Run Codex sweep first, then request final architecture review.

Architecture + Code Review (full pass on PR #43)

I did a full read of the A2A implementation + tests and also reviewed the non-A2A deltas included in this PR.

Overall verdict

Request changes before merge.

The A2A core is directionally strong (types/schema/validator/router/audit/tests are all present and reasonably cohesive), but there are several merge blockers and a few reliability gaps that should be fixed first.

✅ What looks good

• Clean split of protocol layers: schema/types/validator/router/audit/SDK.
• Router pipeline is easy to follow and test.
• Circuit breaker includes both pair-rate and correlation-depth protection.
• Integration tests cover realistic multi-agent flows (task lifecycle, review loop, rate limit, circuit breaker, audit integrity).

❌ Blockers

1) Scope contamination in a mega-branch that is labeled as A2A consolidation

This PR includes 21 non-A2A files in addition to A2A changes (session auto-labeling, sessions_send timeout config, config schema changes, retry-policy/export fixes, etc.).

Examples:

• src/sessions/session-auto-label.ts
• src/agents/tools/sessions-send-tool.ts
• src/config/*
• src/infra/retry-policy.ts

For a consolidation PR intended to represent A2A workstreams (#37–#41), this significantly increases blast radius and makes regression triage harder.

Required fix: split non-A2A changes into separate PR(s), or rebase this branch to include only A2A scope.

2) Router has an unsafe no-validator path and inconsistent audit behavior

In src/gateway/a2a/router.ts:

• route() allows validate to be omitted, then casts input as A2AMessage and dereferences message.from.agentId directly (lines 135-141). Invalid input can throw before controlled error handling/audit.
• On validation_failed, router returns early without audit logging (lines 123-133), while all other non-success outcomes are audited.

This creates observability blind spots and can produce uncaught runtime failures when validation is disabled.

Required fix:

• Make validation mandatory in production path (or add hard runtime guards before dereference).
• Audit validation failures too (at least metadata + error summary).

3) SDK silently drops messages when send function is not set

In src/gateway/a2a/sdk.ts, send() returns the message even if sendFn is null (lines 88-93).

That is dangerous: callers can interpret this as successful send while nothing was delivered.

Required fix: throw explicit error when sendFn is unset (similar to the context guard pattern already used for getContext()).

⚠️ Important reliability gaps (should fix before merge if possible)

4) Circuit breaker pair-key collision risk

pairKey() uses "${sorted[0]}:${sorted[1]}" (src/gateway/a2a/circuit-breaker.ts). If agent IDs can include :, distinct pairs can collide.

Fix: use a collision-safe key encoding (e.g., JSON stringify tuple of sorted IDs, or a length-prefixed separator).

5) Envelope validation is looser than comments/spec imply

In src/gateway/a2a/validator.ts, timestamp is only checked as non-empty string (lines 126-139), despite error text claiming ISO 8601. Also top-level unknown envelope properties are not rejected.

Fix: enforce timestamp format and consider envelope additionalProperties: false behavior.

6) Audit query pagination bounds and corruption tolerance

In src/gateway/a2a/audit-query.ts, limit/offset are not clamped to non-negative (lines 117-118), and in audit.ts, one malformed JSONL line causes readLogFile() to throw entire file read (line 141).

Fix:

• clamp limit >= 0, offset >= 0
• tolerate malformed lines (skip+count/report) to keep query usable under partial log corruption.

Test coverage notes (gaps)

Current tests are solid on happy paths and key controls, but missing cases for:

• validator-omitted router path with malformed input
• validation-failed audit emission
• SDK behavior when send function not registered
• circuit-breaker pair-key collision inputs
• negative pagination inputs in audit query
• malformed JSONL line tolerance

Merge readiness

Given the blockers above, especially scope contamination + unsafe/no-audit behaviors, I do not recommend merging this as-is.

If you want, I can provide a concrete split plan (A2A-only PR cut + follow-up PRs for session-label/timeout/config changes) so this can be landed safely and quickly.

Run Codex sweep first, then request final architecture review.