docs: v0.9.0 code-freeze plan

[trinity-ability]

Output of /release-plan. Work order for the v0.9.0 freeze (2026-08-12): 3 MUST-LAND items (#2095, enterprise pin bump to bce1175, TRINITY_DEFAULT_SKILL_SOURCE compose passthrough), 7 SHOULD-LAND ready PRs, evaluations FENCE checklist, and the Known-limitations block for /release Step 5.

🤖 Generated with Claude Code

[dolho]

ent#286 (portal streaming) — recommend cut from v0.9.0, design recorded

Investigated it today. It is not the small item the plan implies, and I do not think it should land before the freeze. Full write-up on abilityai/trinity-enterprise#286; the short version:

portal_chat awaits execute_task and returns {response, cost, session_id}. The turn is over when the response arrives and no execution_id reaches the client — so an SSE endpoint added alone is unreachable dead code. Real streaming means splitting the chat into dispatch → stream → finalize, which changes a shipped contract in three places: the response shape, the history-persist timing (currently after the await), and the 429/504/502 mapping (currently derived from the completed result). It also needs "public" added to ASYNC_DISPATCH_ELIGIBLE_TRIGGERS, which is a decision about the #1083 fire-and-forget path rather than a config tweak.

The issue's own body rates it P3: "a latency/UX improvement, not a missing capability." It was raised to P1 for this plan. Meanwhile ent#362 and ent#218 / ent#220 have no code at all and genuinely block the Workspace story.

Trading a shipped-contract change for a latency win, 8 days out, while two P1s are unstarted, looks like the wrong risk. Design is settled on the issue so it is pick-up-and-build in v0.9.1.

---

Other corrections to the plan, verified today

Already done — the plan overstates the remaining work:

Item	Plan says	Verified
PR #2024	"just merge it"	MERGED
#1976	"fix red CI, then merge"	MERGED
ent#292 (P0, PAT leak)	blocks the cut	status-in-dev — fix merged (bbc6b304)
ent#297 / #293 / #334	need a ship/block decision	all three status-in-dev, fixes merged

So the agent-key-admin cluster decision is largely retrospective — worth recording, not blocking.

The Alembic check passes — bump is safe. d1c5ebb → bce1175 is one commit (the ent#356 removal) and zero migration files:

d1c5ebb: 16 revisions, 1 head (0015_merge_a2a_control)
bce1175: 16 revisions, 1 head (0015_merge_a2a_control)
OSS:     37 revisions, 1 head (0036_client_portal_oss)

The #2068 two-heads failure mode does not apply to this bump. (Note for anyone re-checking: a naive head count reports 3 heads because 0015_merge_a2a_control is a merge revision whose down_revision is a tuple — parse it as one or you will get a false alarm.)

The compose passthrough gap is real and confirmed:

config.py                reads TRINITY_DEFAULT_SKILL_SOURCE(_REF)   yes
.env.example             documents it                              yes
docker-compose.yml       references                                0
docker-compose.prod.yml  references                                0

Inert on every deployed instance. This is the seventh recurrence of that class in learnings.md (VOIP_, LOG_, #1039, AGENT_LOG_, CANARY_, TEMPLATE_REGISTRY_*), and prod compose is the load-bearing file.

A DIRTY PR runs no CI at all. #2099 showed zero checks — not green, not red, unrun. Any "CI green" line in the plan needs re-checking after a rebase. #2095, #2098, #2099 and #1976 were all conflicted and are now rebased and running; #2076 is still DIRTY and failing lint (sys.modules pollution check).

[vybe]

Docs-only release artifact under docs/releases/ (explicitly out of scope for enterprise-docs-guard). Security scan clean: no secrets, emails, or public IPs. All 4 required checks green.

Merging as the point-in-time record of the /release-plan run. Noting @dolho's corrections above are not folded into the doc — ent#286 cut recommendation, and the four already-done items (#2024, #1976, ent#292, ent#297/#293/#334 all status-in-dev). The comment thread carries them; if the plan is going to be worked from tomorrow rather than just archived, it wants a follow-up commit.