Forbidden invocations

NEVER invoke /sdd-* skills from this workflow. SDD is an optional user-installed ceremony; this skill ships self-contained and does not chain SDD under any condition. If you need to refactor KATA, fixtures, cli/, scripts/, or api/schemas/ pipeline, exit this skill first and invoke /framework-development.

This boundary is mechanical, not advisory: scripts/lint-skills.ts rejects any /sdd- mention outside this section. See: .agents/skills/agentic-qa-core/references/skill-composition-strategy.md §4 (governs users who manually install SDD).

Exploratory Testing — Find Bugs, Discover Test Candidates, Without a Ticket

Freestyle but precise testing across a critical flow, a module, or the whole product: UI, API and DB together, never one alone. No ticket required. This skill hunts for real bugs in the highest-risk surfaces (data integrity and security first) and produces candidate scenarios that were executed live and checked against existing coverage, so test-documentation receives new work, not a second copy of what the repository already holds.

The method rests on four disciplines, and skipping any of them is how a plausible-looking anomaly becomes a wrong bug report: a parallel reconnaissance fan-out, a refutation pass whose job is to kill findings, a main-thread tie-break when verifiers disagree, and evidence-backed reporting. This repository adds the fifth: the coverage view, built before anyone explores, so every candidate says whether it is NEW, WEAK or already COVERED.

Position in the pipeline

This is the second entry into the Execution stage (agentic-qa-core/references/stage-gates.md §"The stages, by name"): same stage, entry by SCOPE instead of by ticket. It is the explore and confirm steps of the discovery-driven flow (agentic-qa-core/references/discovery-driven-doctrine.md §1).

project-context coverage (gap list) --+
                                      v
                       [exploratory-testing]
                         |            |             |
                         v            v             v
            confirmed bugs   candidate scenarios    behaviour with no Story
            (person signs,   NEW / WEAK, oracle     -> coverage Tech Story
            sprint-testing   confirmed -> ROI in    (person signs; doctrine §3)
            retests later)   test-documentation

Dependencies

Requires agentic-qa-core. Loads on demand:

• agentic-qa-core/references/adversarial-verification.md: MANDATORY before the Refute and Tie-break steps.
• agentic-qa-core/references/discovery-driven-doctrine.md: MANDATORY before marking any candidate (the oracle rule §4, the outcome vocabulary and blocking ladder §5, the coverage anchor §3, risk vs ROI §7).
• agentic-qa-core/references/defect-management-doctrine.md: MANDATORY before proposing any Bug / Defect / Improvement.
• agentic-qa-core/references/test-design-doctrine.md: techniques applied to the code, the schema and the observed behaviour instead of an AC (Part 2).
• agentic-qa-core/references/briefing-template.md, agentic-qa-core/references/dispatch-patterns.md, agentic-qa-core/references/orchestration-doctrine.md, agentic-qa-core/references/session-management.md, agentic-qa-core/references/preflight-gate.md, agentic-qa-core/references/decision-elicitation-doctrine.md: cited inline by the sections that use them.

Compact Rules

Exploration operational rules (binding):

• Ticket-free entry into the Execution stage: the scope is a flow, a module or a sweep, never a ticket key (a named ticket routes to sprint-testing). Mode from the first $ARGUMENTS token (flow | module | sweep), or from a request that clearly names one flow, one module or a sweep (args-as-answers); otherwise ASK.
• Five steps, always in order: Charter (main thread) -> Recon (Parallel) -> Refute (Parallel) -> Tie-break (main thread) -> Handoff (main thread). Planning and tie-breaks are never delegated. The charter is presented and confirmed before any Recon dispatch, unless the request already named the scope and said go.
• Build the coverage view BEFORE Recon: the gap list of project-context mode coverage when it exists, else coverage-view.json from kata-manifest.json + the TMS Tests read by bun run context:coverage collect (never by jira:sync-issues: Tests are sync: never); a TMS not read is stated TMS not read and no manual cover counts. Every candidate is marked NEW / WEAK / COVERED; only a counted exercises cover makes COVERED; a weak, stale or unverified cover never dedupes a candidate away.
• Compare at least 2-3 rows plus a control group before calling anything anomalous, and state the detection query verbatim.

Verification doctrine (binding — full canon: agentic-qa-core/references/adversarial-verification.md):

• Refute mandate: every Medium+ anomaly gets a Parallel verifier told to REFUTE it (join fan-out, data quirk, intended behaviour, harness artifact); it survives only if it cannot be explained away. Check a row's own invariant, not just that it differs from another row; execute competing hypotheses side by side, never infer them.
• Verifier disagreement goes to a main-thread tie-break: a decisive query plus a direct read of the live code path. Never a third verifier, never an average, never one agent's verdict alone on a data-integrity, money or security finding.

Oracle and outcomes (binding — full canon: agentic-qa-core/references/discovery-driven-doctrine.md §4, §5):

• Oracle: behaviour with no AC is oracle: observed; "is this intended?" goes to a person (decision-protocol kind 1), a batch through the mkd deck, a single question through the question tool. oracle: confirmed (who, where, when) makes the observation PASSED; not intended makes it FAILED and a proposed defect.
• A candidate is PASSED or FAILED only when observed in the environment its precondition names; a code read, unit test, green CI or another environment leaves it NOT RUN with exactly one reason, BLOCKED or DEFERRED (execution), and it climbs the blocking ladder; with no coverable yet, its follow-up lives on the coverage Tech Story once approved, and until then as an item of the Handoff deck.

Safety and signatures (binding):

• Never call anything that sends, notifies or pushes downstream; reads and previews are fine. A project's own list (## Never-call endpoints in agent-project.md) is carried verbatim in briefing component 7.
• Mutations only to set up a precondition, within the charter's budget. Every created row goes into created_rows the moment it exists; cleanup re-reads each one (a delete's response code is not proof).
• Person-signed writes: every bug (sprint-testing/references/reporting-templates.md §1, gate §1.12, classified per defect-management-doctrine.md Part 1), every coverage Tech Story (discovery-driven-doctrine.md §3), every module-context write. This skill creates no Test and scores no ROI.
• Evidence that reaches a person goes through the evidence-artifact skill; raw captures stay in the session evidence/ until a coverable exists.
• On any subagent failure: STOP, report which surface or finding failed, offer retry / skip-surface / abort; a skipped surface is recorded "not covered this round", never dropped silently. A skill that itself broke is reported upstream per agentic-qa-core/references/upstream-feedback.md.

Read full SKILL.md when: starting a round cold, resuming a session, or handling a sweep, a tie-break or a Handoff not covered by the rules above.

Inputs — read these first, in this order

1. .agents/project.yaml: environments, testing.default_env, testing.tc_creation_stage, the backend and frontend repositories, orchestration.max_workers, qa.qa_epics.
2. The gap list of project-context mode coverage (<<PRIMARY_ROOT>>/.session/project-context/coverage/gap-list.json; schema: project-context/references/coverage.md §"gap-list.json"). Optional: without it, the Charter step builds the coverage view itself.
3. kata-manifest.json: the ATCs already automated (Critical Rule #12).
4. .context/PBI/qa-artifacts/master-test-plan.md: the risk level of each flow, which orders the queue.
5. The business maps, read with bun run context:map <slug>: business-e2e-context first, then business-api-context and business-data-context per touched level; business-domain-context for the vocabulary of every scenario title.
6. The module's synced module-context.md under .context/PBI/epics/ (fixtures, gotchas, trap instances, prior rounds).
7. agent-project.md ## Never-call endpoints, when the project has written one.

Subagent Dispatch Strategy

Orchestration & Session contracts: this skill follows agentic-qa-core/references/orchestration-doctrine.md (mandatory subagent dispatch — main thread is command center) AND agentic-qa-core/references/session-management.md (Phase 0 resume check, plan-first persistence at .session/<skill-slug>/<scope>/, archive on completion). Phase 0 (resume check) and Phase 1 (plan write) are NOT optional.

This skill is per-round scope: <scope> = <YYYY-MM-DD>-<scope-slug> (e.g. 2026-05-20-orders, 2026-05-20-checkout-flow, 2026-05-20-sweep). Session state lives at <<PRIMARY_ROOT>>/.session/exploratory-testing/<scope>/ per agentic-qa-core/references/session-management.md §3 + §9; its files are listed in references/handoff-and-reporting.md. The plan write is the charter: plan.md follows the §6 schema and points at charter.md.

Every dispatch follows the 7-component briefing (agentic-qa-core/references/briefing-template.md); the pattern per step follows agentic-qa-core/references/dispatch-patterns.md. The orchestrator verifies each step's DoD (agentic-qa-core/references/stage-gates.md §"Execution (exploration without a ticket)") before recording its checkpoint and advancing.

Why Parallel for Recon and Refute: each surface and each finding is independent: no shared state, no ordering. A sweep of many modules run serially would take hours; the fan-out is the point.

Why the main thread for Charter and Tie-break: the charter is planning; the tie-break exists because a single agent's verdict on a high-cost finding is not trustworthy enough to be final.

On any subagent failure: STOP, report the partial state (which surfaces and findings completed, which rows were created), present retry / skip-surface / abort. A skipped surface is recorded in session.md as "not covered this round", never silently omitted. No auto-fix, no auto-rollback. A skill that itself broke (a wrong step, a missing verifier, a stale rule) is reported upstream per ../agentic-qa-core/references/upstream-feedback.md: drafted and redacted locally, filed only on explicit OK, verified with gh issue view.

Fleet seam (optional)

Recon runs on in-turn subagents by default. A sweep large enough to outgrow one turn may run one persistent worker per module instead, only when the user asks for it:

• Topology: same checkout. Recon writes nothing to the repository, only the session directory, one recon/<surface>.json per worker. No worktree is needed.
• The conductor writes launch.txt in the session directory, one self-contained line per module, always. Launching and supervising those sessions is orca-orchestration/SKILL.md ([ORCHESTRATION_TOOL]); launch.txt is the payload for the human-paste fallback.
• Refute, Tie-break and Handoff stay with the conductor. Workers return Recon reports and their created_rows; they never file, never ask the person, never tie-break.
• Silence rule: the absence of an orchestration transport is never named to the user and never appears in the charter or the session record.

Readiness Preflight Gate (MANDATORY — runs before Phase 0)

Full doctrine: agentic-qa-core/references/preflight-gate.md. Two laws: (1) args-as-answers: the mode, the scope and a named environment are provided args; ask only the gaps. (2) probe, don't assume. Surface gaps and REDs as ONE checklist question; STOP on a blocking RED. The generic baseline (environment, test-user credentials by name, secret and restart handling) is inherited from §3.1. Below is only this skill's delta.

Phase 0 — Session resume check + Session Init

0.0 Resume check (agentic-qa-core/references/session-management.md §4). Compute <scope> = <YYYY-MM-DD>-<scope-slug>. Check <<PRIMARY_ROOT>>/.session/exploratory-testing/<scope>/progress.md (from a worktree, say so in one line and look in the primary). If it exists, read plan.md, charter.md and the tail of progress.md, surface the last completed step and the next one, and offer resume / restart / abort. On restart, archive the directory to .session/.archive/<YYYY-MM-DD>-exploratory-testing-<scope>-aborted/ first.

0.1 Mode. The first $ARGUMENTS token flow, module or sweep IS the mode and the rest is forwarded; a request that clearly names one flow, one module or a whole-product sweep is the answer too (args-as-answers); a ticket key routes to sprint-testing; otherwise ask the mode question (references/scope-and-recon-planning.md §1).

0.2 Create the session directory, write plan.md (§6 schema, the ## Inputs H2 names the coverage view source and the environment) and append ## Phase 0 — Session Init — <ISO-8601 UTC> with next: Charter to progress.md.

Workflow — five steps, always in this order

Charter -- mode + scope source + charter.md + coverage view + recon plan + env gate
    -> PRESENT the charter, WAIT for confirmation            references/scope-and-recon-planning.md
Recon (Parallel) -- one agent per surface: DB + API [+ UI], Trifuerza technique
    -> structured schema incl. observed[] with covered_by and created_rows
    -> compare 2-3 rows + a control group before any anomaly references/recon-fanout.md
Refute (Parallel) -- one verifier per Medium+ anomaly, mandate REFUTE
    -> own invariant, competing hypotheses executed           agentic-qa-core/references/adversarial-verification.md
Tie-break (main thread) -- only where verifiers disagree
    -> decisive query + direct read of the live code path     adversarial-verification.md §4
Handoff (main thread) -- findings/FNN-*.md (confirmed AND refuted)
    -> candidate-scenarios.md, each NEW / WEAK / COVERED, oracle: observed
    -> ONE deck for the person: oracle answers, bugs to file, Tech Stories, module context
    -> approved writes, cleanup re-read, session.md, footer, archive
                                                              references/handoff-and-reporting.md

Every step closes with its checkpoint in progress.md. A step that returns a partial result is NEEDS REVISION, never advanced (stage-gates.md §"The gate rule").

Handoff decisions, by weight

The instrument follows the weight of the decisions, never their count (AGENTS.md §2). A round's Handoff is usually a report the person reacts to point by point, each item with its own option set: the oracle answer per candidate, file or discard per finding, the grain and approval of each Tech Story, the module context harvest. That report goes into one mkd decision deck (agentic-qa-core/references/decision-elicitation-doctrine.md); a round with a single small item asks it through the question tool. A blocking item (a row that cannot be cleaned up, a credential failure) is asked at once, never held for the deck.

Gotchas — apply every invocation

1. A weak cover is not coverage. A test that runs through a behaviour without checking it leaves the candidate WEAK, and WEAK is handed off. Only a counted exercises cover makes COVERED.
2. Observed is not intended. Never write what the product did as the expected result of a scenario before a person confirms it; that freezes whatever defect was observed (doctrine §4).
3. Compare 2-3 rows before calling an anomaly. One row, or a loose GROUP BY, is a lead (references/recon-fanout.md §2).
4. Check the row's own invariant. Two rows differing proves nothing; a row that breaks its own formula and matches another row's value is a copy (adversarial-verification.md §2).
5. Execute the simulation, never infer it. Same rule as sprint-testing/references/exploration-patterns.md and doctrine §5.2.
6. Size by reachability. A defect in a hidden or unlinked route is real, but its severity reflects who can reach it: map endpoint -> handler -> caller -> navigation entry first (adversarial-verification.md §4, §5).
7. Sparse environment data breaks "regressed this sprint". Say so in the finding instead of guessing at recency.
8. Read the live schema before a query leans on a column name: project docs drift.
9. Language: every artifact (charter, findings, candidate scenarios, bug drafts, commits) in English; the conversation mirrors the user.
10. Session footer, chat only: capture paths the instant they exist and again at close, plus skills, MCPs, CLIs and testing levels, execution framing (agentic-qa-core/references/session-footer-contract.md).

Skills to load

Cross-skill handoff — what this skill does NOT do

This skill produces no ATP, ATR or TC, scores no ROI and writes no test code. A named ticket routes to sprint-testing.

Pseudocode tags used here

References — read the narrow one for the situation

Pre-flight checklist

• Mode picked (Critical-Flow-Targeted / Module-Targeted / Whole-App Sweep), not a ticket key
• module-context.md read for every in-scope module
• Coverage view built (gap list run_id, or coverage-view.json, with TMS not read stated when it was not) and passed to every Recon agent
• Charter written and confirmed; environment gate and tool probe passed
• Recon dispatched Parallel, one agent per surface, every report in the schema (created_rows even if empty)
• Every Medium+ anomaly refuted or survived; every disagreement settled by the tie-break
• One findings/FNN-*.md per finding, confirmed and refuted
• Every candidate marked NEW / WEAK / COVERED and carrying its outcome and oracle state
• The Handoff decisions asked by weight; only approved writes made, each read back
• Module-context harvest asked for every touched module, or recorded as skipped ("nothing cleared the bar")
• Every created_rows entry re-read gone
• session.md written; session footer printed (execution framing); Refinements proposed: line present