docs(vrs): record observed harness state experiments

[schickling]

Why

st2 cannot see what a harness is doing. Presence is agent-declared — status::refresh proves only that a process is alive and preserves whatever the model last wrote — so a wedged agent and a working one are indistinguishable for up to fifteen minutes. Meanwhile CodexObservedState (src/codex_app_server.rs:166) already models exactly this concept for one harness, typed and persisted under st2.codex-control-state.v1, in a hash-keyed host-local file that nothing outside that module reads. This is the activity-lease slice of #162 — "activity leases (idle, active, child, unknown)" — measured before being specified, as DQ3 requires.

These are records of what was measured, not a specification. No requirements text, no ontology entry, no code.

What

• Record that Codex already carries blocked-on-human on the wire: ThreadStatus.active has a required activeFlags field, and observe_thread_status (:770) reads only the status word.
• Record that Claude can prove the same signal: the binary carries 31 hook events against 10 documented, and PermissionRequest fires only when a human is about to be asked — in 2 of 9 captures, and in zero bypassPermissions/auto runs.
• Record that Claude death is silent for SIGTERM and SIGKILL alike, and that a wrapper-owned terminal write does not close the gap, because stop_provider_group escalates into the wrapper's own process group.
• Record that a shared append-only history conflicts under Git while host-named segments do not, and that retention disagreement between hosts resolves to the minimum.
• Record the storage question as unresolved: the transport was not identified, so no cost model is claimed.
• Propose 0004, asserting only the half that survived independent adversarial review: observed state is inadmissible as transport authorization.

Impact

Nothing changes. This adds five documents and their captures under docs/vrs/, and nothing else.

• No source file is modified. No requirements, ontology, or spec text is touched.
• INVARIANTS.md is unchanged; no invariant is claimed, relaxed, or added.
• No Agent Spec field is proposed. Agent Spec authoring authority remains with compoundingtech/evals.
• 0004 is Status: proposed and explicitly says it is not accepted.

Checks

git diff --check origin/main..HEAD
  → trailing whitespace reported ONLY in docs/vrs/.experiments/2026-08-17-captures/**/*.raw
    These are verbatim PTY screen captures; the trailing bytes are part of the evidence.

nix develop --command cargo test --locked --test invariants
  → test qualified_proof_references_resolve ... ok
    test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 0 filtered out

No Darwin run: no code changed.

Limits, stated up front

• activeFlags is unverified on a supported version. Corrected 2026-08-18: schemas generated from 0.145.0 and 0.146.0 binaries — both already admitted by SUPPORTED_CODEX_CLI_VERSIONS — carry ActiveThreadStatus with required: ["activeFlags","type"] and an identical two-value ThreadActiveFlag. The field is available on a supported version today and no finding here depends on admitting 0.147.0. What stands: a populated activeFlags has never been observed on the wire on any version — the live capture confirmed only "activeFlags":[] before the account hit its usage limit — so the steer-into-approval consequence remains inferred, not observed.
• The prototype PTY observer is broken for Claude and is not proposed here. expected = "" collides with the Claude arm's representation of an empty composer, so it proves idle once and reports unproven thereafter.
• DQ3's supervisor-following half is unproven. Only stale-state behavior was measured.
• The CodexHoldReason mapping is lossy, not residue-free: four of six rows report unknown, two of which Codex positively reported as active.

Ready for review.

Related

• Observed harness state: the activity-lease slice of #162, specified and made readable #268 — the proposal these records support, opened as a child of Define Axe harness drivers and extract all harness-specific behavior from st2 core #162. This PR is the measurement it rests on; DQ3 requires that ordering.
• A failed Codex turn is reported as two conflicting turns and stops accepting messages until the next idle #264, observe_thread_status takes only the status word, so a required ThreadStatus field is parsed and discarded at all three call sites #265, exitedReviewMode is never matched, so Held{Review} clears only when the thread reports idle #266 — shipped-code defects found while measuring. Independent of the proposal.
• st2 refuses the locally installed codex-cli, and the version pin is now behind by one release #267 — the version pin. All Codex wire evidence here comes from a release st2 currently refuses to launch.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.

[schickling-assistant]

Superseded by the merged observed-harness-state implementation (#319–#326). The experimental conclusions are now represented by the accepted architecture.

Posted on behalf of @schickling

field	value
agent_identity	dev3.direct.omp.jmg9x29x
session	dev3.jmg9x29x
agent_persona	generalist
agent_supervisor	unavailable
agent_tool	OMP
agent_tool_version	18.0.11
agent_runtime	OMP 18.0.11
tooling_profile	dotfiles@2161b9c