fix(session): reconcile before-agent-start ordering after default-model admission redesign

[Yeachan-Heo]

Exact regression

Dev CI run 31736419453 at 96e8f23ea808844ea27a4dc59468bc9f111dc4fc, shard-2 job 94572361145 fails packages/coding-agent/test/agent-session-before-agent-start-attribution.test.ts after merged PR #4496 / commit e4f4281ac80dc2d7deb0ae31cbbe4e5570760308 changed setDefaultModelSelection admission ordering.

Exact failures:

1. :224 — orders an accepted prompt before a later default selection through provider start: expected selection credential probe not to overtake active prompt preflight, but it now does (Expected false, Received true).
2. orders an accepted default selection before later prompt preflight: DefaultModelSelectionRecoveryError after durable selection.
3. :336 — disposal drains an active selection: queued prompt rejects with same-session AgentBusyError during dispose and escapes as unhandled error.

The same file was unchanged after #4496; the production diff explicitly moved credential probe and waitForIdle outside the selection admission. #4496 focused suites did not include this cross-contract test. This is distinct from #4518 auth-storage fixture drift and #4501 ACP cleanup classification.

Acceptance

• Reconcile the canonical contract between prompt preflight attribution, default-model selection, and dispose.
• Preserve fix(session): order default-model-selection admission before waitForIdle (#4471) #4496's auto-compaction deadlock fix; do not put waitForIdle back under a lease that blocks continuation prompt admission.
• Preserve fail-fast reentrancy and missing-credential semantics where compatible with causal admission ordering.
• Credential probing/durable mutation must not violate accepted prompt ordering or mutate a successor session.
• Default-selection-first then prompt, prompt-first then selection, same-session reentrancy, abort, and dispose all settle deterministically without unhandled rejection or recovery error.
• Run the exact failing file, fix(session): order default-model-selection admission before waitForIdle (#4471) #4496 default-model-selection suite, auto-compaction continuation suite, SDK host wiring, and exact shard-2.
• One focused PR to current dev, independent exact-head adversarial review, full CI, merge, closure, and postmerge Dev CI.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[Yeachan-Heo]

Active implementation ownership is resumed in the dedicated #4519 lane. The preserved patch currently touches only agent-session.ts, the default-model-selection regression suite, and the auto-compaction continuation suite.

The lane was based at 96e8f23ea808844ea27a4dc59468bc9f111dc4fc; exact current dev is 59ac14e7bd5af05d6ce38dfa80b0b2b2100051e3. No PR exists yet. The dirty patch is being preserved and reconciled onto current dev before focused verification. Required closure remains: original before-agent-start failures, #4496 default-model suite, auto-compaction continuation, SDK host wiring, exact shard 2, package checks/types, one focused PR, independent exact-head review, terminal CI, dev merge, issue closure, and postmerge verification. No timeout/sleep inflation or expectation weakening is allowed.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[Yeachan-Heo]

Signed evidence — issue #4519 owner lane

• Original regression evidence read exactly: Dev CI run 31736419453, shard-2 job 94572361145, head 96e8f23ea808844ea27a4dc59468bc9f111dc4fc.
• Focused PR: fix(session): preserve causal model selection order #4524.
• Exact current PR head: fc08b48884b05a4f29804a37044f970f1808641a.
• Exact current base: 38f3b4077c870e41d061f8f2aa7f5c873a30e25f.
• Ancestry proof: merge-base equals exact base.
• Focused suites: 181 pass, 1 existing skip, 0 fail.
• coding-agent check/types: pass.
• isolated production SDK host: pass.
• exact shard-2 reached only independent login-preset-recommendation.test.ts auth fixture failure (authStorage.listCredentialInventory is not a function), outside this issue's explicitly separated ci(auth): close multi-account SDK inventory and ACP usage fixture regressions #4518 contract.

The implementation preserves #4496: waitForIdle remains outside session admission and inherited continuation prompt admission bypasses only the external selection fence. Same-session hook reentrancy remains fail-fast. Prompt-first, selection-first, cancellation/successor, pending continuation, and disposal contracts settle deterministically.

Signed: Bellman / issue-4519 owner

[probepark]

Reproduction on the current dev tip, from PR-review triage of #4523 (this was the residual shard-2 red after #4523 merged, and it is not caused by #4523).

dev @ 4906d0e29 ("Merge pull request #4523"), detached worktree, fresh bun install --frozen-lockfile, rebuilt natives:

$ bun test packages/coding-agent/test/agent-session-before-agent-start-attribution.test.ts
(fail) AgentSession before_agent_start attribution fallback > orders an accepted prompt before a later default selection through provider start [17.90ms]
(fail) AgentSession before_agent_start attribution fallback > orders an accepted default selection before later prompt preflight [73.13ms]
(fail) AgentSession before_agent_start attribution fallback > disposal drains an active selection [15.78ms]
 6 pass
 3 fail

Same 6/3 split on 65130ce03 (#4523 head) and on 099ca7ec1 (its base), so the regression predates #4523 entirely.

Failure signature for the third one:

AgentBusyError: Agent session admission is busy due to same-session reentrancy.
 code: "busy"
      at #sessionAdmissionBusyError (packages/coding-agent/src/session/agent-session.ts:2548:24)
      at #withSessionAdmission (packages/coding-agent/src/session/agent-session.ts:2641:68)
      at async <anonymous> (packages/coding-agent/test/agent-session-before-agent-start-attribution.test.ts:336:9)

So disposal drains an active selection is hitting the reentrancy fast-fail in #withSessionAdmission instead of draining, which matches the acceptance bullet about dispose and reentrancy needing to be reconciled rather than treated independently.

CI evidence for the same three tests on a PR run: job 94656764773 (Affected path validation / test:@gajae-code/coding-agent:shard-2-of-8), run 31763710478.

Worth noting for anyone triaging red shards: shard-2's failure list is currently dominated by this file, so it will keep marking unrelated PRs red until #4519 lands.

[Yeachan-Heo]

Terminal evidence for issue #4519:

• PR fix(session): preserve causal model selection order #4524 merged to dev at merge commit 34f5db331b484726c447f6ab026f35c8b391a879 from exact independently approved head 3baf82e3ac7ce0143b80e51e934ffd128cb7a982 over base 4906d0e29c18d137d15c19af11a2bb0864f017ad.
• Merge reachability verified: the reviewed head is an ancestor of canonical origin/dev.
• Independent authenticated exact-head approval: probepark review 4933522086 — fix(session): preserve causal model selection order #4524 (review)
• Authoritative normal pull-request CI run 31767310776: terminal success, including bootstrap, native build, focused attribution/default-selection/auto-compaction tests, coding-agent check/types, evidence producer, and aggregate.
• Authoritative postmerge Dev CI run 31768516300 at exact merge commit: terminal success with no failed jobs.
• Postmerge shard 2 job 94670381939: success — https://github.com/Yeachan-Heo/gajae-code/actions/runs/31768516300/job/94670381939
• Postmerge evidence producer job 94672516322: success — https://github.com/Yeachan-Heo/gajae-code/actions/runs/31768516300/job/94672516322
• Postmerge aggregate job 94672547841: success — https://github.com/Yeachan-Heo/gajae-code/actions/runs/31768516300/job/94672547841
• Canonical merged worktree verification: attribution/default-selection/auto-compaction suites 79 pass + 1 existing skip, SDK host wiring 104/104 in isolated execution, coding-agent check/types pass.
• Canonical merged local build: bun run build completed successfully, including native addon and compiled dist/gjc.
• Worktree is clean at canonical dev; no postmerge source push was made because no authoritative postmerge regression exists.

The original run 31736419453, shard-2 job 94572361145, and the later current-dev reproductions are resolved. Prompt-first and selection-first causal ordering, same-session fail-fast reentrancy, inherited continuation admission without holding selection across waitForIdle, successor/cancellation behavior, and deterministic disposal settlement are all covered and green.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]