Triage: this is an O5 (#178 / epic #173) diagnostic-truth prerequisite, separate from test-only #187. Exact reproduction implementation is green main f38d6c7ba52b058a4d2e243e50e4ca4f732a6829, using the existing entrypoint cliRoute seam and a zero-network counting transport. The automatically attached version 0.3.4 / commit 64485cdad2b52830d8c6590053e695aec8377cad identifies the feedback reporter binary, not the source SHA under test.

Both already-cancelled and past-deadline dispatch contexts return exit 1, zero RoundTrip calls, failure_class live_transport_failure, http_attempted true. This differs from cancelling a CLI startup configuration read, which returns configuration_error / http_attempted false. Scope must retain both controls and an actual transport-entry positive control; do not conflate the two cancellation boundaries.

Duplicate checks after bounded live issue refresh: exact http_attempted search and cancellation search found no matching open report. #1 is an old opposite (false-negative) diagnostic report; #86 concerns cache-lock privacy; #187 removes scheduling dependence in a timeout test. The new issue is a false-positive actual-attempt diagnostic. No production fix has yet been applied.

Independent author recheck while #188 exact-SHA review proceeds: reproduced this defect deterministically on an isolated archive of unchanged main f38d6c7ba52b058a4d2e243e50e4ca4f732a6829.

go test -p 1 ./cmd/gitcode-mcp -run TestIssue189PreDispatchCancellationProbe -count=3 is expected-red in both already-cancelled and already-expired route contexts: exit 1, exactly zero fake RoundTrip invocations, but public output claims http_attempted: true. The fake transport cannot contact the network. Startup/configuration was deliberately allowed to complete before injecting cancellation through the existing cliRoute seam. No production fix is applied yet and the reviewed #188 branch is unchanged.

Trace confirms two different inference bugs: CLI promotes a PartialSyncError cancellation/timeout diagnostic directly to attempted HTTP, and CLI/MCP treat ErrNetworkUnavailable as attempted without checking its Attempts=0 evidence. A fix must also preserve true evidence for actual transport entry and cancellation after successful/partial provider progress; simply deleting the inference would create false negatives. Scope/design and transport-parity decisions will be recorded before implementation.

Implementation design and verification plan

Scope: truthful HTTP-attempt evidence for explicit sync cancellation/deadline diagnostics and zero-attempt typed network errors, shared by CLI/MCP and retained service results. No new command, provider endpoint, retry policy, setup action or durable schema.

Contract:

• A bounded operation-local primitive witness is set only at actual HTTP RoundTrip entry; it contains no URL, identity, header, body or error string and does not depend on logging/metrics being enabled.
• Sync aggregates preserve the witness across cancellation after successful/list-only/partial progress. Cancellation before provider entry remains explicitly no-attempt; do not infer traffic from a timeout label, result count or selected live mode.
• Typed network attempt counts must not count context/rate-limit waits or request-construction failures as HTTP attempts.
• The shared diagnostic classifier consumes known evidence after transport-specific legacy inference, so CLI and MCP cannot disagree. Pre-provider operation cancellation/deadline is distinguished from transport failure; startup configuration-read cancellation retains its separate configuration diagnostic.
• Existing error unwrap/PartialSyncError handling, partial-success output, retries, telemetry-disabled behavior and provider traffic bounds must remain intact.

UX: this corrects existing attempt/failure context rather than introducing a new resource or action. CLI text/JSON and MCP error/result projections get contract tests. Admin consumes existing job failure/progress fields rather than this CLI operation envelope; no separate UI control is required. #178 still owns integrated CLI/MCP/Admin incident parity and will not infer job execution failure from an observation-reader cancellation.

Regression matrix: cancelled and expired before dispatch with zero counting-transport calls; actual transport-entry timeout/cancellation positive controls; cancellation after a successful empty/list-only response and after partial progress; request-construction/rate-limit-wait non-attempts; current startup configuration controls; multiple collection aggregation; metric sink disabled; no leaked private text. Use restored serial fakes, no real DNS/credentials or timing races. Then focused race tests, full Go/vet/docs/diff, independent exact-SHA review, branch CI, PR discussion gate and resulting-main CI before closure.

Risk: losing real earlier-attempt evidence during error/result aggregation, accidentally making diagnostics depend on telemetry, or altering http.Client timeout/redirect/transport behavior. Keep the witness primitive and request-local, retain existing transport behavior, and test negative plus positive boundaries. Original issue input remains fixed.

Implementation checkpoint: the deterministic zero-transport cancellation regression now reports http_attempted=false. Focused diagnostics, adapter, shared service, CLI and MCP packages passed. The entrypoint suite exposed the already-tracked #187 dependency: its old 1 ms timeout case expects a transport failure even when startup consumes the deadline before transport entry. This is not a weakened gate; #189 will be stacked on the independently reviewed #187 commit 8f6a387062944410b9c1acda4a96dfd8c6a7908c, retaining its deterministic entered-transport positive controls and separate startup cancellation controls. #187 / !171 must be merged before #189 targets main. No merge or completion is claimed.

Implementation/UX clarification: operation interruption and actual HTTP entry are independent dimensions. Before dispatch, cancellation/deadline reports operation_cancelled/operation_timeout with http_attempted=false. An entered transport failure remains live_transport_failure with true. Cancellation later in local sync/cache work reports the operation interruption with true; an earlier empty list must not disappear. The witness is one operation-local atomic bit, frozen on error return, shared by nested calls and multi-collection CLI/MCP actions, with no telemetry dependency or request/coordinate retention. Partial aggregation preserves the first interruption and positive evidence. CLI compact partial summaries and MCP lifecycle/operational diagnostics expose the known bit; existing MCP lifecycle codes are retained with canonical error_class. Admin remains on the existing job/observation DTOs; no new UI action/resource is introduced, and integrated Admin/CLI/MCP rollout remains #178. Focused regressions passed repeatedly, full go test -p 1 ./... -count=1 passed; touched-package race verification is still running. #187 merge remains a dependency, not waived.

Clean implementation commit: 11e6bebc204b3ad578229e958ba8f55f81a50037 on codex/issue-189-truthful-http-attempt, stacked base 8f6a387062944410b9c1acda4a96dfd8c6a7908c (#187 / !171). Final local gates passed: go test -p 1 ./... -count=1; go test -race -p 1 ./internal/diagnostics ./internal/cli ./internal/mcp ./cmd/gitcode-mcp -count=1; go test -race -p 1 ./internal/service -count=1 (82.551s); go test -race -p 1 ./internal/gitcode -count=1 (5.082s); go vet -p 1 ./internal/diagnostics ./internal/gitcode ./internal/service ./internal/cli ./internal/mcp ./cmd/gitcode-mcp; ./scripts/check-product-docs.sh; git diff --check. Focused deterministic regressions also passed repeatedly. No Admin UI source or assets changed; integrated UI rollout remains #178. An independent reviewer is inspecting the exact commit and authoritative issue record; review and remote exact-SHA CI are pending, so this is not completion. Additional boundary coverage preserves http.Client redirect/timeout/legacy CancelRequest behavior, public-safe contention error projection, and prior provider/schema failure classification when a later local interruption occurs. No live provider compatibility mutation is required for this diagnostics fix, and no release/install action was run.

Exact-head CI passed for 11e6bebc204b3ad578229e958ba8f55f81a50037: GitHub run 38151755920, attempt 1, Go test job 114506478863, push on codex/issue-189-truthful-http-attempt. All required steps succeeded, including embedded assets, frontend checks/unit tests, semantic browser tests, Go tests, release automation tests and whitespace. Independent exact-SHA review remains in progress; a later pre-dispatch interruption after earlier successful HTTP is under investigation. This is CI evidence, not review clearance or merge eligibility. Existing #187 dependency and discussion-evidence gates remain. Run: https://github.com/urandon/gitcode-mcp/actions/runs/38151755920

Independent review of 11e6bebc204b3ad578229e958ba8f55f81a50037 found one confirmed P2: after an earlier successful request, cancellation in a later request's local rate-limit wait was still classified live_transport_failure with http_attempted=true. Actual adapter and shared-service multi-collection probes reproduced it three times; the later request entered transport zero times, while the operation correctly retained one earlier HTTP attempt. Accepted and fixing: separate failed-request transport evidence from the enclosing operation receipt, preserving confirmed provider/schema failures and actually entered failed requests in mixed aggregates. Added deterministic adapter, shared-service, CLI/MCP and mixed-cause controls; author adapter/CLI/MCP regressions were RED x3 before the fix and pass after it. Full local and race gates are being rerun; the old green CI does not qualify the revised patch. Review and exact-head CI will be repeated before PR/merge. Original issue input is unchanged.

Corrected exact SHA: 3b3d758c625759aa74254821d1d7fbd2ac412f09 (stacked on #187's 8f6a387062944410b9c1acda4a96dfd8c6a7908c). The sole independently reported P2 is addressed with request-local failed-transport evidence separate from the operation-wide HTTP receipt. Confirmed provider/schema failures retain precedence; an entered failed request wins in mixed failures, and an unknown branch is not treated as known negative.

Final local gates passed: go test -p 1 ./... -count=1; go test -race -p 1 ./internal/diagnostics ./internal/gitcode ./internal/service ./internal/cli ./internal/mcp ./cmd/gitcode-mcp -count=1; focused actual-boundary/mixed-cause regression controls three times under race across diagnostics, adapter, service, CLI and MCP; affected go vet -p 1; ./scripts/check-product-docs.sh; git diff --check. Regression coverage includes actual issues success followed by PR rate-wait cancellation, alongside cancel/deadline and entered-failure positive controls. No Admin source/assets changed; remote CI will repeat the full frontend gates. Independent review of this new exact SHA and remote CI are pending. The prior SHA's green CI is not substituted. No merge, issue closure, release, install or installed-service changes.

Re-review of 3b3d758c625759aa74254821d1d7fbd2ac412f09 found a residual P2: a later target skipped by SyncResources' context guard retained only raw context cancellation, which looked like an unknown failing-request boundary in a mixed aggregate. Independent actual single-batch reproduction: first target succeeds with two HTTP requests; second target cancels locally before transport; third target is skipped; result retains success=1/failure=2 but wrongly classifies transport failure. Accepted. The explicit owner context guard now retains typed skipped-before-dispatch evidence, preserving genuinely unknown branches as unknown rather than changing conservative aggregation.

Author actual shared-action target regression was RED x3 on prior production after valid setup; a failed first fixture precondition is not counted as a product reproduction. Final actual shared-action and single-batch controls, CLI/MCP mixed projections and strict sentinel/unknown controls are being checked. Independent original probes passed on 3b3d758; its CI run 38152690057 attempt 1 also passed, but this rejected SHA is not merge-eligible. New exact-SHA local gates, review and CI must complete before delivery.

New corrected exact SHA: b78df0af89cf8bce502298f4c29eb41d2a984392, stacked base #187/8f6a387062944410b9c1acda4a96dfd8c6a7908c. Both independently identified P2 cases are addressed. Explicit owner guards now retain typed skipped-before-dispatch evidence; this is never inferred from unknown or wrapped errors and is not applied to provider-returned cancellation. The actual single batch preserves its first successful target (two HTTP requests), second locally cancelled target, and third skipped target; operation cancellation retains overall http_attempted=true.

Passed on final content: go test -p 1 ./... -count=1; full go test -race -p 1 ./internal/diagnostics ./internal/gitcode ./internal/service ./internal/cli ./internal/mcp ./cmd/gitcode-mcp -count=1; focused actual-boundary/mixed-cause/skipped-target controls x3 under race; affected go vet -p 1; ./scripts/check-product-docs.sh; git diff --check. CLI/MCP partial projections cover cancelled/deadline local tails and entered-failure positive controls. New exact-SHA independent review and remote CI are running. Prior-head CI is not reused as final evidence; no PR merge or closure is claimed.

Additional local observation-cohort checks passed on b78df0af89cf8bce502298f4c29eb41d2a984392 without source edits: CI=true npm run check (0 errors/warnings), CI=true npm run test (18 unit tests), and CI=true npm run test:e2e -- tests/service-observation.spec.ts (19 semantic Chromium scenarios). Browser artifact guard passed; no screenshot, video or trace output. These are existing Service-view tests, not a claim that the full integrated O5 rollout is qualified. Final independent review and exact-head CI remain in progress.

Final corrected head b78df0af89cf8bce502298f4c29eb41d2a984392 has exact-head CI success: GitHub run 38153405184, attempt 1, Go test job 114511339260, push on codex/issue-189-truthful-http-attempt. Independently confirmed all required steps succeeded, including product docs, embedded assets, frontend types/unit tests, semantic browser tests, visual-artifact guard, Go tests, release automation tests and whitespace. Existing configured mirror completed with zero failures and fresh successful readback. Independent review of this exact head is still pending; this does not authorize merge or replace discussion/main integration gates. Run: https://github.com/urandon/gitcode-mcp/actions/runs/38153405184

Independent review of b78df0af89cf8bce502298f4c29eb41d2a984392 found two additional P2 cause-retention defects: bounded wiki cancellation discarded an entered transport failure; bounded issue synchronization could replace it with a subsequent cancelled queue-summary read. Accepted both findings. A real-parent-cancellation regression failed for issues/wiki on the reviewed SHA (three race-enabled repetitions), while PR and successful-read/local-cancellation controls passed. The correction preserves typed causes privately, including collection aggregation, without changing public failure counts or introducing provider-body output. Focused race regressions now pass; actual CLI/MCP parent-cancellation coverage and full gates are being repeated before a new exact-SHA independent review. Previous green CI does not make the rejected SHA merge eligible.

Fourth correction verification: actual parent cancellation is now covered through bounded issues/wiki/PR service owners and CLI/MCP, with successful-empty-read/local-cancellation negative controls. Private cause-retention tests also cover nested aggregates, stable public JSON/counts, and direct/wrapped self-reference avoidance. Focused six-package race regressions passed in three repetitions. Full go test -p 1 ./... -count=1, affected go vet -p 1, product-doc contracts and whitespace checks passed on the final working content. Full six-package race gate is finishing; exact-SHA independent review and CI will be repeated after commit. No delivery or merge claim yet.

Correction committed and pushed: 8cc54f5cf692eb4091b00edc9e42f089e680732d; remote branch verified at the same SHA, stacked base 8f6a387062944410b9c1acda4a96dfd8c6a7908c. Final-content gates passed: go test -p 1 ./... -count=1; go test -race -p 1 ./internal/diagnostics ./internal/gitcode ./internal/service ./internal/cli ./internal/mcp ./cmd/gitcode-mcp -count=1; focused actual-boundary/mixed-cause/aggregate regressions under race, three repetitions; affected go vet -p 1; ./scripts/check-product-docs.sh; git diff --check. Independent exact-SHA re-review requested. Configured CI mirror synchronization finished; exact-head CI watcher dispatched. Required review/CI are pending, not passed. No PR merge, issue closure, release, installation or installed-service change.

Exact-head CI passed for 8cc54f5cf692eb4091b00edc9e42f089e680732d: https://github.com/urandon/gitcode-mcp/actions/runs/38154979277, attempt 1, Go test job 114515986240. Root readback verified the job head SHA and terminal success of all required steps: product docs, embedded assets, UI types/unit tests, semantic browser suite/artifact guard, full Go tests, release automation fixture tests and whitespace. Independent exact-SHA re-review is still running; CI success is not a substitute for that gate.

Independent exact-SHA review CLEAR for 8cc54f5cf692eb4091b00edc9e42f089e680732d: no P0–P2 findings remain in the cumulative #189 patch. Reviewer independently reproduced the previously failing actual-parent-cancellation cases; bounded issues/wiki/PR, actual parent-deadline inside RoundTrip, unbounded issues and successful-empty-read/local-interruption controls passed in race-enabled repetitions. Independent full Go, full affected six-package race, focused regressions, vet/docs/diff gates passed; exact CI run 38154979277/job 114515986240 was independently read back as all-step success. Review was of an immutable archive of this full SHA, not the author's summary. Opening a stacked PR against #187/!171 next. This is a code/head gate only: prerequisite discussion holds, PR verification, intended-main integration/re-review/re-CI, main CI and O5 acceptance remain separate.

PR opened and attached: https://gitcode.com/urandon/gitcode-mcp/pull/173. Bounded canonical provider readback confirms repository-local PR 173 is open, exact head 8cc54f5cf692eb4091b00edc9e42f089e680732d, stacked base codex/issue-187-deterministic-cli-timeout at 8f6a387062944410b9c1acda4a96dfd8c6a7908c, and the exact expected title/body. Fresh targeted PR-comment sync returned zero records; cached unresolved-discussion view returned zero, but those projections alone are not proof of complete provider discussion history. Independent exact-head review is CLEAR and head CI 38154979277 is green. Delivery remains HOLD: #187/!171 and #188/!172 have unresolved discussion-evidence gates. Do not merge this PR into the temporary base. After #187 merges onto green main, retarget/rebase this branch to intended main and repeat exact-SHA review/CI if the SHA changes, then verify discussions and resulting main CI before closure. Clean pushed fallback safepoint; #189 remains open, with no release/install/service mutation.

Integration gate status: !173 remains open, clean and pushed at independently reviewed SHA 8cc54f5cf692eb4091b00edc9e42f089e680732d; exact-head CI 38154979277 passed. It still targets codex/issue-187-deterministic-cli-timeout, not main. #187 has merged and closed; #188 merged into main at 57a877f2553458cc4fb2d59ad187c8bb3b19c498 (resulting-main CI currently pending).

A real tooling gap is now isolated in #190: installed CLI and exposed MCP cannot retarget an existing PR, and the official v5 PATCH contract does not establish a base/target input. UI inspection could not initialize because the executor rejects a symlinked writable-root setup. No sandbox setting was changed; no guessed remote endpoint was called. Current source adapter UpdatePRRequest also has no base field, so this is not merely an old installed-client schema.

Do not merge !173 into its temporary base. Safe continuation requires a proven retarget route, or an explicitly agreed replacement-PR workflow that closes the old PR and preserves the issue/review history. Original issue description remains unchanged. No release or installed-service mutation was performed.

Recovery decision following the target-branch tooling assessment: replacement of the author's own creation-only stacked PR is a normal delivery repair within the authorized autonomous workflow. There are no human discussion/review records to discard; #189 and its original input stay unchanged. Preserve !173 as a closed superseded historical record, reference it from one replacement PR targeting green main, and retain one active PR for the issue. This does not merge into the temporary base or bypass review.

First merge current green main 57a877f2553458cc4fb2d59ad187c8bb3b19c498 into the issue branch, run applicable local gates, obtain independent review for the new exact SHA, then open the replacement with explicit lineage. New head CI and resulting-main CI remain mandatory. The missing in-place retarget capability remains tracked by #190. This normal delivery repair does not require a human attestation that discussion lists are empty.

Integration validation at 8cb4302a509b81df84d66563b00dd6e8f2148cb8 (parents: prior reviewed 8cc54f5cf692eb4091b00edc9e42f089e680732d and green main 57a877f2553458cc4fb2d59ad187c8bb3b19c498):

Passed: go test ./...; go test -race ./internal/gitcode ./internal/service ./internal/cli ./internal/mcp ./internal/cache ./internal/servicectl; go test -race -p 1 ./internal/diagnostics ./cmd/gitcode-mcp -count=1; go vet ./internal/gitcode ./internal/service ./internal/cli ./internal/mcp ./internal/diagnostics ./internal/servicectl; ./scripts/check-product-docs.sh; git diff --check origin/main...HEAD. Branch is clean and pushed without force.

Independent renewed exact-SHA review is running. Reviewer inspection confirms all #189 implementation and regression blobs are unchanged by main integration, with only accepted #188 source-authority changes added from main. Exact-head CI run 38157867498 is pending. Neither old head review nor old CI is being substituted for the new SHA gates. Replacement PR will explicitly target main and retain !173 lineage.

Renewed independent exact-SHA review is CLEAR for 8cb4302a509b81df84d66563b00dd6e8f2148cb8 against main 57a877f2553458cc4fb2d59ad187c8bb3b19c498. No P0-P2 findings. Reviewer independently verified merge parents and unchanged #189 blobs; all retained probes repeated three times under race passed, as did full Go, six-package race, vet/docs/diff. All reviewer process handles are terminal.

Exact-head CI 38157867498 attempt 1, job 114524459633 was independently and author-confirmed successful, all 17 steps. Main base CI 38157369017 is green. Proceeding with reversible state-close of the author's historical !173, preserving its body/history, and one replacement PR to main. Replacement canonical/discussion gates remain separate before any merge.

Replacement !174 is now the active PR targeting main: https://gitcode.com/urandon/gitcode-mcp/pull/174 . Historical !173 was state-closed through audited update only; its body/history remain intact. New creation was audited/provider-confirmed; canonical readback independently from the creation receipt confirms open state, exact reviewed head 8cb4302a509b81df84d66563b00dd6e8f2148cb8, main base 57a877f2553458cc4fb2d59ad187c8bb3b19c498, title and full body.

Code review and exact-head CI are clear as recorded above. Fresh replacement metadata/discussion verification is now with the independent reviewer. The author fresh timeline shows the known creation-event action digest/schema, canonical entity/time association, one terminal system record, terminal empty user cursor, and no inline/reply/resolution/pagination markers. It is compared with earlier controlled empirical evidence, not assumed empty from v5 alone. No user discussion confirmation is required.

No merge into the old temporary base occurred. Resulting-main exact-SHA CI remains mandatory after any merge.

Fresh independent !174 canonical/discussion gate: CLEAR. Reviewer performed bounded live GETs and independently confirmed open state, exact title/full body, reviewed source SHA 8cb4302a509b81df84d66563b00dd6e8f2148cb8 and main base 57a877f2553458cc4fb2d59ad187c8bb3b19c498. V5 comments are empty; the sole v4 row matches the fixed earlier empirical creation-action digest and exact 11-key schema, canonical entity and creation time within five seconds. User cursor is terminal; system cursor identifies that sole lifecycle row; no next-page/has-more, inline/reply/resolution/nested-note markers. This is positive classified lifecycle evidence, not a cache-zero or self-match assumption.

Code review is CLEAR and exact-head CI 38157867498 is green; current main CI 38157369017 is green and fresh refs are unchanged. All pre-merge gates are satisfied; proceeding with the audited expected-head merge. Closure waits for resulting-main CI.

!174 merged through the audited exact-head lifecycle. Provider-confirmed merge SHA 964eeefd2eb5d81965f3f55993a1fce89830863e matches current remote main. The replacement targeted main; the historical !173 remains superseded rather than merged into the temporary branch.

Independent exact-head code review, canonical/discussion review and head CI are clear as recorded above. CI mirroring was explicitly triggered. Resulting-main exact-SHA CI is pending; final closure remains gated on terminal success. No release/install/installed-service mutation occurred.

Delivery complete. Active replacement !174 merged into main at 964eeefd2eb5d81965f3f55993a1fce89830863e; historical !173 is superseded, with its body/history preserved.

Reviewed source SHA: 8cb4302a509b81df84d66563b00dd6e8f2148cb8. Independent renewed code review CLEAR (no P0-P2), full/focused race probes and Go/static/docs gates passed; independent replacement canonical/discussion gate CLEAR. Exact-head CI38157867498 attempt1/job114524459633 passed.

Resulting-main CI38158838412 attempt1/job114527316291 passed for the exact merge SHA, all 17 steps successful: full Go, embedded assets, frontend types/unit tests, semantic browser tests, artifact guard, docs, release automation fixtures and whitespace. Both author and independent CI watcher verified terminal success.

Main CI: https://github.com/urandon/gitcode-mcp/actions/runs/38158838412

#189 acceptance is complete; closure is now justified. #190 records the separate in-place retarget capability gap but replacement recovery resolved this delivery blocker. #178 integrated qualification remains the next epic child; whole-feature rollout is not declared complete. No release, install or installed-service upgrade occurred.