[JacobStephens2]

Summary

agy, Muse and OpenCode now supply model/tool lifecycles, messages, usage and failures through the shared normalized Event interface from #983. Retained Muse records and OpenCode exports enrich the same completion timeline without counting a call twice.

Interpretation
├── decode → progress + shared Events; capture completion timeline
├── condense → existing progress, without event buffering
└── finish → reconcile retained Muse records / OpenCode export by identity

Message snapshots carry an optional identity separate from their causal parent. Existing Claude/Grok/Codex message JSON stays unchanged when that identity is absent.

Closes #984

Evidence

• Before: all six Spec reproduction commands failed exactly as reported: retained Muse task/terminal facts were missing, OpenCode text parts were lost, agy repeated a reply, Claude child replies collapsed, and Muse terminal fallback text disappeared.
  After: all six exact commands pass; all six regression tests remain. The reconciled harness tests pass (166 passed, two intentional skips). The final Muse helper extraction also passes its 11 focused tests.
• The agy/Muse/OpenCode session suites passed all 35 tests after upstream reconciliation. Formatting and clippy pass on the published head. CI run 39706539172 passed the full Linux Rust suite, site browser tests, Console browser tests and generation checks for exact head 3061557f2b81e44300edf20a65c335f1283973a3; packaged artifact builds also passed.
• Final local selection: NEXTEST_TEST_THREADS=2 /usr/bin/time -p scripts/affected-tests origin/issue-700 selected 2,265 Rust tests. It was stopped by its own session after 781.06 s, once equivalent exact-head CI passed, following docs/testing.md; this is an interrupted local run, not a local full-suite pass. Local compilation took 14.46 s. CI's successful cargo test step took 657 s, including 36.60 s compilation (execution time, not queue delay).
• Three existing tests remain intentionally ignored: authenticated agy calls, the optional short-capture latency probe, and the upstream Codex boundary test covered by Decide Codex per-request evidence for Session traces #1047. macOS full-suite jobs are skipped for this Spec-branch target under CI policy.
• Tested seams: live decode and progress-only condense; completion reconciliation through Muse transaction files; OpenCode standalone export acquisition and fallback; shared message identity across child scopes.

Review coverage

Independent Standards and Spec reviews used issue-700 pinned at dd4b988392ac5f30fb159b524ceff883f6b4906c, including the #982/#983 changes that landed during implementation. Both axes read all 20 changed files. No generated exclusions or unread changed files. Reports with final files-read lists are saved in .thirdshift-review-dMqXYE/standards.md and .thirdshift-review-dMqXYE/spec.md in the worktree.

Standards: three P3 duplication findings addressed. Spec: six reproduced behavior findings fixed, with their tests retained. The final reply-ID helper extraction was checked locally and by the Spec reviewer after the Standards follow-up.

Unaddressed findings

Standards

None. All three heuristic findings were addressed.

Spec

None. All six findings were reproduced and fixed; no decision was deferred.

Merge Danger

Door: two-way

Blast Radius: Sessions

Existing progress, Model evidence, summaries, diagnostics, Resume facts and final-message behavior remain covered by the existing tests. Normalized timelines add retained-record facts and message identity; trace writing and Console adoption remain later work.

Built with codex · gpt-6.1-sol · xhigh