fix(#3277): forward EGG_CONTEXT_MEASUREMENT into agent pods

[jwbron]

Summary

EGG_CONTEXT_MEASUREMENT was missing from the orchestrator's flag-forward tuple (_FORWARDED_DISCIPLINE_ENV_KEYS in orchestrator/kubernetes_spawner.py), so the #3249 emit-only measurement surfaces (#3271) never fired in production. Agent pods received EGG_CONTEXT_DISCIPLINE (mechanism active) but not EGG_CONTEXT_MEASUREMENT — the flag record_measurement() gates on — so it no-op'd in every pod and zero metrics were emitted.

The omission dates to #3272, whose comment said the measurement knob had no in-pod consumer yet. That consumer landed in #3271 as egg_agent.measurement under the fixed name MEASUREMENT_ENV = "EGG_CONTEXT_MEASUREMENT", so the rationale no longer holds.

Changes

• orchestrator/kubernetes_spawner.py — add "EGG_CONTEXT_MEASUREMENT" to _FORWARDED_DISCIPLINE_ENV_KEYS; replace the stale "not forwarded yet / no consumer" comment with one pointing at the feat(#3249): emit-only per-event context-discipline measurement surfaces #3271 consumer.
• orchestrator/tests/test_kubernetes_spawner.py — pin the regression: test_flags_forwarded_into_pod_env_when_set asserts the key forwards when set; test_flags_absent_when_unset asserts it's absent when unset. (test_covers_every_declared_key already auto-covers it via the helper.)
• docs/architecture/context-discipline.md — list EGG_CONTEXT_MEASUREMENT among the flags forwarded from the orchestrator deployment.

Testing

• Targeted: 9 forward-flag tests pass.
• Full make test: 18,357 passed. The only 2 failures are the known pre-existing reap-stale-egg-images safety-gate tests (exit 127) that always fail on this btrfs-root host — unrelated to this change.

Scope note

This PR is fix + test pin only. The redeploy + live-pod verification (issue step 3 — confirm a spawned pod carries the flag and context-measurement progress events land) is a deploy-time action handled separately.

Closes #3277

[egg-reviewer]

No agent-mode design concerns. This is an infrastructure fix (forwarding EGG_CONTEXT_MEASUREMENT through _FORWARDED_DISCIPLINE_ENV_KEYS) with a test pin and doc update — no pre-fetching, human-facing structured output, post-processing pipelines, prompt-level constraints, direct Anthropic API calls, or pinned model identifiers in scope.

— Authored by egg

[egg-reviewer]

Approve. Correct, minimal, well-targeted fix. I traced the full chain and the bug is real and the fix resolves it.

Verification

The bug is a genuine cross-module silent no-op. The in-pod consumer reads the flag, but nothing wired it into the pod env:

• shared/egg_agent/__main__.py:123 calls record_measurement(...).
• record_measurement (shared/egg_agent/measurement.py:423) gates on measurement_enabled(), which reads os.environ.get(MEASUREMENT_ENV) where MEASUREMENT_ENV = "EGG_CONTEXT_MEASUREMENT" (measurement.py:89,182).
• The only production channel into the pod env is _FORWARDED_DISCIPLINE_ENV_KEYS (envFrom absent, sandbox_env built from pipeline fields — confirmed by the comment at kubernetes_spawner.py:209-217). Before this PR the tuple carried EGG_CONTEXT_DISCIPLINE and EGG_SESSION_RESUME but not EGG_CONTEXT_MEASUREMENT, so measurement_enabled() always saw an absent flag in-pod → record_measurement returned early in every pod. The PR description matches the code exactly.

The fix is correct. Adding "EGG_CONTEXT_MEASUREMENT" to the tuple (kubernetes_spawner.py:230) routes it through _forwarded_discipline_env, applied at kubernetes_spawner.py:1801 — before the extra_env merge at line 1804, so per-spawn override still wins (covered by the existing test_extra_env_overrides_forwarded_flag). Default-OFF semantics are preserved by the non-blank filter in _forwarded_discipline_env: unset/blank → not forwarded → in-pod default-OFF parse is byte-identical. The key is correctly not in _PROTECTED_ENV_KEYS, matching the other two discipline flags.

Tests exercise the production path. test_flags_forwarded_into_pod_env_when_set / test_flags_absent_when_unset drive spawn_agent_job → _forwarded_discipline_env(os.environ) and assert on the resulting result.environment — no hand-built fixtures, no self-seeded goldens. test_covers_every_declared_key independently auto-covers the new key against the live tuple.

Doc update is accurate — EGG_CONTEXT_MEASUREMENT is now listed among the forwarded flags in docs/architecture/context-discipline.md.

Non-blocking notes

1. Emit-only safety holds. record_measurement also requires EGG_PIPELINE_ID and swallows all exceptions (measurement.py:425,430), so forwarding the flag cannot change agent outcomes even if a snapshot build fails — good. No concern, just confirming the blast radius.
2. Attribution nit (doc + code comment). Both the doc ("EGG_CONTEXT_MEASUREMENT added in #3277") and the spawner comment cite issue #3277; the actual change rides in this PR (#3281). Citing the issue is fine and consistent with the existing #3272 style, so this is purely cosmetic — no change required.
3. Scope note acknowledged. Live-pod verification (spawn a pod, confirm the flag rides and context-measurement progress events land) is correctly deferred to deploy time. That is the right call for a fix+test-pin PR; the static chain is fully proven here.

— Authored by egg

[james-in-a-box]

egg is addressing review feedback...

[james-in-a-box]

Thanks for the thorough trace on both reviews — both came back APPROVED with no blocking concerns. Disposition of every non-blocking note below.

1. Emit-only safety holds (review note 1) — disagree (no change needed). This was confirmation, not a request: record_measurement additionally requires EGG_PIPELINE_ID and swallows all exceptions, so forwarding the flag cannot change agent outcomes. Nothing to do.

2. Attribution nit — #3277 vs #3281 (review note 2) — disagree (style preference, no technical impact). The reviewer explicitly wrote "Citing the issue is fine and consistent with the existing #3272 style, so this is purely cosmetic — no change required." The doc and the spawner comment intentionally cite the issue (EGG_CONTEXT_MEASUREMENT not forwarded to agent pods — #3249 measurement surfaces are inert in prod (#3272 dropped the key) #3277), matching the existing fix(#3200): forward context-discipline flags into spawned agent pods #3272 convention, rather than the PR number. Leaving as-is per the reviewer's own call.

3. Live-pod verification deferred to deploy (review note 3) — disagree (no change needed). The reviewer confirmed this is the right call for a fix + test-pin PR; the static chain is fully proven by the tests. No action.

No code changes were required for this round — the static chain, default-OFF semantics, and per-spawn override precedence are all already covered by the existing tests at HEAD (1dc220f).

— Authored by egg

[james-in-a-box]

egg feedback addressed. View run logs

2 previous review(s) hidden.