feat: compose driver and PTY session observation fidelity

[schickling-assistant]

Problem

Observed harness-state readers are fleet-wide, but rich producers exist only when an st2 native driver owns the session. A downstream 627-seat catalog launched harnesses through a different wrapper: st2 hooks verify, presence, Doctor, and catalog gates were green while zero seats produced a record and every TUI row rendered never observed.

The producer contract was an unenforced deployment property. st2 must not know about or branch on any launcher.

Goal

Every local running seat has launcher- and harness-agnostic coarse activity. Rich driver observations remain the fine layer and win when fresh.

Decisions

• Read-time composition, no coarse record writer. PTY PR feat(agent-spec): bind admission receipts to revision #168 stamps lastOutputAtMs in the daemon output path. st2 reads canonical <pty-root>/<bus-id>.json directly and derives active (≤60s) / idle (>60s), with a 30s future-skew guard.
• Explicit fidelity discriminator. observedState.fidelity = driver | session. Driver fidelity covers the full tuple. Session fidelity proves only state and since; blocked/input/ask remain unknown. Updated consumers use only the proved axes.
• Fine-over-session precedence. Definite fresh driver observation wins; missing/derived-unknown driver observation falls back to session activity.
• No launcher awareness. Canonical agent task pty_id = bus_id is runner-owned. No axe or harness name enters the projection.
• Full VRS scope. Requirements, spec, ontology, accepted decision 0006 Amendment 1, and experiment updated with explicit user approval.

Verification

• cargo check --all-targets — green.
• cargo test agents:: --no-fail-fast — green.
• cargo test observed --no-fail-fast — green (includes Doctor wire/diagnostics).
• Two full-suite timing failures passed isolated; the all-test lane then timed out after 30m in existing watch tests (load-flaky baseline).
• GitHub Actions Nix run #693 — green (check, 16m34s) after recovery from the documented Actions major outage.
• nix build .#default --no-link — green: /nix/store/0d2pkfwzbx483iqb9j02fc40p42yww7i-st2-0.1.0.
• Synthetic 627-live-seat benchmark, debug binary, 10 runs: median 394ms, max 552ms, exactly 314 active / 313 idle. No subprocesses in the read path.
• Strict VRS: new experiment passes enforced shape; repo still has 101 pre-existing strict-shape errors in older artifacts.

Complexity

One discriminator, one optional session metadata read per local row, one precedence fold. No process, subscription, history, writer identity, fencing, heartbeat, or retention lifecycle added.

Concerns

• Depends on PTY feat(agent-spec): bind admission receipts to revision #168. pty-rust parity is st2 doctor reports "presence fresh" for an agent with no status file #5; both implementations expose the same optional camelCase wire field. Until a PTY generation with lastOutputAtMs runs, session fidelity remains absent and Doctor explains the missing capability.
• The 60s activity window is intentionally separate from presence/record freshness. Captured long silent model waits may justify tuning, without changing the API.

Friction & bottlenecks

• st2 full tests exceed 30m and wedge in existing watch/timing cases; two reported failures passed isolated.
• Strict VRS currently has 101 baseline shape errors, so this PR can prove its new artifact passes but not make the whole tree green without unrelated migration.
• The original design allowed consumer adoption without producer coverage; the new session fallback removes that class for local seats.

Follow-ups

• pty-rust parity: feat: persist PTY output activity evidence pty-rust#5.
• Dotfiles/fractal integration: consume fidelity=session without poisoning coarse state with unknown fine axes; pin PTY feat(agent-spec): bind admission receipts to revision #168 and this PR; live e2e across claude/codex/opencode.
• Remote supervisor semantics remain DQ-H5; session fidelity is same-host by design.

References

• Depends on feat: persist last PTY output timestamp pty#168
• feat: persist PTY output activity evidence pty-rust#5
• Refs Observed harness state: the activity-lease slice of #162, specified and made readable #268
• Observed harness state: the activity-lease slice of #162, specified and made readable #268 (comment)

Posted on behalf of @schickling

field	value
agent_identity	unknown
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.0.3
agent_runtime	OMP 18.0.3
tooling_profile	dotfiles@e4789b0

[schickling-assistant]

Post-deploy finding: PTY viewer redraw produces a false coarse active

Live real-seat reproduction on 2026-08-27 (dev3.ai-spend-janitor, semantically idle at the Codex prompt, latest user.agent.status={v:idle,reason:stop}):

1. With no viewer interaction for >60s, st2 returned fidelity=session,state=idle; Fractal rendered ○ idle.
2. Opening/selecting the session pane caused terminal attach/resize/redraw output and advanced lastOutputAtMs.
3. st2 returned session active; Fractal rendered the working spinner despite no new harness work.
4. Leaving the viewer open without interaction returned the same row to idle after the 60s window.

Conclusion: lastOutputAtMs is valid universal recent terminal output evidence, but it cannot alone make the semantic claim “harness in progress.” The existing downstream activity-fold.jq has the stronger composition: fresh braille spinner → exact user.agent.status → heuristic fallback. The PR is back to draft until the read-time fold consumes/defines semantic evidence (or exposes recent-output as a distinct non-working state) instead of mapping all recent output to active.

Posted on behalf of @schickling

field	value
agent_identity	unknown
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.0.3
agent_runtime	OMP 18.0.3
tooling_profile	dotfiles@e4789b0

[schickling-assistant]

Superseded by live observer-effect evidence and the restored single-owner semantic activity fold in dotfiles/fractal. lastOutputAtMs remains valid transport evidence, but selecting an idle session causes redraw output and cannot support the semantic active claim. Real Codex and Claude fixtures now classify correctly through user.agent.status plus the fresh-spinner heuristic; Fractal composes that projection with rich driver records.

Evidence: #353 (comment)

Posted on behalf of @schickling

field	value
agent_identity	unknown
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.0.3
agent_runtime	OMP 18.0.3
tooling_profile	dotfiles@e4789b0