overall net saving or long-horizon benefit. It cannot be added to packet or selective-read percentages with different denominators. Unsupported versions, corrupt source, a self-consistently rehashed substituted adapter and null complete input still refuse through the existing typed owner.

Matched managed Session validation observation

Only quota.turn_envelope.evaluate calls are counted here. Preparation retains zero such calls by reusing the dispatch identity. At the post-model boundary the complete-source reader adds one, with 42,911 / 42,959 B canonical argument-plus-result JSON for File/SQLite in the same-source native-fixture comparison. These component observations are pinned to the pre-integration source and are not a new integrated full-cost rerun. The separate eight fixed CLI-source cases add 36,320 / 36,404 B for complete inputs and no such call for historical inputs. These populations are separate and are not added together.

This measured component cost increase supports source refusal; it is not compression. Framing, latency, allocation, canonical Task/admission reads, independent argv, model usage, recovery and accepted outcomes are excluded. Zero counted calls never means zero total cost. No whole-net-benefit value, including zero, is inferred from these observations.

Matched canonical Todo projection observations

The original output-budget failures retain their original unmarked Goal workload. An additional matched observation uses the existing canonical todo project-markdown CLI, with a fresh provider revision, on main f941b5939 and runtime C1 c59777597. File/SQLite each cover one Todo, 36 Todos without/with Vision, and 18 Todos across three agents. Across 16 projections and 64 quota/plan responses, complete canonical rows remain identical and independent full-file/region hashes reconstruct the delivered Goal narrative, including a final stop. Required emitted Goal reads are actually executed and counted. This qualifies an existing route; there is no new runtime compactor or fixture replacement.

C1 observation	Measured argv JSON + stdout before → after projection	C1 minus main after projection
File crowded plan, no Vision	24,352 → 16,654 B	+23 B
File crowded plan, Vision	27,815 → 20,118 B	+23 B
SQLite crowded plan, no / with Vision	24,366 → 16,668 / 27,829 → 20,132 B	+23 / +23 B
Small quota / plan	Grows by 364 / 75 B on both backends	Quota +959 B; plan +23–24 B

The one-time projection component costs 1,667–1,700 B. Counting both complete canonical readbacks as one-time audit cost puts the positive cases' conditional break-even at 7–9 repeated observations; regenerating every Turn is not recommended by these data. After projection C1 crowded quota still costs 1,958 / 1,759 B more than main without / with Vision. Ordinary multi-agent quota supplies neither inline Goal nor a Goal read in either snapshot and is excluded from Goal-equivalent comparisons. RPC, CPU, model use, invoice, decision/outcome equivalence and complete net benefit remain unqualified. These are separate source-workflow observations, not the same-capture full-to-selected-body reduction metric and not new C1 savings.

The new bilingual ledger passes staged/committed mirror, whitespace and public-boundary checks. Its prior 632c05a 131-file combined scope was 3e17c3033cbed6bbdb396872afa48afa90df309b8a597bad9c0b3fe5f121f270, review_required, without a passing current receipt. Existing runtime cohorts retain their exact snapshots. The original five holds, incomplete writer/reader cutover and original-quota single-session acceptance remain open.

Remaining C1 acceptance

Repair the measured output/delivery gaps through their owners, migrate the remaining actual consumers and qualify refusal/recovery/rollback before final writer cutover. Compare necessary semantics, actual decisions, accepted outcomes and complete costs on the same single-session path against original quota; evaluate two-session operation separately. The controller/executor RFC remains open. This reversible consumer increment does not certify complete C1 delivery or long-horizon benefit.

Bounded refactoring: complete planning reuses the version gate already inside its typed lineage operation. The current managed Session increment replaces two complete-input compatibility-envelope selections with the same complete owner: preparation reuses only invocation-local metadata, while the post-model boundary checks its source again. Historical decoding and standalone diagnostics remain; there is no new runtime helper, vocabulary, parallel parser or cross-call cache. The prior profile fixture keeps its shared local CLI invocation path. Original Goal content, negative cases, output limits and all five budget holds are retained. Earlier CLI/driver qualification retains its own snapshot. Earlier author validation retains its own pinned native/ordinary/private/Antigravity snapshots. The previous preference/readback integration retains its six-module 143-case and 124-TypeScript snapshots. Current reservation/capture qualification covers the four-module 357-case run, two actual native cases and exact-head checks above. The whole output-budget module still fails, and this head lacks its own passing quality receipt. This remains an unfinished C1 proposal, with no current merge-readiness or installed-host/full-cost qualification.

[mergify]

This pull request has merge conflicts with main and cannot be merged
until they are resolved. Please rebase or merge the base branch, @loopx-agent.

Choose the remote for the base repository, not an out-of-date fork.
For a fork clone, first inspect git remote -v; upstream must point
to https://github.com/loopx-project/loopx.git. If it is absent, add it
with git remote add upstream https://github.com/loopx-project/loopx.git.
Then run:

git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEAD

For a same-repository clone whose origin points to
https://github.com/loopx-project/loopx.git, use origin instead of
upstream for fetch/rebase. If you prefer merging the base, use
git merge <base-remote>/main and push normally.

Keep the DCO Signed-off-by trailer on every commit when you rebase.
https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts with main and cannot be merged
until they are resolved. Please rebase or merge the base branch, @loopx-agent.

Choose the remote for the base repository, not an out-of-date fork.
For a fork clone, first inspect git remote -v; upstream must point
to https://github.com/loopx-project/loopx.git. If it is absent, add it
with git remote add upstream https://github.com/loopx-project/loopx.git.
Then run:

git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEAD

For a same-repository clone whose origin points to
https://github.com/loopx-project/loopx.git, use origin instead of
upstream for fetch/rebase. If you prefer merging the base, use
git merge <base-remote>/main and push normally.

Keep the DCO Signed-off-by trailer on every commit when you rebase.
https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts with main and cannot be merged
until they are resolved. Please rebase or merge the base branch, @loopx-agent.

Choose the remote for the base repository, not an out-of-date fork.
For a fork clone, first inspect git remote -v; upstream must point
to https://github.com/loopx-project/loopx.git. If it is absent, add it
with git remote add upstream https://github.com/loopx-project/loopx.git.
Then run:

git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEAD

For a same-repository clone whose origin points to
https://github.com/loopx-project/loopx.git, use origin instead of
upstream for fetch/rebase. If you prefer merging the base, use
git merge <base-remote>/main and push normally.

Keep the DCO Signed-off-by trailer on every commit when you rebase.
https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts with main and cannot be merged
until they are resolved. Please rebase or merge the base branch, @loopx-agent.

Choose the remote for the base repository, not an out-of-date fork.
For a fork clone, first inspect git remote -v; upstream must point
to https://github.com/loopx-project/loopx.git. If it is absent, add it
with git remote add upstream https://github.com/loopx-project/loopx.git.
Then run:

git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEAD

For a same-repository clone whose origin points to
https://github.com/loopx-project/loopx.git, use origin instead of
upstream for fetch/rebase. If you prefer merging the base, use
git merge <base-remote>/main and push normally.

Keep the DCO Signed-off-by trailer on every commit when you rebase.
https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts with main and cannot be merged
until they are resolved. Please rebase or merge the base branch, @loopx-agent.

Choose the remote for the base repository, not an out-of-date fork.
For a fork clone, first inspect git remote -v; upstream must point
to https://github.com/loopx-project/loopx.git. If it is absent, add it
with git remote add upstream https://github.com/loopx-project/loopx.git.
Then run:

git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEAD

For a same-repository clone whose origin points to
https://github.com/loopx-project/loopx.git, use origin instead of
upstream for fetch/rebase. If you prefer merging the base, use
git merge <base-remote>/main and push normally.

Keep the DCO Signed-off-by trailer on every commit when you rebase.
https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts with main and cannot be merged
until they are resolved. Please rebase or merge the base branch, @loopx-agent.

Choose the remote for the base repository, not an out-of-date fork.
For a fork clone, first inspect git remote -v; upstream must point
to https://github.com/loopx-project/loopx.git. If it is absent, add it
with git remote add upstream https://github.com/loopx-project/loopx.git.
Then run:

git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEAD

For a same-repository clone whose origin points to
https://github.com/loopx-project/loopx.git, use origin instead of
upstream for fetch/rebase. If you prefer merging the base, use
git merge <base-remote>/main and push normally.

Keep the DCO Signed-off-by trailer on every commit when you rebase.
https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork