Halo Campaign Evolved CU4 under the hood: every tag the August 17 update touched

Halo Campaign Evolved updated to CU4 (2026.08.11.1121610.2-Rel-i343-Meteorite-2607-CU4) on August 17 — the executable's version stamp says August 11, which is when it was built, not when Steam shipped it. We snapshotted the build the day it landed and ran every one of the game's 12,292 Blam tags through a field-level diff against CU3. The verdict: 1 tag added, 0 removed, 358 changed, 11,931 byte-identical — and inside those 358 is a real gameplay rebalance the patch notes didn't mention.

What actually changed for gameplay

The AI got better at shooting you. The shared AI tuning tag (ai/generic-character) carries 199 field-level changes: two new firing patterns (69 → 71), and an existing pattern rebuilt end-to-end — rate of fire 0 → 3, projectile error 0.035 → 0.009, maximum error angle halved twice over, burst duration tightened from (0.56, 0.75) to (0.25, 0.45), target tracking relaxed 0.9 → 0.5. The Flood got the same treatment: floodcombat_base-character adds a firing pattern of its own and retunes over 250 values. If firefights feel different on CU4, they are.

Jackals have new hitboxes. jackal-model drops from 10 model targets to 8, retargets target_elbow_r/l to target_hand_r/l, flags a target as headshot, and rebalances targeting relevance (one target 0.5 → 1, another 0.5 → 0.01). Aim-assist and auto-aim against Jackals now resolve to different bones than they did on CU3.

Active camo tolerates faster movement. In globals, the camo biped speed reference rose 3.25 → 3.7 — the speed at which movement degrades your camouflage moved up.

Mission scripting was touched in three places. Mission E10 gained an entire new script (526 → 527) and rewires placement scripts across a half-dozen squads' spawn points; e30 swaps the unit definitions bound to six scripted vehicle seats; C45 swaps one squad's template — back to its pre-CU3 value, undoing a CU3 edit. Targeted encounter surgery, not a rewrite.

The encounter remix seeds rolled again, exactly as they did in CU3 — they move every update, so treat reseeding as routine churn rather than a deliberate shuffle.

One new tag in the whole update: an animation graph for the A10 cryo capsule (objects/levels/halo1/solo/a10/unsc_cryo_capsule) — the Pillar of Autumn's cryo bay picked up a newly animated prop.

What changed but doesn't matter

The other ~330 changed tags are the sound of a cooker, not a designer: 176 animation graphs whose bytes moved without a single visible field changing, 120 level-geometry BSPs whose differences are runtime pointers and mopp bookkeeping, 31 lighting-info checksums, and seam-ID renumbering. Seventeen of the 31 IoStore container sets were re-cooked — some with their directory casing changed, which our diff treats as the same asset. Among the re-cooks: the curious launch-era jorge_turret cinematic graph we flagged in the CU3 report moved bytes again without changing a field. At the file level the update also rebuilt the host executable, the simulation DLL and a sweep of third-party libraries (nine boost DLLs, libHttpClient, OpenColorIO, tbb).

The stack still works

Re-verified on CU4, in order of how much we were braced for it to break: all four UE4SS signatures resolve uniquely; all 13 MJOLNIR mods load with zero Lua errors and the in-game bridge answers; the tag reader parses the CU4 containers clean and round-trips them byte-for-byte; live tag patching still locates payloads in the running game. The tag schema is untouched — 101 groups, 1,779 structs, 13,250 fields, byte-identical definitions — so every MJOLNIR tool carries over. The two level-geometry chunks our Oodle decoder couldn't decompress fail identically here as on every build since launch — a long-standing gap in our tooling, not a CU4 regression. (Update, August 19: the decoder is fixed — both chunks now decode on every build.)

The timeline so far

How this post was made

CU4 overwrote CU3 in place, as Steam updates do — but CU3 was already in our content-addressed snapshot store, recovered bit-for-bit from Steam's depot history and verified against a lock of the live install. mjolnir tagdiff compared every tag and decoded every changed one. Every build from here on gets archived the day it lands, so the next one of these posts is a tagdiff away.