feat(runtime): validate provider request traces

[Nyvo-io]

Summary

• version provider-request captures as schema v2 while preserving legacy v1 reads
• hash protocol-independent request semantics while preserving normalized reasoning effort
• parse capture and physical-attempt rows and fail closed on malformed, torn, missing, or mismatched evidence
• validate exact expected identity sets across continuation runs and autonomous retries, including retries that use different sessions
• preserve the first trace-write failure in memory, attempt header and sentinel persistence independently, and carry the latch into the terminal header

The exported artifact is intentionally scoped to top-level Harbor invocations. Semantic-compaction and child-agent provider dispatches are outside this contract; the Runtime normalization and attempt-codec ownership cleanups are deferred to the broader Runtime refactor rather than duplicated here.

This is the second flat split requested in #1451 and addresses the request-trace evidence findings on that PR.

Refs #1269

Verification

• provider-request-trace: 23 passed
• SessionManager: 235 passed
• terminal ledger: 26 passed
• Runtime full suite: 2,558 passed, 7 skipped
• Headless full suite: 1,365 passed, 2 skipped
• runtime and headless typecheck passed
• lint: 2,160 files passed
• format check: 1,116 files passed
• git diff check passed
• local review-agent and independent Codex diff review: no findings

Not run: the paid Kimi/Harbor benchmark. That belongs to the separate stacked benchmark PR.

[Astro-Han]

Requesting changes for the P1 trace-completeness issue noted inline.

Non-blocking suggestion: the attempt parser requires integer timestamps, but AiSdkBackendInput.now, ProviderRequestAttemptRecord, and the AgentRun codec allow any finite number. A fractional monotonic clock produces a valid attempt that the reader rejects as invalid_attempt. Either accept finite non-negative values or normalize the timestamps at the Runtime producer. Keeping the decoder next to the Runtime record type would also prevent the two contracts from drifting again.

[Nyvo-io]

All requested changes are addressed and the inline thread is resolved. d62b136 fails closed on trace write/corruption sentinels and accepts the Runtime finite timestamp contract; 3e2c222 normalizes equivalent cross-protocol request evidence; 370c437 validates same-session continuation executions. The focused regressions, full workspace suite, and GitHub typecheck/test/e2e checks all pass. @Astro-Han, please re-review when convenient.

[Nyvo-io]

Latest P1 is addressed in af6079f and ab70f7f, with the requested real FileAgentRunStore export/read/assert regression chain plus torn-tail fail-closed coverage. The branch is rebased onto current main; npm test, typecheck, lint, format:check, diff --check, and the completed independent review all pass. @Astro-Han, please re-review when convenient.

[Astro-Han]

The earlier trace-latch and torn-tail failures are fixed. I found two other ways the trace can still pass when the experiment is not comparable or complete; details inline.

[Nyvo-io]

All four inline findings are addressed and resolved at c1917e8f:

• trace-write header latch and torn-tail evidence are preserved through the real FileAgentRunStore export path;
• protocol-independent hashes retain normalized reasoning effort and reject max/low/none mismatches;
• missing evidence for an earlier continuation invocation now emits an identity-bearing diagnostic and fails closed.

The branch is clean, rebased onto current main, and the focused/full tests plus typecheck, lint, format, diff checks, and independent review pass. Please re-review the current head when convenient.

[Astro-Han]

I did not find a P0 or P1 issue, so I am approving this PR. The earlier trace-integrity findings are fixed. I found these follow-ups:

• [P2] A failed attempt append can still be missed if the first header-latch write also fails and later terminal writes recover.
• [P2] Autonomous attempts from different sessions are exported into one file, but the reader treats the second session as mixed_identity.
• [P2] Semantic-compaction provider calls bypass the request tracker.
• [P2] Child-agent runs are outside the exported run set, so their provider calls are absent without a missing-evidence diagnostic.
• [P3] Request-shape normalization repeats provider-option and output-limit rules owned by Runtime.
• [P3] Headless still maintains a separate decoder for Runtime-owned attempt records.

These affect compound I/O failures, optional execution modes, or benchmark evidence rather than production state.

Please push back if provider-request traces intentionally exclude compaction or child-agent dispatches. If that is the intended boundary, narrowing the public contract may be enough. There is no need to accept a code change solely to clear the comment.

[Nyvo-io]

Addressed the latest review in d73dc7c2:

• the first trace-write failure is now latched in memory, header and sentinel writes are attempted independently, and the terminal header carries the latch;
• autonomous retries across different sessions are accepted while run/turn identity conflicts still fail closed, with exact expectedIdentities validation available;
• the public contract now explicitly covers exported top-level Harbor invocations only, so semantic compaction and child-agent dispatches are intentionally outside this artifact;
• the two P3 ownership cleanups are left for the broader Runtime refactor instead of adding more duplication in this PR.

Focused tests, full Runtime/Headless suites, typecheck, lint, format, diff checks, and independent reviews pass. GitHub typecheck, test, and e2e are also green.

[Astro-Han]

Reviewed d73dc7c2; merged as b308282d1.

The last commit fixes trace-write error latching and multi-session identity handling. The scope for compaction and child-agent calls is now explicit instead of being counted as complete provider accounting. I tested the merge with the then-current main; 544 focused tests passed. No new P0–P3.

[Nyvo-io]

Summary

• version provider-request captures as schema v2 while preserving legacy v1 reads
• hash protocol-independent request semantics while preserving normalized reasoning effort
• parse capture and physical-attempt rows and fail closed on malformed, torn, missing, or mismatched evidence
• validate exact expected identity sets across continuation runs and autonomous retries, including retries that use different sessions
• preserve the first trace-write failure in memory, attempt header and sentinel persistence independently, and carry the latch into the terminal header

The exported artifact is intentionally scoped to top-level Harbor invocations. Semantic-compaction and child-agent provider dispatches are outside this contract; the Runtime normalization and attempt-codec ownership cleanups are deferred to the broader Runtime refactor rather than duplicated here.

This is the second flat split requested in #1451 and addresses the request-trace evidence findings on that PR.

Refs #1269

Verification

• provider-request-trace: 23 passed
• SessionManager: 235 passed
• terminal ledger: 26 passed
• Runtime full suite: 2,558 passed, 7 skipped
• Headless full suite: 1,365 passed, 2 skipped
• runtime and headless typecheck passed
• lint: 2,160 files passed
• format check: 1,116 files passed
• git diff check passed
• local review-agent and independent Codex diff review: no findings

Not run: the paid Kimi/Harbor benchmark. That belongs to the separate stacked benchmark PR.

[Astro-Han]

Requesting changes for the P1 trace-completeness issue noted inline.

Non-blocking suggestion: the attempt parser requires integer timestamps, but AiSdkBackendInput.now, ProviderRequestAttemptRecord, and the AgentRun codec allow any finite number. A fractional monotonic clock produces a valid attempt that the reader rejects as invalid_attempt. Either accept finite non-negative values or normalize the timestamps at the Runtime producer. Keeping the decoder next to the Runtime record type would also prevent the two contracts from drifting again.

[Nyvo-io]

All requested changes are addressed and the inline thread is resolved. d62b136 fails closed on trace write/corruption sentinels and accepts the Runtime finite timestamp contract; 3e2c222 normalizes equivalent cross-protocol request evidence; 370c437 validates same-session continuation executions. The focused regressions, full workspace suite, and GitHub typecheck/test/e2e checks all pass. @Astro-Han, please re-review when convenient.

[Nyvo-io]

Latest P1 is addressed in af6079f and ab70f7f, with the requested real FileAgentRunStore export/read/assert regression chain plus torn-tail fail-closed coverage. The branch is rebased onto current main; npm test, typecheck, lint, format:check, diff --check, and the completed independent review all pass. @Astro-Han, please re-review when convenient.

[Astro-Han]

The earlier trace-latch and torn-tail failures are fixed. I found two other ways the trace can still pass when the experiment is not comparable or complete; details inline.

[Nyvo-io]

All four inline findings are addressed and resolved at c1917e8f:

• trace-write header latch and torn-tail evidence are preserved through the real FileAgentRunStore export path;
• protocol-independent hashes retain normalized reasoning effort and reject max/low/none mismatches;
• missing evidence for an earlier continuation invocation now emits an identity-bearing diagnostic and fails closed.

The branch is clean, rebased onto current main, and the focused/full tests plus typecheck, lint, format, diff checks, and independent review pass. Please re-review the current head when convenient.

[Astro-Han]

I did not find a P0 or P1 issue, so I am approving this PR. The earlier trace-integrity findings are fixed. I found these follow-ups:

• [P2] A failed attempt append can still be missed if the first header-latch write also fails and later terminal writes recover.
• [P2] Autonomous attempts from different sessions are exported into one file, but the reader treats the second session as mixed_identity.
• [P2] Semantic-compaction provider calls bypass the request tracker.
• [P2] Child-agent runs are outside the exported run set, so their provider calls are absent without a missing-evidence diagnostic.
• [P3] Request-shape normalization repeats provider-option and output-limit rules owned by Runtime.
• [P3] Headless still maintains a separate decoder for Runtime-owned attempt records.

These affect compound I/O failures, optional execution modes, or benchmark evidence rather than production state.

Please push back if provider-request traces intentionally exclude compaction or child-agent dispatches. If that is the intended boundary, narrowing the public contract may be enough. There is no need to accept a code change solely to clear the comment.

[Nyvo-io]

Addressed the latest review in d73dc7c2:

• the first trace-write failure is now latched in memory, header and sentinel writes are attempted independently, and the terminal header carries the latch;
• autonomous retries across different sessions are accepted while run/turn identity conflicts still fail closed, with exact expectedIdentities validation available;
• the public contract now explicitly covers exported top-level Harbor invocations only, so semantic compaction and child-agent dispatches are intentionally outside this artifact;
• the two P3 ownership cleanups are left for the broader Runtime refactor instead of adding more duplication in this PR.

Focused tests, full Runtime/Headless suites, typecheck, lint, format, diff checks, and independent reviews pass. GitHub typecheck, test, and e2e are also green.

[Astro-Han]

Reviewed d73dc7c2; merged as b308282d1.

The last commit fixes trace-write error latching and multi-session identity handling. The scope for compaction and child-agent calls is now explicit instead of being counted as complete provider accounting. I tested the merge with the then-current main; 544 focused tests passed. No new P0–P3.

[Nyvo-io]

Summary

• version provider-request captures as schema v2 while preserving legacy v1 reads
• hash protocol-independent request semantics while preserving normalized reasoning effort
• parse capture and physical-attempt rows and fail closed on malformed, torn, missing, or mismatched evidence
• validate exact expected identity sets across continuation runs and autonomous retries, including retries that use different sessions
• preserve the first trace-write failure in memory, attempt header and sentinel persistence independently, and carry the latch into the terminal header

The exported artifact is intentionally scoped to top-level Harbor invocations. Semantic-compaction and child-agent provider dispatches are outside this contract; the Runtime normalization and attempt-codec ownership cleanups are deferred to the broader Runtime refactor rather than duplicated here.

This is the second flat split requested in #1451 and addresses the request-trace evidence findings on that PR.

Refs #1269

Verification

• provider-request-trace: 23 passed
• SessionManager: 235 passed
• terminal ledger: 26 passed
• Runtime full suite: 2,558 passed, 7 skipped
• Headless full suite: 1,365 passed, 2 skipped
• runtime and headless typecheck passed
• lint: 2,160 files passed
• format check: 1,116 files passed
• git diff check passed
• local review-agent and independent Codex diff review: no findings

Not run: the paid Kimi/Harbor benchmark. That belongs to the separate stacked benchmark PR.

[Astro-Han]

Requesting changes for the P1 trace-completeness issue noted inline.

Non-blocking suggestion: the attempt parser requires integer timestamps, but AiSdkBackendInput.now, ProviderRequestAttemptRecord, and the AgentRun codec allow any finite number. A fractional monotonic clock produces a valid attempt that the reader rejects as invalid_attempt. Either accept finite non-negative values or normalize the timestamps at the Runtime producer. Keeping the decoder next to the Runtime record type would also prevent the two contracts from drifting again.

[Nyvo-io]

All requested changes are addressed and the inline thread is resolved. d62b136 fails closed on trace write/corruption sentinels and accepts the Runtime finite timestamp contract; 3e2c222 normalizes equivalent cross-protocol request evidence; 370c437 validates same-session continuation executions. The focused regressions, full workspace suite, and GitHub typecheck/test/e2e checks all pass. @Astro-Han, please re-review when convenient.

[Nyvo-io]

Latest P1 is addressed in af6079f and ab70f7f, with the requested real FileAgentRunStore export/read/assert regression chain plus torn-tail fail-closed coverage. The branch is rebased onto current main; npm test, typecheck, lint, format:check, diff --check, and the completed independent review all pass. @Astro-Han, please re-review when convenient.

[Astro-Han]

The earlier trace-latch and torn-tail failures are fixed. I found two other ways the trace can still pass when the experiment is not comparable or complete; details inline.

[Nyvo-io]

All four inline findings are addressed and resolved at c1917e8f:

• trace-write header latch and torn-tail evidence are preserved through the real FileAgentRunStore export path;
• protocol-independent hashes retain normalized reasoning effort and reject max/low/none mismatches;
• missing evidence for an earlier continuation invocation now emits an identity-bearing diagnostic and fails closed.

The branch is clean, rebased onto current main, and the focused/full tests plus typecheck, lint, format, diff checks, and independent review pass. Please re-review the current head when convenient.

[Astro-Han]

I did not find a P0 or P1 issue, so I am approving this PR. The earlier trace-integrity findings are fixed. I found these follow-ups:

• [P2] A failed attempt append can still be missed if the first header-latch write also fails and later terminal writes recover.
• [P2] Autonomous attempts from different sessions are exported into one file, but the reader treats the second session as mixed_identity.
• [P2] Semantic-compaction provider calls bypass the request tracker.
• [P2] Child-agent runs are outside the exported run set, so their provider calls are absent without a missing-evidence diagnostic.
• [P3] Request-shape normalization repeats provider-option and output-limit rules owned by Runtime.
• [P3] Headless still maintains a separate decoder for Runtime-owned attempt records.

These affect compound I/O failures, optional execution modes, or benchmark evidence rather than production state.

Please push back if provider-request traces intentionally exclude compaction or child-agent dispatches. If that is the intended boundary, narrowing the public contract may be enough. There is no need to accept a code change solely to clear the comment.

[Nyvo-io]

Addressed the latest review in d73dc7c2:

• the first trace-write failure is now latched in memory, header and sentinel writes are attempted independently, and the terminal header carries the latch;
• autonomous retries across different sessions are accepted while run/turn identity conflicts still fail closed, with exact expectedIdentities validation available;
• the public contract now explicitly covers exported top-level Harbor invocations only, so semantic compaction and child-agent dispatches are intentionally outside this artifact;
• the two P3 ownership cleanups are left for the broader Runtime refactor instead of adding more duplication in this PR.

Focused tests, full Runtime/Headless suites, typecheck, lint, format, diff checks, and independent reviews pass. GitHub typecheck, test, and e2e are also green.

[Astro-Han]

Reviewed d73dc7c2; merged as b308282d1.

The last commit fixes trace-write error latching and multi-session identity handling. The scope for compaction and child-agent calls is now explicit instead of being counted as complete provider accounting. I tested the merge with the then-current main; 544 focused tests passed. No new P0–P3.

[Nyvo-io]

Summary

• version provider-request captures as schema v2 while preserving legacy v1 reads
• hash protocol-independent request semantics while preserving normalized reasoning effort
• parse capture and physical-attempt rows and fail closed on malformed, torn, missing, or mismatched evidence
• validate exact expected identity sets across continuation runs and autonomous retries, including retries that use different sessions
• preserve the first trace-write failure in memory, attempt header and sentinel persistence independently, and carry the latch into the terminal header

The exported artifact is intentionally scoped to top-level Harbor invocations. Semantic-compaction and child-agent provider dispatches are outside this contract; the Runtime normalization and attempt-codec ownership cleanups are deferred to the broader Runtime refactor rather than duplicated here.

This is the second flat split requested in #1451 and addresses the request-trace evidence findings on that PR.

Refs #1269

Verification

• provider-request-trace: 23 passed
• SessionManager: 235 passed
• terminal ledger: 26 passed
• Runtime full suite: 2,558 passed, 7 skipped
• Headless full suite: 1,365 passed, 2 skipped
• runtime and headless typecheck passed
• lint: 2,160 files passed
• format check: 1,116 files passed
• git diff check passed
• local review-agent and independent Codex diff review: no findings

Not run: the paid Kimi/Harbor benchmark. That belongs to the separate stacked benchmark PR.

[Astro-Han]

Requesting changes for the P1 trace-completeness issue noted inline.

Non-blocking suggestion: the attempt parser requires integer timestamps, but AiSdkBackendInput.now, ProviderRequestAttemptRecord, and the AgentRun codec allow any finite number. A fractional monotonic clock produces a valid attempt that the reader rejects as invalid_attempt. Either accept finite non-negative values or normalize the timestamps at the Runtime producer. Keeping the decoder next to the Runtime record type would also prevent the two contracts from drifting again.

[Nyvo-io]

All requested changes are addressed and the inline thread is resolved. d62b136 fails closed on trace write/corruption sentinels and accepts the Runtime finite timestamp contract; 3e2c222 normalizes equivalent cross-protocol request evidence; 370c437 validates same-session continuation executions. The focused regressions, full workspace suite, and GitHub typecheck/test/e2e checks all pass. @Astro-Han, please re-review when convenient.

[Nyvo-io]

Latest P1 is addressed in af6079f and ab70f7f, with the requested real FileAgentRunStore export/read/assert regression chain plus torn-tail fail-closed coverage. The branch is rebased onto current main; npm test, typecheck, lint, format:check, diff --check, and the completed independent review all pass. @Astro-Han, please re-review when convenient.

[Astro-Han]

The earlier trace-latch and torn-tail failures are fixed. I found two other ways the trace can still pass when the experiment is not comparable or complete; details inline.

[Nyvo-io]

All four inline findings are addressed and resolved at c1917e8f:

• trace-write header latch and torn-tail evidence are preserved through the real FileAgentRunStore export path;
• protocol-independent hashes retain normalized reasoning effort and reject max/low/none mismatches;
• missing evidence for an earlier continuation invocation now emits an identity-bearing diagnostic and fails closed.

The branch is clean, rebased onto current main, and the focused/full tests plus typecheck, lint, format, diff checks, and independent review pass. Please re-review the current head when convenient.

[Astro-Han]

I did not find a P0 or P1 issue, so I am approving this PR. The earlier trace-integrity findings are fixed. I found these follow-ups:

• [P2] A failed attempt append can still be missed if the first header-latch write also fails and later terminal writes recover.
• [P2] Autonomous attempts from different sessions are exported into one file, but the reader treats the second session as mixed_identity.
• [P2] Semantic-compaction provider calls bypass the request tracker.
• [P2] Child-agent runs are outside the exported run set, so their provider calls are absent without a missing-evidence diagnostic.
• [P3] Request-shape normalization repeats provider-option and output-limit rules owned by Runtime.
• [P3] Headless still maintains a separate decoder for Runtime-owned attempt records.

These affect compound I/O failures, optional execution modes, or benchmark evidence rather than production state.

Please push back if provider-request traces intentionally exclude compaction or child-agent dispatches. If that is the intended boundary, narrowing the public contract may be enough. There is no need to accept a code change solely to clear the comment.

[Nyvo-io]

Addressed the latest review in d73dc7c2:

• the first trace-write failure is now latched in memory, header and sentinel writes are attempted independently, and the terminal header carries the latch;
• autonomous retries across different sessions are accepted while run/turn identity conflicts still fail closed, with exact expectedIdentities validation available;
• the public contract now explicitly covers exported top-level Harbor invocations only, so semantic compaction and child-agent dispatches are intentionally outside this artifact;
• the two P3 ownership cleanups are left for the broader Runtime refactor instead of adding more duplication in this PR.

Focused tests, full Runtime/Headless suites, typecheck, lint, format, diff checks, and independent reviews pass. GitHub typecheck, test, and e2e are also green.

[Astro-Han]

Reviewed d73dc7c2; merged as b308282d1.

The last commit fixes trace-write error latching and multi-session identity handling. The scope for compaction and child-agent calls is now explicit instead of being counted as complete provider accounting. I tested the merge with the then-current main; 544 focused tests passed. No new P0–P3.

[Nyvo-io]

Summary

• version provider-request captures as schema v2 while preserving legacy v1 reads
• hash protocol-independent request semantics while preserving normalized reasoning effort
• parse capture and physical-attempt rows and fail closed on malformed, torn, missing, or mismatched evidence
• validate exact expected identity sets across continuation runs and autonomous retries, including retries that use different sessions
• preserve the first trace-write failure in memory, attempt header and sentinel persistence independently, and carry the latch into the terminal header

The exported artifact is intentionally scoped to top-level Harbor invocations. Semantic-compaction and child-agent provider dispatches are outside this contract; the Runtime normalization and attempt-codec ownership cleanups are deferred to the broader Runtime refactor rather than duplicated here.

This is the second flat split requested in #1451 and addresses the request-trace evidence findings on that PR.

Refs #1269

Verification

• provider-request-trace: 23 passed
• SessionManager: 235 passed
• terminal ledger: 26 passed
• Runtime full suite: 2,558 passed, 7 skipped
• Headless full suite: 1,365 passed, 2 skipped
• runtime and headless typecheck passed
• lint: 2,160 files passed
• format check: 1,116 files passed
• git diff check passed
• local review-agent and independent Codex diff review: no findings

Not run: the paid Kimi/Harbor benchmark. That belongs to the separate stacked benchmark PR.

[Astro-Han]

Requesting changes for the P1 trace-completeness issue noted inline.

Non-blocking suggestion: the attempt parser requires integer timestamps, but AiSdkBackendInput.now, ProviderRequestAttemptRecord, and the AgentRun codec allow any finite number. A fractional monotonic clock produces a valid attempt that the reader rejects as invalid_attempt. Either accept finite non-negative values or normalize the timestamps at the Runtime producer. Keeping the decoder next to the Runtime record type would also prevent the two contracts from drifting again.

[Nyvo-io]

All requested changes are addressed and the inline thread is resolved. d62b136 fails closed on trace write/corruption sentinels and accepts the Runtime finite timestamp contract; 3e2c222 normalizes equivalent cross-protocol request evidence; 370c437 validates same-session continuation executions. The focused regressions, full workspace suite, and GitHub typecheck/test/e2e checks all pass. @Astro-Han, please re-review when convenient.

[Nyvo-io]

Latest P1 is addressed in af6079f and ab70f7f, with the requested real FileAgentRunStore export/read/assert regression chain plus torn-tail fail-closed coverage. The branch is rebased onto current main; npm test, typecheck, lint, format:check, diff --check, and the completed independent review all pass. @Astro-Han, please re-review when convenient.

[Astro-Han]

The earlier trace-latch and torn-tail failures are fixed. I found two other ways the trace can still pass when the experiment is not comparable or complete; details inline.

[Nyvo-io]

All four inline findings are addressed and resolved at c1917e8f:

• trace-write header latch and torn-tail evidence are preserved through the real FileAgentRunStore export path;
• protocol-independent hashes retain normalized reasoning effort and reject max/low/none mismatches;
• missing evidence for an earlier continuation invocation now emits an identity-bearing diagnostic and fails closed.

The branch is clean, rebased onto current main, and the focused/full tests plus typecheck, lint, format, diff checks, and independent review pass. Please re-review the current head when convenient.

[Astro-Han]

I did not find a P0 or P1 issue, so I am approving this PR. The earlier trace-integrity findings are fixed. I found these follow-ups:

• [P2] A failed attempt append can still be missed if the first header-latch write also fails and later terminal writes recover.
• [P2] Autonomous attempts from different sessions are exported into one file, but the reader treats the second session as mixed_identity.
• [P2] Semantic-compaction provider calls bypass the request tracker.
• [P2] Child-agent runs are outside the exported run set, so their provider calls are absent without a missing-evidence diagnostic.
• [P3] Request-shape normalization repeats provider-option and output-limit rules owned by Runtime.
• [P3] Headless still maintains a separate decoder for Runtime-owned attempt records.

These affect compound I/O failures, optional execution modes, or benchmark evidence rather than production state.

Please push back if provider-request traces intentionally exclude compaction or child-agent dispatches. If that is the intended boundary, narrowing the public contract may be enough. There is no need to accept a code change solely to clear the comment.

[Nyvo-io]

Addressed the latest review in d73dc7c2:

• the first trace-write failure is now latched in memory, header and sentinel writes are attempted independently, and the terminal header carries the latch;
• autonomous retries across different sessions are accepted while run/turn identity conflicts still fail closed, with exact expectedIdentities validation available;
• the public contract now explicitly covers exported top-level Harbor invocations only, so semantic compaction and child-agent dispatches are intentionally outside this artifact;
• the two P3 ownership cleanups are left for the broader Runtime refactor instead of adding more duplication in this PR.

Focused tests, full Runtime/Headless suites, typecheck, lint, format, diff checks, and independent reviews pass. GitHub typecheck, test, and e2e are also green.

[Astro-Han]

Reviewed d73dc7c2; merged as b308282d1.

The last commit fixes trace-write error latching and multi-session identity handling. The scope for compaction and child-agent calls is now explicit instead of being counted as complete provider accounting. I tested the merge with the then-current main; 544 focused tests passed. No new P0–P3.

[Nyvo-io]

Summary

• version provider-request captures as schema v2 while preserving legacy v1 reads
• hash protocol-independent request semantics while preserving normalized reasoning effort
• parse capture and physical-attempt rows and fail closed on malformed, torn, missing, or mismatched evidence
• validate exact expected identity sets across continuation runs and autonomous retries, including retries that use different sessions
• preserve the first trace-write failure in memory, attempt header and sentinel persistence independently, and carry the latch into the terminal header

The exported artifact is intentionally scoped to top-level Harbor invocations. Semantic-compaction and child-agent provider dispatches are outside this contract; the Runtime normalization and attempt-codec ownership cleanups are deferred to the broader Runtime refactor rather than duplicated here.

This is the second flat split requested in #1451 and addresses the request-trace evidence findings on that PR.

Refs #1269

Verification

• provider-request-trace: 23 passed
• SessionManager: 235 passed
• terminal ledger: 26 passed
• Runtime full suite: 2,558 passed, 7 skipped
• Headless full suite: 1,365 passed, 2 skipped
• runtime and headless typecheck passed
• lint: 2,160 files passed
• format check: 1,116 files passed
• git diff check passed
• local review-agent and independent Codex diff review: no findings

Not run: the paid Kimi/Harbor benchmark. That belongs to the separate stacked benchmark PR.

[Astro-Han]

Requesting changes for the P1 trace-completeness issue noted inline.

Non-blocking suggestion: the attempt parser requires integer timestamps, but AiSdkBackendInput.now, ProviderRequestAttemptRecord, and the AgentRun codec allow any finite number. A fractional monotonic clock produces a valid attempt that the reader rejects as invalid_attempt. Either accept finite non-negative values or normalize the timestamps at the Runtime producer. Keeping the decoder next to the Runtime record type would also prevent the two contracts from drifting again.

[Nyvo-io]

All requested changes are addressed and the inline thread is resolved. d62b136 fails closed on trace write/corruption sentinels and accepts the Runtime finite timestamp contract; 3e2c222 normalizes equivalent cross-protocol request evidence; 370c437 validates same-session continuation executions. The focused regressions, full workspace suite, and GitHub typecheck/test/e2e checks all pass. @Astro-Han, please re-review when convenient.

[Nyvo-io]

Latest P1 is addressed in af6079f and ab70f7f, with the requested real FileAgentRunStore export/read/assert regression chain plus torn-tail fail-closed coverage. The branch is rebased onto current main; npm test, typecheck, lint, format:check, diff --check, and the completed independent review all pass. @Astro-Han, please re-review when convenient.

[Astro-Han]

The earlier trace-latch and torn-tail failures are fixed. I found two other ways the trace can still pass when the experiment is not comparable or complete; details inline.

[Nyvo-io]

All four inline findings are addressed and resolved at c1917e8f:

• trace-write header latch and torn-tail evidence are preserved through the real FileAgentRunStore export path;
• protocol-independent hashes retain normalized reasoning effort and reject max/low/none mismatches;
• missing evidence for an earlier continuation invocation now emits an identity-bearing diagnostic and fails closed.

The branch is clean, rebased onto current main, and the focused/full tests plus typecheck, lint, format, diff checks, and independent review pass. Please re-review the current head when convenient.

[Astro-Han]

I did not find a P0 or P1 issue, so I am approving this PR. The earlier trace-integrity findings are fixed. I found these follow-ups:

• [P2] A failed attempt append can still be missed if the first header-latch write also fails and later terminal writes recover.
• [P2] Autonomous attempts from different sessions are exported into one file, but the reader treats the second session as mixed_identity.
• [P2] Semantic-compaction provider calls bypass the request tracker.
• [P2] Child-agent runs are outside the exported run set, so their provider calls are absent without a missing-evidence diagnostic.
• [P3] Request-shape normalization repeats provider-option and output-limit rules owned by Runtime.
• [P3] Headless still maintains a separate decoder for Runtime-owned attempt records.

These affect compound I/O failures, optional execution modes, or benchmark evidence rather than production state.

Please push back if provider-request traces intentionally exclude compaction or child-agent dispatches. If that is the intended boundary, narrowing the public contract may be enough. There is no need to accept a code change solely to clear the comment.

[Nyvo-io]

Addressed the latest review in d73dc7c2:

• the first trace-write failure is now latched in memory, header and sentinel writes are attempted independently, and the terminal header carries the latch;
• autonomous retries across different sessions are accepted while run/turn identity conflicts still fail closed, with exact expectedIdentities validation available;
• the public contract now explicitly covers exported top-level Harbor invocations only, so semantic compaction and child-agent dispatches are intentionally outside this artifact;
• the two P3 ownership cleanups are left for the broader Runtime refactor instead of adding more duplication in this PR.

Focused tests, full Runtime/Headless suites, typecheck, lint, format, diff checks, and independent reviews pass. GitHub typecheck, test, and e2e are also green.

[Astro-Han]

Reviewed d73dc7c2; merged as b308282d1.

The last commit fixes trace-write error latching and multi-session identity handling. The scope for compaction and child-agent calls is now explicit instead of being counted as complete provider accounting. I tested the merge with the then-current main; 544 focused tests passed. No new P0–P3.

[Nyvo-io]

Summary

• version provider-request captures as schema v2 while preserving legacy v1 reads
• hash protocol-independent request semantics while preserving normalized reasoning effort
• parse capture and physical-attempt rows and fail closed on malformed, torn, missing, or mismatched evidence
• validate exact expected identity sets across continuation runs and autonomous retries, including retries that use different sessions
• preserve the first trace-write failure in memory, attempt header and sentinel persistence independently, and carry the latch into the terminal header

The exported artifact is intentionally scoped to top-level Harbor invocations. Semantic-compaction and child-agent provider dispatches are outside this contract; the Runtime normalization and attempt-codec ownership cleanups are deferred to the broader Runtime refactor rather than duplicated here.

This is the second flat split requested in #1451 and addresses the request-trace evidence findings on that PR.

Refs #1269

Verification

• provider-request-trace: 23 passed
• SessionManager: 235 passed
• terminal ledger: 26 passed
• Runtime full suite: 2,558 passed, 7 skipped
• Headless full suite: 1,365 passed, 2 skipped
• runtime and headless typecheck passed
• lint: 2,160 files passed
• format check: 1,116 files passed
• git diff check passed
• local review-agent and independent Codex diff review: no findings

Not run: the paid Kimi/Harbor benchmark. That belongs to the separate stacked benchmark PR.

[Astro-Han]

Requesting changes for the P1 trace-completeness issue noted inline.

Non-blocking suggestion: the attempt parser requires integer timestamps, but AiSdkBackendInput.now, ProviderRequestAttemptRecord, and the AgentRun codec allow any finite number. A fractional monotonic clock produces a valid attempt that the reader rejects as invalid_attempt. Either accept finite non-negative values or normalize the timestamps at the Runtime producer. Keeping the decoder next to the Runtime record type would also prevent the two contracts from drifting again.

[Nyvo-io]

All requested changes are addressed and the inline thread is resolved. d62b136 fails closed on trace write/corruption sentinels and accepts the Runtime finite timestamp contract; 3e2c222 normalizes equivalent cross-protocol request evidence; 370c437 validates same-session continuation executions. The focused regressions, full workspace suite, and GitHub typecheck/test/e2e checks all pass. @Astro-Han, please re-review when convenient.

[Nyvo-io]

Latest P1 is addressed in af6079f and ab70f7f, with the requested real FileAgentRunStore export/read/assert regression chain plus torn-tail fail-closed coverage. The branch is rebased onto current main; npm test, typecheck, lint, format:check, diff --check, and the completed independent review all pass. @Astro-Han, please re-review when convenient.

[Astro-Han]

The earlier trace-latch and torn-tail failures are fixed. I found two other ways the trace can still pass when the experiment is not comparable or complete; details inline.

[Nyvo-io]

All four inline findings are addressed and resolved at c1917e8f:

• trace-write header latch and torn-tail evidence are preserved through the real FileAgentRunStore export path;
• protocol-independent hashes retain normalized reasoning effort and reject max/low/none mismatches;
• missing evidence for an earlier continuation invocation now emits an identity-bearing diagnostic and fails closed.

The branch is clean, rebased onto current main, and the focused/full tests plus typecheck, lint, format, diff checks, and independent review pass. Please re-review the current head when convenient.

[Astro-Han]

I did not find a P0 or P1 issue, so I am approving this PR. The earlier trace-integrity findings are fixed. I found these follow-ups:

• [P2] A failed attempt append can still be missed if the first header-latch write also fails and later terminal writes recover.
• [P2] Autonomous attempts from different sessions are exported into one file, but the reader treats the second session as mixed_identity.
• [P2] Semantic-compaction provider calls bypass the request tracker.
• [P2] Child-agent runs are outside the exported run set, so their provider calls are absent without a missing-evidence diagnostic.
• [P3] Request-shape normalization repeats provider-option and output-limit rules owned by Runtime.
• [P3] Headless still maintains a separate decoder for Runtime-owned attempt records.

These affect compound I/O failures, optional execution modes, or benchmark evidence rather than production state.

Please push back if provider-request traces intentionally exclude compaction or child-agent dispatches. If that is the intended boundary, narrowing the public contract may be enough. There is no need to accept a code change solely to clear the comment.

[Nyvo-io]

Addressed the latest review in d73dc7c2:

• the first trace-write failure is now latched in memory, header and sentinel writes are attempted independently, and the terminal header carries the latch;
• autonomous retries across different sessions are accepted while run/turn identity conflicts still fail closed, with exact expectedIdentities validation available;
• the public contract now explicitly covers exported top-level Harbor invocations only, so semantic compaction and child-agent dispatches are intentionally outside this artifact;
• the two P3 ownership cleanups are left for the broader Runtime refactor instead of adding more duplication in this PR.

Focused tests, full Runtime/Headless suites, typecheck, lint, format, diff checks, and independent reviews pass. GitHub typecheck, test, and e2e are also green.

[Astro-Han]

Reviewed d73dc7c2; merged as b308282d1.

The last commit fixes trace-write error latching and multi-session identity handling. The scope for compaction and child-agent calls is now explicit instead of being counted as complete provider accounting. I tested the merge with the then-current main; 544 focused tests passed. No new P0–P3.

[Nyvo-io]

Summary

• version provider-request captures as schema v2 while preserving legacy v1 reads
• hash protocol-independent request semantics while preserving normalized reasoning effort
• parse capture and physical-attempt rows and fail closed on malformed, torn, missing, or mismatched evidence
• validate exact expected identity sets across continuation runs and autonomous retries, including retries that use different sessions
• preserve the first trace-write failure in memory, attempt header and sentinel persistence independently, and carry the latch into the terminal header

The exported artifact is intentionally scoped to top-level Harbor invocations. Semantic-compaction and child-agent provider dispatches are outside this contract; the Runtime normalization and attempt-codec ownership cleanups are deferred to the broader Runtime refactor rather than duplicated here.

This is the second flat split requested in #1451 and addresses the request-trace evidence findings on that PR.

Refs #1269

Verification

• provider-request-trace: 23 passed
• SessionManager: 235 passed
• terminal ledger: 26 passed
• Runtime full suite: 2,558 passed, 7 skipped
• Headless full suite: 1,365 passed, 2 skipped
• runtime and headless typecheck passed
• lint: 2,160 files passed
• format check: 1,116 files passed
• git diff check passed
• local review-agent and independent Codex diff review: no findings

Not run: the paid Kimi/Harbor benchmark. That belongs to the separate stacked benchmark PR.

[Astro-Han]

Requesting changes for the P1 trace-completeness issue noted inline.

Non-blocking suggestion: the attempt parser requires integer timestamps, but AiSdkBackendInput.now, ProviderRequestAttemptRecord, and the AgentRun codec allow any finite number. A fractional monotonic clock produces a valid attempt that the reader rejects as invalid_attempt. Either accept finite non-negative values or normalize the timestamps at the Runtime producer. Keeping the decoder next to the Runtime record type would also prevent the two contracts from drifting again.

[Nyvo-io]

All requested changes are addressed and the inline thread is resolved. d62b136 fails closed on trace write/corruption sentinels and accepts the Runtime finite timestamp contract; 3e2c222 normalizes equivalent cross-protocol request evidence; 370c437 validates same-session continuation executions. The focused regressions, full workspace suite, and GitHub typecheck/test/e2e checks all pass. @Astro-Han, please re-review when convenient.

[Nyvo-io]

Latest P1 is addressed in af6079f and ab70f7f, with the requested real FileAgentRunStore export/read/assert regression chain plus torn-tail fail-closed coverage. The branch is rebased onto current main; npm test, typecheck, lint, format:check, diff --check, and the completed independent review all pass. @Astro-Han, please re-review when convenient.

[Astro-Han]

The earlier trace-latch and torn-tail failures are fixed. I found two other ways the trace can still pass when the experiment is not comparable or complete; details inline.

[Nyvo-io]

All four inline findings are addressed and resolved at c1917e8f:

• trace-write header latch and torn-tail evidence are preserved through the real FileAgentRunStore export path;
• protocol-independent hashes retain normalized reasoning effort and reject max/low/none mismatches;
• missing evidence for an earlier continuation invocation now emits an identity-bearing diagnostic and fails closed.

The branch is clean, rebased onto current main, and the focused/full tests plus typecheck, lint, format, diff checks, and independent review pass. Please re-review the current head when convenient.

[Astro-Han]

I did not find a P0 or P1 issue, so I am approving this PR. The earlier trace-integrity findings are fixed. I found these follow-ups:

• [P2] A failed attempt append can still be missed if the first header-latch write also fails and later terminal writes recover.
• [P2] Autonomous attempts from different sessions are exported into one file, but the reader treats the second session as mixed_identity.
• [P2] Semantic-compaction provider calls bypass the request tracker.
• [P2] Child-agent runs are outside the exported run set, so their provider calls are absent without a missing-evidence diagnostic.
• [P3] Request-shape normalization repeats provider-option and output-limit rules owned by Runtime.
• [P3] Headless still maintains a separate decoder for Runtime-owned attempt records.

These affect compound I/O failures, optional execution modes, or benchmark evidence rather than production state.

Please push back if provider-request traces intentionally exclude compaction or child-agent dispatches. If that is the intended boundary, narrowing the public contract may be enough. There is no need to accept a code change solely to clear the comment.

[Nyvo-io]

Addressed the latest review in d73dc7c2:

• the first trace-write failure is now latched in memory, header and sentinel writes are attempted independently, and the terminal header carries the latch;
• autonomous retries across different sessions are accepted while run/turn identity conflicts still fail closed, with exact expectedIdentities validation available;
• the public contract now explicitly covers exported top-level Harbor invocations only, so semantic compaction and child-agent dispatches are intentionally outside this artifact;
• the two P3 ownership cleanups are left for the broader Runtime refactor instead of adding more duplication in this PR.

Focused tests, full Runtime/Headless suites, typecheck, lint, format, diff checks, and independent reviews pass. GitHub typecheck, test, and e2e are also green.

[Astro-Han]

Reviewed d73dc7c2; merged as b308282d1.

The last commit fixes trace-write error latching and multi-session identity handling. The scope for compaction and child-agent calls is now explicit instead of being counted as complete provider accounting. I tested the merge with the then-current main; 544 focused tests passed. No new P0–P3.