Corpus migration: 40 cells declare `env { ST_AGENT "<team>.<seat>" }`, which st2 now rejects at boot (the runner owns ST_AGENT)

[schickling]

st2 now owns ST_AGENT and forces it to the runner-derived bus id <host>.<team>.<seat>. Any authored value that differs is rejected at admission, before any model starts. This is an intentional upstream boundary, not an st2 bug (see Why this is intended below) — but it silently invalidates most of this corpus.

40 of the 59 cells declare the short env { ST_AGENT "<team>.<seat>" } form, across 94 seat declarations, and every one of them now aborts its eval at boot:

Error: agent '<host>.t.alpha' task 'agent' declares conflicting ST_AGENT 't.alpha'; expected runner-owned value '<host>.t.alpha'

The migration is a one-line deletion per seat. The runner injects ST_AGENT unconditionally, so every prompt that uses $ST_AGENT keeps working with no other change.

Reproduction

Free and self-contained — every seat is exec sleep 5, no model is ever launched. Needs only st2 on PATH; does not touch this repo.

#!/usr/bin/env bash
cell=$(mktemp -d); mkdir -p "$cell/fixture/alpha" "$cell/fixture/beta" "$cell/fixture/j"
echo 'Any body.' > "$cell/task.md"
HOST=$(hostname)

run() { # run <label> <team-body> [<extra-eval-body>]
  printf '\n--- %s\n' "$1"
  cat > "$cell/cell.kdl" <<KDL
team "t" {
$2
}
eval { copy "./fixture"; message { from "requester"; to "t.alpha"; content "./task.md" }; max-timeout "8s"
${3:-}
  judges { judge "the team booted" { exec "true" } } }
KDL
  st2 eval "$cell" 2>&1 | grep -E '^(== boot|Error:|  \[|SCORE|VERDICT)' | sed "s/$HOST/<host>/g;s/^/    /"
}

st2 --version
run 'CASE A — a team seat declares the short form ST_AGENT "t.alpha"' \
    '  agent "alpha" { workspace "./alpha"; env { ST_AGENT "t.alpha" }; command "exec sleep 5" }'
run 'CASE B — control: the env block omitted entirely' \
    '  agent "alpha" { workspace "./alpha"; command "exec sleep 5" }'
run "CASE C — control: ST_AGENT declared as the runner-owned <host>.t.alpha" \
    "  agent \"alpha\" { workspace \"./alpha\"; env { ST_AGENT \"$HOST.t.alpha\" }; command \"exec sleep 5\" }"
run 'CASE D — blast radius: alpha valid, beta conflicting' \
    '  agent "alpha" { workspace "./alpha"; command "exec sleep 5" }
  agent "beta"  { workspace "./beta";  env { ST_AGENT "t.beta" }; command "exec sleep 5" }'
run 'CASE E — an agent declared inside eval{} with a bare ST_AGENT' \
    '  agent "alpha" { workspace "./alpha"; command "exec sleep 5" }' \
    '  agent "judge" { workspace "./j"; env { ST_AGENT "judge" }; command "exec sleep 5" }'
rm -rf "$cell"

Actual output (hostname elided as <host>):

st2 0.1.0+b868a07 — committed 4 days ago

--- CASE A — a team seat declares the short form ST_AGENT "t.alpha"
    == boot team (1 agents) ==
    Error: agent '<host>.t.alpha' task 'agent' declares conflicting ST_AGENT 't.alpha'; expected runner-owned value '<host>.t.alpha'

--- CASE B — control: the env block omitted entirely
    == boot team (1 agents) ==
      [PASS] the team booted  (exit 0)
    SCORE: 1 PASS / 0 FAIL / 1 gating judges
    VERDICT: PASS

--- CASE C — control: ST_AGENT declared as the runner-owned <host>.t.alpha
    == boot team (1 agents) ==
      [PASS] the team booted  (exit 0)
    SCORE: 1 PASS / 0 FAIL / 1 gating judges
    VERDICT: PASS

--- CASE D — blast radius: alpha valid, beta conflicting
    == boot team (2 agents) ==
    Error: agent '<host>.t.beta' task 'agent' declares conflicting ST_AGENT 't.beta'; expected runner-owned value '<host>.t.beta'

--- CASE E — an agent declared inside eval{} with a bare ST_AGENT
    == boot team (2 agents) ==
    Error: agent '<host>.judge' task 'agent' declares conflicting ST_AGENT 'judge'; expected runner-owned value '<host>.judge'

What each case establishes:

• A — the form this corpus uses is rejected at boot. The eval exits non-zero with the reason printed, so it fails closed and loudly (no silent pass).
• B — this is the migration. Deleting the env block entirely boots cleanly, and $ST_AGENT is still set in the task environment by the runner.
• C — declaring the runner-owned value also boots, but it hard-codes a machine hostname into a cell, so it is not portable. Do not migrate this way.
• D — one conflicting seat aborts the whole eval; the valid sibling is never launched. A partial migration therefore buys nothing: a cell has to be fully clean to run.
• E — an agent declared inside eval { } rather than team { } goes through the identical check and fails identically. license-mit / license-mit-codex declare env { ST_AGENT "judge" } on exactly this kind of eval-level judge — there is no special case for it, and it needs the same edit.

Why this is intended upstream

• compoundingtech/st2#64 — Decide ownership of stable agent identity at PTY creation — poses the choice explicitly. Direction A. Runner-derived identity: "st2 forces ST_AGENT=<host.identity> into every task of the agent; authored conflicts either fail validation or lose to the runner." That is the boundary that shipped.
• compoundingtech/st2#192 — One conflicting ST_AGENT aborts the whole reconcile pass, and st2 up --once still exits 0 — records the provenance: "The behaviour arrives with the runner-owned task identity work; the previous revision supervises both specs and does not fail the pass." The open question there is about st2 up --once's exit status and the scope of the abort, not about whether the runner owns the value. st2 eval (Case A/D above) already exits non-zero with a clear diagnostic.

The relevant st2 code, on main at a8dc061485ec38e8aca2f3118a05b131254c9663:

• src/reconcile.rs:220-242, validate_task_identities — for every running local spec, task.env.get("ST_AGENT") must either be absent or equal spec.bus_id(this_host); otherwise TaskIdentityAdmissionError::Conflict. It is called from reconcile (:395, :505), from src/materialize.rs (:620, :824, :884) and from the eval/up paths in src/run.rs (:1472, :1643, :1746, :1795), so there is no path that skips it.

• src/reconcile.rs:245-257, runner_task_env — unconditionally inserts ST_AGENT=<bus_id> into the launched task's environment:

  let mut env = task.env.clone();
  env.insert("ST_AGENT".to_owned(), bus_id.to_owned());

  This is why deleting the declaration is safe: the variable the prompts read is supplied by the runner, with the correct host-qualified value that a cell cannot know at authoring time.

st2 validate reports the same conflict, so a migration can be checked without running an eval:

$ st2 validate --catalog <root> --host h1
ERROR  agents/h1/alpha/agent.kdl: agent 'h1.alpha' task 'agent' declares conflicting ST_AGENT 'alpha'; expected runner-owned value 'h1.alpha'
─ 1 error, 0 warnings across 1 agent

Scope in this repo

Measured at 12c96acbbdc654f27ccb27cae064ef4311d4f7f7:

grep -lE 'env[[:space:]]*\{[[:space:]]*ST_AGENT' cells/*/*.kdl | wc -l   # -> 40 cells
grep -hcE 'env[[:space:]]*\{[[:space:]]*ST_AGENT' cells/*/*.kdl | paste -sd+ | bc   # -> 94 seats

Every declared value is short: 92 are two-part <team>.<seat>, 2 are bare (ST_AGENT "judge"), and zero are host-qualified — so all 94 conflict on every host.

The 40 affected cells (seat counts)

• assignment-contract-cold-assignment (2)
• assignment-contract-cold-focus (2)
• assignment-contract-cold-resources (2)
• assignment-contract-handoff-assignment (4)
• assignment-contract-handoff-focus (4)
• assignment-contract-handoff-resources (4)
• assignment-contract-hot-assignment (3)
• assignment-contract-hot-focus (3)
• assignment-contract-hot-resources (3)
• crash-ding (7)
• ding-mode (2)
• ding-reply (1)
• docs (2)
• feature-fit (2)
• fork-in-the-road (4)
• fork-in-the-road-codex (4)
• ghost-bug (2)
• ghost-bug-codex (2)
• inbox-hygiene (2)
• incident-response (2)
• license-mit (3)
• license-mit-codex (3)
• migration (2)
• poisoned-pr (2)
• poisoned-pr-codex (2)
• restart-continuity (3)
• security-audit (2)
• signal-rename (4)
• signal-rename-codex (4)
• skill-inheritance (1)
• test-writing (2)
• vrs-cross-file-absent (1)
• vrs-cross-file-present (1)
• vrs-definition-of-done-absent (1)
• vrs-definition-of-done-present (1)
• vrs-scope-drift-absent (1)
• vrs-scope-drift-present (1)
• vrs-scope-pressure-absent (1)
• vrs-scope-pressure-present (1)
• weird-git-setup (1)

Not affected: the 7 canonical agent.kdl files inside cell fixtures that declare ST_AGENT (context-resource-continuity, hook-integrity ×2, managed-agent-color-env ×2, st2-network, st2-doctor-structure) all declare an explicit host and a fully host-qualified value matching it — e.g. host "cr" with ST_AGENT "cr.agent", host "color" with ST_AGENT "color.ambient". Those already equal the runner-owned bus id, so they pass validate_task_identities unchanged. managed-agent-color-env in particular appears to exercise env precedence deliberately; nothing in this issue asks for it to change. (render-target-safety's fixture spec only consumes $ST_AGENT in a render target, file ".st2/generated.txt" "AGENT=$ST_AGENT" — it declares nothing, and the runner-injected value keeps that working.)

Separate, overlapping breakage worth knowing about while migrating

Migrating ST_AGENT will not by itself make every affected cell green. compoundingtech/st2#248 reports that a compact-team eval's flat message bus is disabled by any spec-shaped file under the run catalog, and compoundingtech/st2#247 reports that any package.json with a top-level type key qualifies. 22 of the 59 cells here ship such a fixture manifest and all 22 declare ding. That is an upstream defect, not a corpus problem, and it is tracked there — noted only so a failing post-migration run is not mistaken for an incomplete migration.

Suggested action

Delete the env { ST_AGENT "…" } declaration from all 94 seats (including the eval-level judges), keeping any other variables in the same env block. Do not substitute a hostname. Prompts referring to $ST_AGENT need no change.

Versions

• st2 --version: 0.1.0+b868a07 — b868a0710a59fc40efd44a3e96bc92903d1376bb. All output above was produced by this binary.
• st2 source citations: main at a8dc061485ec38e8aca2f3118a05b131254c9663 (2026-08-11). Caveat: that tree was source-inspected only, not built or run; validate_task_identities and runner_task_env are unchanged there relative to the binary under test.
• Last revision observed booting the short form: st2 9887b2842222def0838c2cd82e6c24c218f7efa6 (2026-07-28), consistent with the provenance note in st2#192.
• Corpus counts: this repo at 12c96acbbdc654f27ccb27cae064ef4311d4f7f7.
• OS: Linux 6.18.33.