State Probing

Status: v0.1 baseline draft — derived from the case catalog and axis definitions; awaiting trace evidence from Fable 5.1 case runs.

When to apply

• You are about to run a command, edit a file, or report a result whose correctness depends on the state of the environment (which version is installed, whether a process is up, what branch you are on).
• A step failed in a way that only makes sense if reality differs from your assumption ("but I installed it").
• You resumed a session, a build ran, or time passed — any earlier probe result may now be stale.
• You are inferring environment state from source, config, or memory instead of from the live system.
• You are about to claim something is true of the world ("the server is running", "the file exists", "we're on main") rather than something you did.
• The plan's next action is destructive or expensive and assumes a precondition you have not confirmed.

Core loop

1. Name the assumption as a checkable proposition. Turn the vague belief into one testable claim about the world: "node is v20", "port 5173 is listening", "the working tree is clean", ".env contains API_KEY". If you cannot phrase it as something a single command could confirm or deny, it is not yet a probe target — sharpen it first.

2. Pick the cheapest probe that discriminates. Choose the command whose output distinguishes the assumption from its negation with the least cost and least side effect. node --version beats reading a lockfile; git status --short beats git log; curl -sI localhost:PORT beats starting a client. A probe that cannot come back false is not a probe.

3. Run the probe and read the actual output. Execute it. Read the real bytes — exit code, version string, PID, branch name — not what you expected them to say. The point of probing is to be surprised; skimming for the expected answer defeats it.

4. Distinguish "absent" from "unknown". A probe that errored, timed out, or returned nothing is not evidence of absence. pgrep returning nothing means the process is not running or the name is wrong. Confirm the probe itself worked before trusting its silence.

5. Act only inside what the probe established. Take the next step only within the state you just confirmed. If the action's precondition was not the thing you probed, you have not earned the action — go back to step 1 for the real precondition.

6. Re-probe on any staleness signal. After a build, install, checkout, resume, long gap, or another agent's turn, treat prior probe results as expired. Re-run the probe rather than reusing a remembered value. State you confirmed 40 tool-calls ago is a memory, not a measurement.

7. Report the ground truth you observed, not the assumption you started with. State what the probe showed ("git status: 3 modified, on feature/x"), and if you could not probe, say the claim is unverified — never launder an assumption into a fact.

Heuristics

• Probe before every irreversible or expensive action whose precondition you have not confirmed this turn.
• The cheapest discriminating probe wins: prefer a one-line status query over spinning up the real client.
• One probe answers one proposition. If your command's output can't be false, it isn't testing anything.
• Version/tool claims cost one command (--version, which) — never assert a version from memory or a manifest.
• Treat "it worked a minute ago" as expired after any state-changing event; re-probe rather than recall.
• Empty output is ambiguous: confirm the probe ran (exit 0, correct target) before reading silence as "none".
• Prefer the runtime source of truth over its declaration: a running process over a config file, git status over what you think you committed, ls over the path you expect exists.
• If two probes disagree (lockfile says X, --version says Y), the live runtime wins and X is the bug.
• Budget the probing: 2–4 targeted probes to ground a task is discipline; probing every trivial fact is noise.
• When you cannot probe (no access, sandbox), downgrade every dependent claim to "assumed", not "confirmed".
• A probe that changes state is not a probe: prefer read-only queries; if the only check has side effects, note them.
• Ground the surprising result, not the expected one: when a probe confirms your belief, move on; when it contradicts it, that contradiction is the highest-value fact in the session — act on it first.

Anti-patterns

• Assumed-state action. Running the fix on the belief the service is up / the branch is right / the dep is installed, without a single confirming command. → Corrective: name the precondition and probe it before the action, not after it fails.
• Stale recall. Reusing a version, PID, or branch you checked long ago and several state-changes back. → Corrective: re-probe after any build/install/checkout/resume; a remembered value is not a measurement.
• Reading the declaration for the reality. Trusting package.json, a Dockerfile, or docs for what is actually installed and running. → Corrective: query the live runtime; let the declaration be the hypothesis.
• Silence as absence. Concluding "not running" / "not there" from an empty or errored probe without confirming the probe itself was valid. → Corrective: verify the probe targeted the right thing and exited cleanly before trusting its emptiness.
• Over-probing. Burning turns confirming facts that don't gate any action, or re-probing unchanged state every step. → Corrective: probe only what a decision depends on, and only after a staleness signal.

Relation to existing axes

• vs exploration-strategy: exploration builds a mental model of the code — what functions exist, how a request flows through the modules. State-probing grounds claims about the runtime/environment — what is installed, running, checked out, present on disk right now. Exploration reads files to understand; state-probing runs commands to confirm reality before acting on it.
• vs verification-discipline: verification comes after a change, to prove the result meets the claim. State-probing comes before an action, to confirm the preconditions the action assumes. One checks the output; the other checks the ground you are standing on.

Worked example

Illustrative construction, not a recorded run.

Task: "The deploy script keeps failing at the migration step — fix it." Autonomous session, resumed from a handoff.

Agent restates preconditions as propositions: (a) DB is reachable, (b) migration tool version matches
    the migration files, (c) we're on the branch the handoff described.
Agent picks cheapest discriminating probes: `git rev-parse --abbrev-ref HEAD`, `migrate --version`,
    `pg_isready -h $DB_HOST`.
Probe 1 → branch is `main`, but the handoff said work was on `release/2.3` → assumption (c) FALSE.
Agent stops: acting on main would migrate the wrong tree. Checks out `release/2.3`, then RE-PROBES version
    because the checkout is a staleness signal.
Probe 2 (post-checkout) → `migrate v4`, but migration files use v5 syntax → real cause found without ever
    editing the deploy script.
Agent confirms the probe wasn't lying: `which migrate` → an old global shim shadowing the project binary.
Exit: reports "root cause was environment, not the script — global `migrate` v4 shadowed project v5 on the
    wrong branch. Verified via `migrate --version` after checkout." No code changed on a false assumption.

Trace evidence

Case	Session	What it showed
awaiting case runs	—	—