Grounding the POV (Phase 1 machinery)

Read this before dispatching scouts. It defines how the Ground step (SKILL.md Phase 1) runs: which model each scout uses, where scratch files go, what each scout receives, which scouts run at each tier, and how grounded facts are kept apart from unconfirmed ones.

Model Tiers

Dispatch is tiered by task shape, never hardcoded to a model name:

• Extraction tier — the project-grounding scout and the precedent-&-activity scout: search-and-quote work. Use the platform's cheapest capable model when the harness exposes a known override; otherwise inherit.
• Generation tier — the external-evidence researcher: web/docs retrieval and entailment checking. Use the platform's mid-tier model when a known override exists; otherwise inherit.
• Ceiling tier — the final assessment stays with the agent running ce-pov, even when another agent delegated the task to it. That agent checks whether the evidence is sufficient, weighs skeptical findings, and produces the required result. Research subagents gather evidence; they do not make this final assessment.

If the user's or project's active instructions or the live prompt name a model for this subagent, use that model instead of the default tier above. They can name it by the subagent's tier, by its prompt-asset name, or for the subagents being dispatched. Without such an instruction, a model configured for this subagent's role takes precedence over the default tier.

When per-agent models cannot be chosen. If the platform's subagent mechanism cannot select a model per agent, dispatch every scout on the inherited model and keep their read budgets. Cost control then comes from the read budgets and the tier-sensitive scout count, not from tiering.

When a scout dispatch is rejected, first check whether an agent launched. If the rejection was a bad argument before launch, correct it once. If the platform is out of capacity, leave the work queued. If a launch still fails after the correction, gather that scout's bounded evidence inline and lower the verdict's stated confidence.

Create the scratch dir once, and reuse the echoed path for every scout this run:

SCRATCH_ROOT="/tmp/compound-engineering-$(id -u)";
[ ! -L "$SCRATCH_ROOT" ] && (umask 077; mkdir -p "$SCRATCH_ROOT") 2>/dev/null && [ ! -L "$SCRATCH_ROOT" ] && [ -O "$SCRATCH_ROOT" ] && [ -w "$SCRATCH_ROOT" ] || SCRATCH_ROOT="${TMPDIR:-/tmp}/compound-engineering-$(id -u)";
if [ -L "$SCRATCH_ROOT" ]; then echo "unsafe scratch root symlink: $SCRATCH_ROOT" >&2; exit 1; fi;
(umask 077; mkdir -p "$SCRATCH_ROOT") || exit 1;
if [ -L "$SCRATCH_ROOT" ] || [ ! -O "$SCRATCH_ROOT" ]; then echo "scratch root is not owned by the current user: $SCRATCH_ROOT" >&2; exit 1; fi;
chmod 700 "$SCRATCH_ROOT" || exit 1;
SCRATCH_DIR="$SCRATCH_ROOT/ce-pov/$(openssl rand -hex 4)";
(umask 077; mkdir -p "$SCRATCH_DIR") || exit 1; chmod 700 "$SCRATCH_DIR" || exit 1;
echo "$SCRATCH_DIR";

Scoping applies on both paths. Use the project's active instructions already in context. If the candidate cannot be scoped from the frame and existing context, allow one targeted root or workspace probe. This holds whether this phase dispatches scouts or resolves the facts with bounded inline reads.

Every scout prompt carries the same context. A fresh subagent does not inherit this conversation, so fill the persona files' {subject} / {scratch-dir} placeholders at dispatch. Pass each scout the framed question (subject + intent), the named incumbent and the reversibility tier, and the resolved <scratch-dir> path, plus any user-supplied links for the external researcher. A scout seeded with only its generic persona grounds "some external thing" and can produce an empty or unfocused dossier.

A generic subagent is one whose whole brief is the prompt you send it. When the host's subagent tool makes you name an agent, use only a name you have seen the host list, and list its agents before the first dispatch if you have not. Pick the plain one that adds no role of its own, which still counts as a generic subagent. Never pick a specialist, such as a reviewer or worker agent, whose own role prompt, output format, or tool limits would replace or narrow your prompt.

Which scouts run depends on the tier. For Tier 1 (reversible), run a single combined grounding pass: seed one subagent with references/agents/project-grounding-scout.md covering the candidate-specific project facts (incumbent, call-sites) at a tight read budget, and one with references/agents/external-evidence-researcher.md. Skip the standalone precedent scout; on this tier the project-grounding scout's prior-decision scan (<root>/solutions/, ADRs, design docs) is the precedent check, so it must run. For Tier 2/3, dispatch all three scouts in parallel:

• project-grounding scout (extraction tier) — read references/agents/project-grounding-scout.md and seed a generic subagent with it. Run the candidate-specific slice fresh: the named incumbent for this candidate, its call-sites/footprint, incumbent-pain, exact runtime or framework constraints that materially affect compatibility, and the project/candidate/dependency license check. Do not start with generic shape discovery; the project floor (the minimum verified project evidence a verdict needs, defined in references/method.md) still requires a freshly verified call-site and current compatibility evidence.
• precedent-&-activity scout (extraction tier) — read references/agents/precedent-activity-scout.md and seed a generic subagent with it. Always run its local-doc precedent pass (<root>/solutions/, ADRs, design docs — file reads, no tools needed); only its tracker/PR portion depends on a reachable tracker and is skipped, with a note, when those interfaces aren't reachable. Do not skip the whole scout for missing tracker access; that would drop the only path that finds a prior local adopt/reject decision.
• external-evidence researcher (generation tier) — read references/agents/external-evidence-researcher.md and seed a generic subagent with it; it runs only when web tools are reachable. Scale its brief to the tier so Tier 3's deeper workup is real, not nominal. At Tier 3, seed it with a deeper brief: a wider source net, a larger read budget, and mandatory two-source corroboration on every claim the verdict depends on (at Tier 3 a single-source claim cannot anchor the verdict). Tier 2 uses the persona's standard budget and its prefer-two-sources default.

Skip work only when nothing can reach it. Skip a scout, or a portion of one, only when it has no reachable interface at all. The project-grounding scout and the precedent scout's local-doc pass are file reads and always run; the tracker/PR reads and the external researcher need tools and are skipped when those tools are absent. Let a scout that loses a tool mid-run report "unavailable" itself. Never block on a missing interface. Record it and let it lower the verdict's stated confidence, or fail the external floor in Phase 2 (Verify Grounding) when the external leg is entirely absent.

Sort every fact by where it came from using the returned dossiers and your own bounded inline-read observations, and keep the groups separate for Phase 2 (Verify Grounding): observed-project-facts and verified-external-facts count as grounding; conversation-claims and unconfirmed-assumptions from a warm invocation do not count until a scout or a bounded inline read of the authoritative source corroborates them. Read dossiers from their paths on demand; do not pull their bulk into this context.