consult-design-thinking

Per-Deliverable Design Thinking

Produce one deliverable by walking it through its own design-thinking loop: empathize → define → ideate → prototype → test. The loop is scoped to a single deliverable inside one action field — there is no engagement-level phase machine. The stage methods live in $CLAUDE_PLUGIN_ROOT/references/methods/ (empathy-mapping.md, hmw-synthesis.md, guided-ideation.md); this skill owns the conversation flow, the artifact, and the state writes. Schemas: $CLAUDE_PLUGIN_ROOT/references/data-model.md.

The framework lens

Each deliverable's field.json entry carries chosen_framework — the structuring framework its argument takes, set once at creation and read-only thereafter (schema in $CLAUDE_PLUGIN_ROOT/references/data-model.md). Its value is one of: a stable slug from $CLAUDE_PLUGIN_ROOT/references/frameworks-registry.md (e.g. pyramid-principle), a "combo:<slugA>+<slugB>" pairing of two, or null.

The Prerequisite Gate (step 1) already reads field.json, so the value is in hand — no extra read is needed. When it is non-null, resolve its one-line Structure signature from the registry's framework table (for a combo:, resolve both and apply them together) and carry that signature into the Define and Prototype stages below: it shapes how the problem is framed and how the artifact body is organized. The registry is thin by design — it pins the signature and the stable key; supply the framework's depth at runtime: where the registry slug cell links to a first-party page, follow it for the framework's substance; otherwise supply that substance from your own knowledge.

When chosen_framework is null — a legacy deliverable created before a framework was chosen, or a deliberate no-framework choice — run those stages exactly as written below, with no framework structuring. The framework only adds shape when present; it never blocks production when absent.

Workflow

Interaction mode

By default this loop runs as an auto-walk: it writes each stage's artifact and log entry as it goes, surfacing the resulting entries — fast, and right for an obvious-shape deliverable. A consultant who wants to steer the work can instead opt into interactive mode (also transparent mode): trigger it at loop entry with a phrase such as "run it interactively", "transparent mode", or "with confirmation gates", or set it as a per-engagement default the consultant states at the start. Interaction mode is ephemeral — it governs only this session's conversation flow and is never written to field.json, consult-project.json, or any log; the state-write ownership and the logging contract below are identical in both modes.

In interactive mode, at each stage gate marked Interactive mode below (start of Empathize, the Empathize pre-write merge confirmation, end of Define, end of Ideate, start of Prototype, Test), before the stage's write: surface the reasoning behind the stage decision (why this HMW spec, why this ideation method, why this prototype shape, why these persona dispositions) ahead of the log entry, name exactly what is about to be written, and pause for the consultant's explicit confirmation — revise on feedback, then write. The start-of-Empathize gate is the exception in kind: it pauses to gather input material to ground the deliverable rather than to narrate a write (see Empathize below). In auto-walk mode, skip the pauses and proceed directly through each gate. The gates only add confirmation seams and reasoning narration around the existing writes; they never change what gets written or who owns it.

Advancing the stage

Every dt_stage boundary below — in both auto-walk and interactive mode — moves the stage through the guarded helper rather than a free-text Edit of field.json. The helper validates the transition (rejects an unknown stage name and a forward jump that skips a stage), writes dt_stage via an idempotent read-modify-write, and appends a per-stage move to .metadata/stage-log.json so the loop's iterations leave a trail (the execution log keeps recording state transitions only, never per-stage moves):

bash $CLAUDE_PLUGIN_ROOT/scripts/dt-stage-advance.sh \
    <engagement-dir> <field-slug> <deliverable-slug> <target-stage>

A single-step forward advance, an idempotent same-stage re-set, and a re-entry to any earlier stage (the loop may iterate) are all permitted; a forward jump that skips a stage is refused with success: false. The helper degrades gracefully on a legacy field.json whose deliverable has no dt_stage (it logs the move with from: null). On a success: false (a refused jump, a missing deliverable, an unreadable manifest) do not proceed past the boundary — surface the error and resolve it before retrying, since an unwritten dt_stage leaves the loop state inconsistent. Should the helper be absent (an older install), fall back to a free-text Edit of field.json setting dt_stage directly. Where a stage below says "advance dt_stage → "X"", run this helper with <target-stage> X.

1. Prerequisite Gate

When arriving via an in-session handoff (e.g. from a WBS dashboard recommendation), the engagement directory and target deliverable are already known — skip discovery. Otherwise locate the engagement:

bash $CLAUDE_PLUGIN_ROOT/scripts/discover-projects.sh --json

and confirm the intended engagement with the user when more than one is registered. When discovery returns zero engagements, treat it the same as a missing consult-project.json.

Read <engagement-dir>/consult-project.json. Branch explicitly:

• If it is missing (or discovery returned zero engagements): dispatch Skill("cogni-consult:consult-setup") and stop — write nothing.
• If workflow_state.scope is not "complete": dispatch Skill("cogni-consult:consult-scope") and stop — write nothing. The WBS must exist before deliverable work starts.

Then identify the target deliverable: the consultant names it, or pick the recommendation handed in. Read the field's action-fields/<field-slug>/field.json and confirm the deliverable entry exists (slug, title, state, dt_stage). If the entry is missing, stop — this skill produces deliverables, it never invents manifest entries. Dispatch Skill("cogni-consult:consult-action-fields") to plan the field's deliverable set (it writes the full planned-entry shape, including producing_route and persona_review), then resume here with the planned entry.

Conduct the conversation in the resolved interaction language (workspace default, overridden by the user's message language) — independent of the engagement's language field, which is the deliverable axis. See $CLAUDE_PLUGIN_ROOT/references/interaction-language.md.

Personas gate (fresh starts only). Before opening the loop for a deliverable whose state is pending (a fresh start — not a resume or rework), the engagement's personas gate must be satisfied, so Empathize (which maps personas) and Test (which challenges as personas) operate on real stakeholders rather than the degraded fallback. Read the derived rollup — this gate does not otherwise call it:

bash $CLAUDE_PLUGIN_ROOT/scripts/engagement-status.sh <engagement-dir>

and branch on data.personas_gate:

• If "pending" (no source: "scope-seeded" persona and no personas/.gate-waiver marker — the seeded default advisors alone do not satisfy it): do not open the loop. Dispatch Skill("cogni-consult:consult-personas") and stop — write nothing. Explain that the engagement's personas gate is unsatisfied, and that it flips to satisfied once scope-specific stakeholders are seeded (define mode) or an explicit waiver is recorded (for engagements with no external stakeholders); the consultant returns here once it is satisfied.
• If "satisfied": proceed to Open the Loop as usual.

This guard applies only to the pending (fresh-start) branch below. A deliverable that is already in-progress (resume) or being reopened from complete (rework) is never blocked — the gate is a start-of-loop check, and once satisfied it stays satisfied, so in practice it only ever gates the engagement's first deliverable. The Empathize/Test fallbacks remain as a last-resort safety net but do not trigger for the scope-seeded flow.

2. Open the Loop

If the deliverable's state is pending: one Edit of field.json sets it to "in-progress", advance dt_stage → "empathize" via the helper (see Advancing the stage above), and one Edit of .metadata/execution-log.json appends to its transitions[] array the entry {"action_field": "<field-slug>", "deliverable": "<deliverable-slug>", "from": "pending", "to": "in-progress", "timestamp": "<ISO>", "triggered_by": "consult-design-thinking"}.

If the deliverable is already in-progress, resume at its current dt_stage — re-entering an earlier stage is normal (the loop may iterate); just keep dt_stage honest at each boundary.

If the deliverable is complete and the consultant wants rework ("continue the deliverable", a revision request), confirm the re-entry first, then one Edit of field.json sets state back to "in-progress", advance dt_stage to the stage the rework needs (often define or ideate) via the helper (see Advancing the stage above — a re-entry to an earlier stage is permitted), and one Edit of .metadata/execution-log.json appends the complete → in-progress transition to transitions[].

Because the deliverable's content is about to change, flag its downstream dependents stale now — before the rework begins — so the consultant sees the blast radius up front:

python3 $CLAUDE_PLUGIN_ROOT/scripts/deliverable-graph.py <engagement-dir> \
    cascade-stale <field-slug>/<deliverable-slug> --trigger deliverable_update

Surface data.newly_flagged (the deliverables now marked stale). A "success": false (node not found, bad dir) is non-blocking — surface the error and proceed with the rework.

3. Empathize

The empathize-stage per-persona empathy mapping runs as a read-only fan-out: one consult-empathy-mapper dispatch per relevant personas/*.json, merged and written by this stage (the agent never writes). It embodies $CLAUDE_PLUGIN_ROOT/references/methods/empathy-mapping.md; the full fan-out, envelope, merge, write, and idempotency contract is authoritative in $CLAUDE_PLUGIN_ROOT/references/orchestration/empathize-empathy-mapping.md. Define and Ideate stay fully inline — only this per-persona mapping is delegated. When personas/ is empty, say so and continue with the consultant's own stakeholder knowledge — persona files can be added later without redoing the loop.

Interactive mode — input material first. Before the gap-check, run the Empathize source-material intake rung — the pre-gap-check intake of consultant-supplied files, pasted text, and URLs, the ingest-into-bound-base vs. read-direct-first-party sink choice, and its re-entry idempotency — per $CLAUDE_PLUGIN_ROOT/references/orchestration/empathize-intake.md. In auto-walk mode, skip this prompt and proceed directly to the gap-check on the scope-time material.

Research for this stage follows the Research Routing Rule in $CLAUDE_PLUGIN_ROOT/references/research-routing.md — the canonical contract for every research run in the engagement. Start with the gap-check rung: dispatch Skill("cogni-knowledge:knowledge-query") with --knowledge-slug <plugin_refs.knowledge_base> for the deliverable's topic. Record each gap-check per the Gap-Check Recording contract in the Research Routing Rule — append one entry to .metadata/decision-log.json's decisions[] array tagged "kind": "gap-check", carrying the verbatim question plus the coverage outcome as discrete keys (never fold the verdict or overlap scores into a prose decision string): {"id": "d-NNN", "kind": "gap-check", "action_field": ..., "deliverable": ..., "question": "<verbatim --question>", "theme_label": <label-or-null>, "verdict": "covered"|"partial"|"uncovered", "top_hit": "<page-slug>"|null, "top_score": <score>|null, "timestamp": ...}. Use kind (not type) and emit verdict/top_hit/top_score as their own keys — no decision prose, no evidence_refs — so gap-checks stay filterable and the routing decision replays programmatically. When the base has no coverage, escalate to the full inverted pipeline (or the --source wiki re-run on a populated base) per the rule, and copy the finalized synthesis to action-fields/<field-slug>/research/<topic-slug>.md so this deliverable — and later ones — find it at a stable path. Evidence comes from the knowledge base, never from raw web search.

Run the fan-out. With the evidence context in hand from the rung above, dispatch consult-empathy-mapper once per relevant personas/*.json (inputs engagement_dir, field_slug, deliverable_slug, persona_slug, plugin_root, and evidence_refs — the research-synthesis and prior-deliverable paths gathered above). Merge the success: true envelopes and apply this stage's write contract per $CLAUDE_PLUGIN_ROOT/references/orchestration/empathize-empathy-mapping.md: per persona, Edit personas/<slug>.json to populate empathy_map and needs, promote maturity to "researched" when the envelope recommends it, and append one idempotent empathy-mapped work_log entry keyed by (action_field, deliverable). Surface the cross-persona overlaps and tensions — they feed the Define spec. Interactive mode: before these writes, present the merged maps and each persona's key insight, confirm with the consultant, then write.

Close the stage by advancing dt_stage → "define" via the helper (see Advancing the stage above).

4. Define

Read $CLAUDE_PLUGIN_ROOT/references/methods/hmw-synthesis.md and sharpen the deliverable's problem spec from the empathize outputs plus the field's framing. Lock 1-3 HMW questions with the consultant. When a framework lens is in play (see The framework lens above), let its Structure signature frame how you organize the problem and the approach — e.g. a mece-issue-tree signature decomposes the problem into mutually-exclusive, collectively- exhaustive branches before drafting, while a pyramid-principle signature pushes you to lead with the answer and group the supporting arguments beneath it. For a combo: choice, let the first signature frame the problem and the second the approach. When chosen_framework is null, frame the spec as you would by default. When sharpening the spec surfaces an evidence gap (an assumption the consultant cannot ground), route the research per $CLAUDE_PLUGIN_ROOT/references/research-routing.md before locking — a spec built on an unverified assumption fails at the test stage anyway.

Interactive mode: before logging, present the locked 1-3 HMW questions and the framing reasoning that produced them (including how the framework lens, if any, shaped them), and confirm them with the consultant; revise until accepted, then write.

Append the locked spec to .metadata/decision-log.json's decisions[] array as a decision ({"id": "d-NNN", "action_field": ..., "deliverable": ..., "decision": "<locked problem framing>", "rationale": ..., "evidence_refs": [...], "timestamp": ...}). Then advance dt_stage → "ideate" via the helper.

5. Ideate

Read $CLAUDE_PLUGIN_ROOT/references/methods/guided-ideation.md and run the diverge→cluster→converge→sketch flow against the locked spec. Keep it proportionate — a deliverable with an obvious shape needs one quick pass, not a full workshop.

Interactive mode: before logging, surface the chosen ideation method and why it fits this spec (proportionate to the deliverable's shape — one quick pass vs. a full workshop), and confirm it with the consultant before writing.

Append the method selection to .metadata/method-log.json's methods[] array ({"action_field": ..., "deliverable": ..., "proposed": [...], "selected": [...], "rationale": ...}). Then advance dt_stage → "prototype" via the helper.

6. Prototype

Interactive mode: before drafting, surface the prototype direction — the chosen approach and how the framework lens (if any) will shape the artifact body — and confirm it with the consultant; revise the direction on feedback, then draft.

Draft the deliverable artifact at action-fields/<field-slug>/<deliverable-slug>.md — Obsidian markdown with YAML frontmatter exactly per the data model: slug, action_field, sources[] (each entry the lineage triple source_url, entity_ref, propagated_at, plus kb_ref when the claim came from the knowledge base — or, for read-direct first-party material carried in from the Empathize intake rung, a file:// source_url with evidence_class: first-party and no kb_ref; canonical shape in $CLAUDE_PLUGIN_ROOT/references/data-model.md), and updated. State is intentionally absent from the frontmatter — it lives in field.json only.

Structure the body from the loop's outputs: the problem (define), options considered (ideate), the chosen approach, and the content itself. When a framework lens is in play (see The framework lens above), organize the artifact body to its Structure signature rather than that default outline — e.g. pyramid-principle → lead with the answer, then MECE-grouped supporting arguments; scqa → Situation → Complication → Question → Answer; journey-process → sequential stages along the path. For a combo: choice, apply both signatures together (typically one frames the opening, the other the body). When chosen_framework is null, use the default outline above. Every evidence-backed claim carries a sources[] entry. Then advance dt_stage → "test" via the helper.

7. Test

Interactive mode: before challenging, surface the persona dispositions you intend to apply — which personas will challenge (per the relevance rule in $CLAUDE_PLUGIN_ROOT/references/orchestration/test-persona-challenge.md: shipped advisors plus context-matching personas) and the objections each is likely to raise — so the consultant can steer the challenge before it runs; then proceed.

Challenge the draft as the stakeholder personas by delegating to the write-contract owner — do not reimplement the persona-challenge writes inline here. Dispatch Skill("cogni-consult:consult-personas") in challenge mode, naming this deliverable (its artifact path under action-fields/<field-slug>/) as the in-session handoff. consult-personas fans out the per-persona in-voice objections (one read-only consult-persona-challenger dispatch per relevant personas/*.json), merges the returned {success, data, error} envelopes, and owns the single append-work_log / append-## Persona Challenges / advance-persona_review write contract — so that contract lives in exactly one place. The full fan-out, merge, idempotency, and zero-personas fallback contract is authoritative in $CLAUDE_PLUGIN_ROOT/references/orchestration/test-persona-challenge.md. Revise the artifact where a challenge lands; the challenge is advisory and never blocks completion — the consultant decides what to revise.

If the draft survives (consultant accepts): first record the deliverable's evidence provenance per $CLAUDE_PLUGIN_ROOT/references/orchestration/test-provenance-gate.md — no deliverable completes without a provenance record: reuse the Empathize gap-check decision-log entry when present, otherwise append an evidence-provenance-waiver, and set the resulting evidence_class on the field.json deliverable entry. Then, when the deliverable's chosen_framework is non-null (already in hand from the Prerequisite Gate), run the framework-adherence review per $CLAUDE_PLUGIN_ROOT/references/orchestration/test-adherence-review.md — dispatch the read-only consult-framework-adherence-reviewer agent (inputs engagement_dir, field_slug, deliverable_slug, plugin_root), surface its advisory adherence band and drift findings to the consultant, and record the outcome as an adherence-review decision-log entry keyed by (action_field, deliverable). This is the framework-adherence rung of the Three-Layer Quality Gate — advisory, never blocking (auto-walk proceeds to completion; interactive mode may elect to loop back). A null-framework deliverable skips it entirely and completes with no adherence stop.

Next, the advisory language-polish rung — the register-quality tier of the Three-Layer Quality Gate, alongside the framework-adherence rung above. Dispatch Skill("cogni-copywriting:copywriter") over the deliverable artifact with FILE_PATH=<engagement-dir>/action-fields/<field-slug>/<deliverable-slug>.md and --scope=tone — register-only, so tone scope skips restructuring and the framework structure is not re-imposed; only the prose register is polished (clause length, Floskel and Denglish reduction per the copywriter's German style rules). The copywriter natively freezes every {{asm:<slug>}} placeholder and the ## Persona Challenges table byte-identical and excludes the frontmatter sources[] lineage, so the tone pass cannot corrupt a resolver-critical token. The copywriter writes the polish in place and backs the original up to <engagement-dir>/action-fields/<field-slug>/.<deliverable-slug>.md (its dotfile backup), so surface the resulting diff to the consultant after the write and let them accept or reject: on reject, restore the deliverable from that backup dotfile so a rejected polish never lingers on disk; on accept (and on skip/error, where no polish was written) the backup dotfile is redundant and may be removed — the copywriter overwrites it on the next polish run either way. The chosen_framework headings are the accept/reject gate's safeguard against any heading rewording. Record the outcome as a copywriter-polish decision-log entry keyed by (action_field, deliverable) for parity with the adherence-review rung (shape: {"id": "d-NNN", "kind": "copywriter-polish", "action_field": <field-slug>, "deliverable": <deliverable-slug>, "outcome": "accepted" | "rejected" | "skipped" | "error", "timestamp": <iso8601>}). This rung is advisory and strictly non-blocking: whether the polish runs, is skipped, is rejected, or errors, the DT loop proceeds to completion unchanged — a language pass must never stall the flow. output_language-agnostic: it polishes whatever language the deliverable is written in, and a copywriter failure is swallowed as a WARN, never a halt.

Next, the promote-check rung — the assumption-registration tier that keeps load-bearing planning numbers in the engagement's single source of truth. Run it per $CLAUDE_PLUGIN_ROOT/references/orchestration/test-promote-check.md: scan the finished artifact for bare quantified planning literals (market sizes, rates, headcounts, price points) written directly rather than cited as a {{asm:<slug>}} placeholder backed by the engagement-root assumptions.json registry; propose registry entries (provenance_type + capped status), convert accepted occurrences to their placeholders, verify with a resolve-assumptions.py dry-run (fail-loud), and record an assumption-promotion decision-log entry keyed by (action_field, deliverable). This rung acts by default — it promotes rather than only warning — but stays advisory and never a hard fail (the consultant may opt out on the record) and is idempotent on a Step-7 resume. A promotable literal must never stall the flow.

Then: one Edit of field.json sets state → "complete" (keep dt_stage at "test") and the evidence_class, and one Edit of .metadata/execution-log.json appends the in-progress → complete transition to transitions[]. Then run the dependency cascade so every downstream deliverable that listed this one in its depends_on[] is flagged stale:

python3 $CLAUDE_PLUGIN_ROOT/scripts/deliverable-graph.py <engagement-dir> \
    cascade-stale <field-slug>/<deliverable-slug> --trigger deliverable_update

Surface data.newly_flagged in the session summary so the consultant knows which deliverables now need revisiting. A "success": false (node not found, bad dir) is non-blocking — the completion stands; surface the error and continue. If it does not survive, loop back — advance dt_stage to the stage the revision needs (often define or ideate) via the helper (a re-entry to an earlier stage is permitted) and continue; state stays in-progress.

8. Close the Session

Summarize: the deliverable's final state, the artifact path, key decisions logged, which personas challenged it, and the Test-stage promote-check outcome (assumptions promoted to the registry, or a recorded opt-out). Recommend the next step — the next unstarted deliverable in the WBS (via the WBS dashboard skill when present in the plugin, or by reading the field manifests directly).

Milestone dashboard refresh. When this session moved the deliverable's state to "complete" (or the delegated persona challenge closed its persona_review), the engagement's status changed — offer the consultant a fresh visual dashboard. If the engagement already has output/design-variables.json (a prior consult-dashboard run set up a theme), regenerate the HTML without prompting by delegating to the consult-dashboard-refresher agent with engagement_dir: <engagement-dir> and plugin_root: $CLAUDE_PLUGIN_ROOT; the agent runs the read-only generator and opens output/dashboard.html. If no theme is configured yet, point the consultant at /cogni-consult:consult-dashboard to set one up. This is a lightweight snapshot — the dashboard reflects the engagement state at this checkpoint, which is exactly what the consultant wants to see before picking the next deliverable.

Milestone README refresh. On the same trigger, also run python3 $CLAUDE_PLUGIN_ROOT/scripts/generate-engagement-readme.py "<engagement-dir>" — unconditional (unlike the theme-gated dashboard, no output/design-variables.json needed) and non-fatal: on failure, warn and continue.

Knowledge-base deposit is elected, not automatic. When this session moved the deliverable to complete, offer to deposit the completed artifact into the engagement's bound knowledge base (default-on: deposit without pausing in auto-walk mode, confirm first in interactive mode, never auto-fire without offering) so future gap-checks and research reuse its findings; the consultant may decline only on the record via a kb-deposit-waiver. The full deposit signature (reusing cogni-knowledge:knowledge-ingest-source verbatim), the provenance carried into the deposit context, and the waiver shape + idempotency are in $CLAUDE_PLUGIN_ROOT/references/orchestration/close-kb-deposit.md.

Publishing is elected, not automatic. When this session moved the deliverable to complete, the consultant may now turn it into a presentation-ready brief for Claude Design with /cogni-consult:consult-publish (slides, web-poster, report, or infographic). Mention it as an available next step only — never auto-fire it from this loop. Publishing is a deliberate consultant judgment call about which deliverables are presentation-worthy and which format fits; the design-thinking loop ends at complete.

Important Notes

• State ownership: deliverable state and dt_stage live only in the field's field.json; the artifact frontmatter never carries state. Field and engagement completion are derived at read time. See $CLAUDE_PLUGIN_ROOT/references/data-model.md.

• Edit, never rewrite: field.json, consult-project.json, personas/*.json, and the .metadata/ logs are all edited surgically; the root consult-project.json is never touched by deliverable work (its updated covers root-file changes only).

• Evidence discipline: research goes through the engagement's bound knowledge base per the Research Routing Rule ($CLAUDE_PLUGIN_ROOT/references/research-routing.md), never raw web search; every evidence-backed claim in the artifact carries the sources[] lineage triple so corrections can cascade.

• Claims-correction cascade: when the consultant surfaces a resolved cogni-claims correction, propagate the corrected ground — the script depends on what the correction touched:

  A deliverable's cited source (a claim in its sources[] lineage was deviated then resolved upstream) — flag that deliverable's downstream dependents stale, so its lineage_status.trigger reads claims_correction:

  python3 $CLAUDE_PLUGIN_ROOT/scripts/deliverable-graph.py <engagement-dir> \
      cascade-stale <field-slug>/<deliverable-slug> --trigger claims_correction
  

  A claim-type assumption's value (its ClaimRecord is resolved with resolution.action: corrected) — write the corrected value back, then fan staleness to every deliverable citing that assumption:

  python3 $CLAUDE_PLUGIN_ROOT/scripts/submit-assumption-claim.py <engagement-dir> \
      resolve-propagate <asm-id> --corrected-value <value> [--claim-id <id>]
  python3 $CLAUDE_PLUGIN_ROOT/scripts/deliverable-graph.py <engagement-dir> \
      cascade-stale --assumption <asm-id> --trigger assumption_update
  

  resolve-propagate mutates the assumption of record (demotes verified → reviewed) and is fail-loud ("success": false / exit 1) — surface data.failed_check and STOP, don't cascade a value that never changed. Every cascade-stale form is non-blocking (exit 0) — surface the error and continue. cogni-claims has no push callback into cogni-consult, so this stays a consultant-initiated step, run when a correction is noticed.

• Loop, not gate: stages may re-enter earlier stages; state stays in-progress until the test stage passes. State transitions go in the execution log; per-stage dt_stage moves are logged separately to .metadata/stage-log.json by the stage-advance helper (see Advancing the stage) — the two logs never mix.