feat(parser): add TraeX (TRAE CLI 2.0) as a first-class agent

[GhostFlying]

TRAE CLI 2.0 (traex) is a closed-source fork of codex-rs. It writes rollout
JSONL that is field-for-field the Codex shape into ~/.trae/cli/sessions,
including the subagent session_meta, token_count records, and a sibling
history.jsonl at the same position relative to the sessions root. Until now
the only way to index it was to declare that root as a codex source, which
gives the sessions the wrong agent and an ID prefix whose one-click resume
command (codex resume <id>) cannot resume them.

This adds traex as its own agent, served by the Codex provider through a
relabel hook — the same shape opencode_provider.go already uses to serve
Kilo, MiMoCode, and Icodemate from one implementation. Parsing, discovery,
fingerprinting, and JSONL-tail incremental appends are shared; only the agent
label and the traex: ID prefix differ. internal/parser/traex.go is the
whole fork-specific surface and is the first place to look.

The relabel covers more than Kilo's does. Besides the session ID, parent, and
source-session ID, the Codex parser stamps codex: onto the
SubagentSessionID of tool calls and their result events, so a partial
relabel would leave a TraeX parent pointing at codex:<uuid> rows the
traex: namespace never stores. The provider's incremental path returns
appended messages with no session, so the hook takes (*ParsedSession, []ParsedMessage) and treats a nil session as that path.

Discovery keys are now agent-scoped (CodexSourceKey takes an agent). Both
agents name files by UUID, so a shared key would let one agent's session
resolve to the other's file.

In internal/sync, the branches that describe the format — same-UUID
duplicate resolution, dated-over-flat layout preference, reconciliation
identity, the incremental file_mtime basis, and the parse-diff raced guard —
move behind a new isCodexFormatAgent predicate. The branches that describe
Codex's out-of-band files stay keyed to parser.AgentCodex alone, because
TraeX writes neither: there is no session_index.jsonl sidecar (so no title
backfill, no index-driven cache invalidation, and no index watch or changed-path
fan-out), and no s3://.../raw/codex/... archive convention, so a traex root
that looks like an S3 URI is skipped rather than imported under Codex's
identity. That split is the main thing worth reviewing.

Two consequences of having no index file are worth stating. Titles come from
the rollout head only, so a TraeX session never picks up a renamed thread name.
And ~/.trae/cli is not watched shallowly, unlike ~/.codex, since the only
reason to watch it would be the index file and the directory otherwise holds
constantly-rewritten SQLite WALs.

Evidence for the format is observational and documented as no-public-source
in the format inventory: TRAE CLI ships only as a binary, and neither Trae's
product site nor the public Trae-AI/Trae repository publishes a session
schema. A de-identified rollout with subagent spawn/wait records is kept as a
fixture.

Switching an existing installation from the codex-source workaround to
agent = "traex" is not handled: the stored rows are keyed by file path, so
the old codex:<uuid> rows would satisfy the freshness skips and the new rows
would never be written. A full resync into a fresh database is the intended
path.

🤖 Generated with Claude Code

[roborev-ci]

roborev: Combined Review (fc9dffd)

Changes requested: one medium-severity provider-isolation issue could corrupt session attribution.

Medium

• internal/sync/engine.go:9962 — Incremental and freshness lookups remain keyed only by file path. If a TraeX rollout was previously indexed as Codex, or configured roots overlap, an unchanged TraeX source may be skipped using the Codex row. After an append, GetSessionForIncremental may return the Codex session ID and append relabeled TraeX messages to that session.

  Fix: Scope incremental lookups, database freshness checks, and skip-cache identities by agent. Reject incremental results whose stored session belongs to another agent, and add a Codex-to-TraeX reassignment regression test.

---

Reviewers: 2 done | Synthesis: codex, 11s | Total: 9m14s

[wesm]

looking

[roborev-ci]

roborev: Combined Review (fe90f9e)

Code changes require fixes before approval: two medium-severity agent-isolation issues remain in TraeX sync handling.

Medium

• Verified-source state is keyed only by path — internal/sync/verified_source_gate.go:216,244

  When Codex and TraeX share or reassign a root, one agent’s trusted signature and current database row can cause the other agent’s sync to be skipped, leaving its missing row unrepaired.

  Fix: Include the agent in verified-source cache keys and add an agent-scoped equivalent of GetSourceRepairStateByPath.

• TraeX incorrectly inherits Codex session-index behavior — internal/sync/engine.go:10152,10437

  TraeX uses Codex-specific session_index.jsonl logic through CodexEffectiveMtime, LookupCodexThreadName, and the index-title skip branch. A copied index or reassigned root can apply Codex titles to TraeX sessions or consult the Codex row when evaluating freshness.

  Fix: Gate index lookup, effective-mtime folding, eviction, and title comparison on AgentCodex; use transcript-only metadata for TraeX.

---

Reviewers: 2 done | Synthesis: codex, 12s | Total: 8m48s

[roborev-ci]

roborev: Combined Review (03c0a34)

Medium-severity synchronization issues remain and should be fixed before merging.

Medium

• internal/sync/engine.go:8998 — Appending ?agent=... breaks remote skip-cache translation for exact-file mappings such as Hermes state.db. The qualified path is treated as a sibling of the mapped file, causing tempPathToRemotePath and remoteArchiveRel to reject it and silently drop the cache entry. Split recognized cache qualifiers before remote path translation and reattach them afterward in both directions; add an ExtraFiles round-trip test with an agent-qualified key.

• internal/sync/engine.go:9970 — TraeX’s incremental gate uses the path-only pathNeedsProjectReparse; the cached-skip check at line 8304 has the same unscoped lookup. If Codex and TraeX rows share a path, another agent’s newer project can mask stale TraeX metadata, allowing an incremental append that preserves it. Pass the provider agent through the repair helpers and use GetProjectByAgentPath, GetFileInfoByAgentPath, and GetDataVersionByAgentPath.

---

Reviewers: 2 done | Synthesis: codex, 10s | Total: 8m52s

[roborev-ci]

roborev: Combined Review (caa475c)

No issues found.

---

Reviewers: 2 done | Synthesis: codex | Total: 11m54s

[roborev-ci]

roborev: Combined Review (ebe92ac)

Medium-severity issue found in TraeX repeat-sync handling.

Medium

• internal/sync/engine.go:10933 — An unchanged TraeX rollout always enters the equal-size incremental branch. Because that branch exempts only Codex, it sets forceReplace and bypasses later database freshness checks, causing every periodic sync to reparse and replace all messages indefinitely.
  • Fix: Use isCodexFormatAgent(agent) so TraeX reaches the fingerprint-based freshness check. Add a repeat-sync test verifying that an untouched TraeX session is skipped.

---

Reviewers: 2 done | Synthesis: codex, 11s | Total: 13m54s

[roborev-ci]

roborev: Combined Review (eff90e8)

Code review passed with no medium-or-higher severity findings.

---

Reviewers: 2 done | Synthesis: codex, 10s | Total: 23m27s