fix(codex): show sub-agent models by rynfar · Pull Request #136 · pylon-code/pylon

rynfar · 2026-08-28T20:32:31Z

Codex sub-agents did not carry their model metadata into task events, so the
Agents panel could not show a child's model or reasoning effort.

Reads model and effort from each spawned child thread, updates them when Codex
reports a settings change or a model reroute, and includes them in later task
events. Unknown values are omitted rather than inherited from the parent, so a
child never claims a model it is not running.

Adopted from T3 Code 49f6241ddeee2a5d20ec181b5b94ec0a9ac72f7b
(pingdotgg/t3code#8502).

Adaptation

Two conflicts, both in the same place conceptually: Pylon and upstream had each
added fields to the child-identity payload.

CodexAdapter.ts — kept Pylon's toolUseId alongside upstream's model
and effort. The linkage object immediately below already spread all three,
so this was a "both added a const" conflict with an unambiguous answer.
CodexSessionRuntime.ts — adopted upstream's new collabChildIdentity
helper and spread Pylon's toolUseId at the call site rather than folding it
into the helper. The helper describes the child; toolUseId names the
child's live turn, and turn/started overrides it with the id the
notification itself carries. Merging the two would have conflated different
lifetimes.

Verification

vp test run across CodexAdapter.test.ts,
CodexCollabRuntime.integration.test.ts, CodexCollabWire.test.ts,
ThreadBackgroundLiveness.test.ts, subagentRuntime.test.ts — 121 passed
vp run -F t3 typecheck and -F @t3tools/client-runtime typecheck — clean
vp lint on both conflicted files — clean
Mutation probes on both halves of the resolution, since silently dropping
one side is the whole risk here:
- removing Pylon's toolUseId spread fails "terminal rows retain the
  activation id for deterministic persisted folding";
- removing upstream's model/effort fails two assertions with
  expected { …(3) } to have deep property 'model'.
So each side is independently guarded, and the merge kept both.

Claude Opus 5 via Claude Code.

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

github-actions · 2026-08-28T20:36:55Z

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.3 KiB	13.3 KiB	+13 B (+0.1%)	15.1 KiB	✅
Codex	Thread snapshot wire	6.9 KiB	6.9 KiB	+2 B (+0.0%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	6.4 KiB	6.4 KiB	+11 B (+0.2%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	55.5 KiB	55.5 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	9	9	0 (0.0%)	21	✅
Claude	Total thread wire	13.3 KiB	13.2 KiB	−161 B (−1.2%)	15.1 KiB	✅
Claude	Thread snapshot wire	6.9 KiB	6.9 KiB	+15 B (+0.2%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	6.5 KiB	6.3 KiB	−176 B (−2.7%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	56.4 KiB	55.5 KiB	−928 B (−1.6%)	66.4 KiB	✅
Claude	Live turn messages	11	9	−2 (−18.2%)	21	✅

Baseline: 3d6035f · PR result: 08820bc · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

Codex decoded thread snapshot: 109.4 KiB
Claude decoded thread snapshot: 110.1 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

rynfar · 2026-08-29T00:52:20Z

Independent review pass. Five findings — one fixed, four are upstream bugs I am
documenting rather than changing blind, since none can be verified without a real
Codex 0.150 app-server.

Fixed in 03db9fab3 — the adopted paragraph in docs/user/providers-codex.md
said "T3 Code leaves it out" in a shipped-product page. My conflict resolution
missed it. (The install.md mentions of T3 Code are deliberate: they describe
Pylon installing beside it.)

thread/resume is sent with a parameter that does not exist.
CodexSessionRuntime.ts:1360 passes excludeTurns: true, and the comment above
it claims that "avoids loading or replaying its history". excludeTurns appears
0 times in the generated schema; includeTurns appears 15, and only on
thread/read. client.raw.request is untyped, so this compiles. Against a real
app-server it either errors — swallowed by Effect.catch, meaning no sub-agent
ever shows a model while the mock-peer test still passes — or is ignored, loading
the child's full rollout history and resuming a possibly mid-turn child. This is
upstream's line, added by 49f6241dd, not a Pylon adaptation. I left it because
both failure modes are plausible and removing it blind could as easily break the
feature as fix it.

A transient timeout permanently loses a child's model.lookupStarted is
latched before the request and never reset on failure or timeout, so one 5s
timeout leaves that child without model/effort for the session. The
closed + lookupStarted pair likewise blocks re-lookup for a re-registered
child.

Cleared effort keeps a stale value.updateCollabChildMetadata only assigns
truthy values, so an effort cleared by thread/settings/updated retains the old
one even with overwriteKnown: true.

collabChildMetadataRef grows without bound. Entries are only flagged
closed, never deleted, and the new intercept inserts one for any non-root
thread id — unlike the sibling maps, which clean up on thread/closed.

The review separately cleared the parts I most wanted checked: the contracts
already carry model/effort on task payloads and the client fold consumes them,
the thread/settings/updated and model/rerouted schemas match the field
accesses, the routing/suppression change is consistent, and the
ThreadBackgroundLiveness widening is required by the new status-free
task.updated without regressing another adapter path.

Codex sub-agents did not carry their model metadata into task events, so the Agents panel could not show a child's model or reasoning effort. Reads model and effort from each spawned child thread, updates them when Codex reports a settings change or a model reroute, and includes them in later task events. Unknown values are omitted rather than inherited from the parent. Adopted from T3 Code 49f6241ddeee2a5d20ec181b5b94ec0a9ac72f7b (pingdotgg/t3code#8502). Adaptation: two conflicts, both places where Pylon and upstream had each added fields to the same child-identity payload. CodexAdapter kept Pylon's toolUseId alongside upstream's model and effort; the linkage object below already spread all three. CodexSessionRuntime adopted upstream's collabChildIdentity helper and spreads Pylon's toolUseId at the call site. The helper describes the child, while toolUseId names the child's live turn, and turn/started overrides it with the id the notification carries — so folding it into the helper would have conflated two different lifetimes.

Review follow-up. The paragraph adopted from upstream #8502 named T3 Code in a docs/user page, which is shipped-product voice. The remaining T3 Code mentions in install.md are deliberate: they describe Pylon installing beside T3 Code.

rynfar · 2026-08-28T20:32:31Z