fix(agent): an environment hard stop persists nothing, and recall follows the status lineage

[using-system]

Summary

The mission side of environment targeting:

• An environment hard stop persists nothing (observe-run Setup step 5 covers its own divergence and step 4's split; the self-check's persistence item is exempted; the observe-run-report reference says so; /odd-observe and /odd-verify relay the stop — both values, the queries, the remedies — instead of running show when no stored path comes back).
• On a replay, the handoff's environment= is the report's, and the configured one plays no part (one clause in Setup step 5).
• grafana-discover.py prints each service's deployment environment, read from the log stream and metric series labels it already fetches (no extra query); several values print as a list. Verified live on Grafana Cloud.
• Recall matches the same service set, the lineage get-status keys by, so a run's baseline sits in the same status row as the run.

Choices amended against the issue text are recorded on the issue: #658 (comment)

Tests

• tests/skills/odd-memory/test_odd_recall.py: equal-set matching; tests/skills/observability-cli-guides/test_grafana_scripts.py: environment read off streams and series (both, logs only when series fail, several values, none) on masked fake-gcx replays.
• pytest tests/skills 878 passed (re-run after rebasing on fix(mcp): refuse the environment writes that resolve to a wrong or hidden entry #660), ruff 0.16.4 check/format clean, check_stack_reference.py green, apm install --target claude + apm audit green.

Review

A separate reviewer sub-agent checked the whole branch against the issue under the bound-review rule (recall callers, /odd-verify, benchmark recall, stored reports, a live re-run of the discover script): no finding, ready to merge on the first round. The rebase onto #660 was conflict-free.

Harness measurement

test-plugin-harnessing, copilot / openai/gpt-5.6-luna / medium, /odd-observe drive mission on the local llms-benchmark stack, ABBA (base1, after1, after2, base2); every changed file was in the deployed path.

Side (mean of 2)	Turns	Commands	Tokens in	Tokens out	nanoAIU	Wall	Preflight	Observation
main	39.5	38.5	2.81 M	15.7 k	11.53 G	341 s	43 s	169 s
branch	38	47.5	2.37 M	14.8 k	10.35 G	365 s	23.5 s	202 s

Main's spread today over eight samples: turns 35–45, commands 30–51, tokens in 2.66–3.41 M, wall 310–431 s, preflight 43–75 s. No degradation on the four axes: turns, output and wall inside main's spread; input tokens and AIU slightly below it on both branch samples; the shorter preflight is paid back in the observation, so wall is level — not claimed as a gain on two samples. One branch sample made two --help calls, all on scripts this branch does not touch (replay_benchmark.py, probe_services.py, grafana-context.py, gcx_local.py). No sample had a stored baseline for these services, so the recall change was not exercised by the mission (it is covered by the unit tests). Every sample wrote a report (4–5 anomalies, 2–5 gaps; main's range today 4–7 / 4–6).

Closes #658

🤖 Generated with Claude Code