fix(proxy): persist complete HTTP bridge replay transcripts

[shaqman]

Closes #1898

Summary

This PR makes HTTP Responses bridge recovery replayable when upstream continuity is lost, while keeping the recovery fail-closed when the transcript is incomplete or ambiguous.

The branch is rebased onto main at b311aea760aa639fd96f63bd118f775e9b4a89f9 (currently v1.24.0-beta.4).

Changes

• Persist root operations in the durable transcript.
• Materialize terminal response output from the event spool before marking a turn replayable.
• Persist bounded, account-neutral replay snapshots only after output and replay input are complete.
• Match continuations while tolerating upstream-omitted reasoning/hosted-search envelopes.
• Deduplicate echoed tool output.
• Track synthetic snapshot roots that already contain terminal output so a tool-output-only continuation does not duplicate the preceding function call.
• Preserve fail-closed behavior when parent identity or transcript completeness cannot be proven.
• Add unit coverage and OpenSpec updates.

No upstream API or Codex CLI changes are required.

Validation

Local:

• pytest -q tests/unit/test_complete_transcript.py: 10 passed
• Ruff check and format check on changed Python files: passed
• Ty check on changed production files: passed
• python -m compileall and git diff --check: passed

The full bridge unit module was also attempted. The local test container currently lacks the git executable required by an unrelated worktree-path test, so it stops at that environment failure rather than indicating a regression in this change.

Live deployment evidence:

• Affinity-aware tool continuation completed across a reconstructed response chain.
• Root and child operations both recorded response_output_items_complete=true and response_replay_input_complete=true.
• The child replay snapshot contained exactly one function_call and one function_call_output; no duplicate tool call was emitted.
• Since deployment: 1,401 HTTP successes and no 502/503/cooldown/retry-circuit responses.
• Remaining failures are predominantly upstream WebSocket clean closes before response.completed (27 HTTP stream_incomplete responses); seven stale sessions had no durable replay snapshot and correctly failed closed.

Related work

This is intentionally narrower than the existing recovery work:

• bug: hard HTTP bridge sessions stall after repeated response.create acknowledgement timeouts #1655, fix(proxy): recover anchored HTTP bridge turns after silent response.create #1735, and HTTP bridge can loop or time out after silent and clean upstream closes #1393 cover hard acknowledgement-timeout and clean-close recovery.
• fix(proxy): durably recover hard HTTP bridge operations #1657 is the merged durable hard-bridge recovery foundation.
• fix(proxy): recover stale previous response anchors #1863 covers stale previous_response_id recovery.
• feat(proxy): fall back to HTTP transport when the upstream websocket is unavailable #1886 and feat(proxy): fall back to raw HTTP on replay-safe bridge cooldown suppressions #1890 cover HTTP transport/cooldown fallbacks.

Those changes do not provide complete transcript materialization plus tool-output deduplication for the replay path addressed here.

Rebase follow-up

The release branch advanced after the initial PR creation. The branch was force-with-lease updated after a clean rebase onto current main at b311aea760aa639fd96f63bd118f775e9b4a89f9 (currently v1.24.0-beta.4); no conflicts occurred. The focused replay suite and static checks were rerun successfully after the rebase (10/10 tests passed).

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: f3c13db9-a40a-4c41-ab65-3c9574cc2d40

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[shaqman]

This PR was closed automatically when the temporary release/beta-1.24.0-beta.4 base branch was deleted. The implementation has since been cleanly rebased onto upstream main at b311aea7 and recreated as #1900: #1900