Spec Pov

Phase 1: Ground with authorized scouts or bounded inline fallback

Grounding searches code, git, the issue tracker, PRs, and docs. When the package-local boundary permits dispatch, use scout sub-agents that return only a dossier path plus a short gist. Otherwise apply the same persona budgets serially in the orchestrator, keep raw search notes in the scratch directory, and carry only compact evidence into verdict reasoning.

Resolve current project orientation first. Derive stack, dependency/license surface, conventions, and structure from the current target repo/worktree for this run. Record current git identity and dirty state when available, carry direct source refs, and never persist or reuse the orientation across runs, branches, or worktrees. If git or a required source cannot be read, record the concrete degraded fact and narrow the project-floor claim; do not substitute conversation claims or stale orientation.

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

umask 077
SCRATCH_DIR="$(mktemp -d "${TMPDIR:-/tmp}/spec-first-pov.XXXXXX")"
[ -d "$SCRATCH_DIR" ] && [ ! -L "$SCRATCH_DIR" ] || { echo 'private scratch creation failed' >&2; exit 1; }
chmod 700 "$SCRATCH_DIR"
echo "$SCRATCH_DIR"

This directory is owner-only, ephemeral scratch. Recheck that it remains a non-symlink directory before atomic publication; durable POV evidence must use its canonical artifact owner rather than this path.

Every scout payload 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.

Tier-sensitive execution. For Tier 1 (reversible), run a combined project-grounding and external-evidence pass at tight budgets; use subagents only when authorized, otherwise run the two lenses serially inline. Skip the standalone precedent lens because the project-grounding pass includes the prior-decision scan. For Tier 2/3, use the full fleet when authorized or the same three lenses serially inline:

• project-grounding scout (extraction tier) — read references/agents/project-grounding-scout.md and seed a generic subagent with it. With the agnostic profile already loaded from the cache, this scout runs only the candidate-specific slice: the named incumbent for this candidate, its call-sites/footprint, incumbent-pain, and the license/compat check against the profile's dependency-license set. Do not re-derive stack, conventions, or structure — those are in the profile. But note the profile may name an incumbent dependency, and a named dep is only a lead — it does not satisfy the project floor (see references/method.md), which still requires a freshly verified call-site the cache never holds. Do not let a cache-named incumbent short-circuit the fresh touchpoint check.
• 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 (docs/solutions/, ADRs, design docs — file reads, no tools needed); only its tracker/PR portion is capability-gated and degrades gracefully when those interfaces aren't reachable. Do not skip the whole scout for missing tracker access — that would drop the only path that surfaces 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; capability-gated on web tools. Scale the remit to the tier so Tier 3's deeper-workup promise 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 load-bearing claim (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.

Capability gating is two-level: skip only a scout (or scout-portion) with no reachable surface 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 are tool-gated and degrade. Let a scout that loses a tool mid-run self-report "unavailable." Never block on a missing surface — record it and let it lower the verdict's stated confidence, or trip the external floor (Phase 2) when the external leg is entirely absent.

Populate the provenance buckets from the returned dossiers, keeping them separate for Phase 2: observed-project-facts and verified-external-facts (these count as grounding) vs. conversation-claims and unconfirmed-assumptions from a warm invocation (these do not count until a scout corroborates them). Read dossiers from their paths on demand; do not pull their bulk into this context.

Phase 2: Verify against the two floors

Read references/method.md now, before reasoning about the verdict — it defines the Verify and Verdict steps, the skeptic stance and reversibility tiering as cross-cutting properties, and the two-floor Invalid-Verdict gate. Apply that gate as a pass/fail checklist over the dossiers: a failed floor forbids Adopt/Reject and returns the matching Hold subtype. Do this reasoning on the clean context — read a dossier on demand, never pull its bulk in.

Phase 3: Verdict

When the user or visible upstream handoff explicitly authorized cross-model/delegated work, read references/cross-model-panel.md before final synthesis. Apply its canonical authorization, external-data, allowlisted-input, redaction, source-identity, provider-independence, bounded lifecycle, and reap gates. Use references/agents/pov-peer.md, references/pov-schema.json, and the Skill-local adapter/runner only after every gate passes. Missing authorization or any safety fact means zero peer processes and no independent coverage claim. Reconcile valid peer disagreement against the two floors and the frozen approach set; never decide by vote.

Emit the verdict contract defined in references/method.md — grade vocabulary, schema fields, tier sizing, and output economy are all specified there. The verdict is a compact chat block, not a research report: lead with the grade, keep each schema field terse, and never reprint scout dossiers or raw search output. Size it to the tier — a Tier 1 verdict fits one screen; Tier 2/3 carries the full workup but still leads with the verdict and cites evidence rather than pasting it.

Phase 4: Follow-up

The chat verdict (the TL;DR) is the deliverable. What you offer next is reasoned from the verdict and sized to the tier — never a fixed menu, and never an assumption that everything routes to a plan.

Compute the next step. From the grade and the verdict's Handoff field, reason about the single best next move and a one-clause why — it is not always obvious between plan and brainstorm, so decide in context:

• Adopt, scope clear → take it into spec-plan.
• Adopt, scope still fuzzy → spec-brainstorm to pin down what "adopt" means before planning.
• Trial → scope a timeboxed spike (spec-work).
• Hold / Reject / Not-our-problem → no handoff; there is nothing to take forward.

Tier-gate the offer (anti-ritual):

• Tier 1, or a Reject / Not-our-problem grade → end with a single prose line — e.g. "Want the full write-up, or <computed next step>? Otherwise we're done." No blocking menu; silence means done.
• Tier 2/3 with an actionable grade → ask via the platform's blocking question tool, with the computed next step as the first, dynamically-labeled option:
  1. <computed next step> (e.g. "Plan the adoption with spec-plan") — seeded with the verdict substance, not a file pointer.
  2. Full write-up — the expanded, shareable artifact.
  3. Done. Add spec-compound as a one-line prose nudge under the menu, not a slot: "Want it in our decision history? say 'compound it.'" It is the least-frequent path and is never the first thing offered.

On each selection:

• Computed next step → invoke that skill via the platform's skill-invocation primitive, seeding it with the verdict substance (the decision, conditions, and verified facts).
• Full write-up → read references/report.md and follow it (HTML by default; opened locally via an available HTML tool). Opt-in; the default stays chat-only.
• "compound it" → invoke spec-compound with mode:headless, seeding it with the structured verdict for tooling_decision / architecture_pattern storage (no schema change; headless avoids its interactive prompts). Never mandatory.

Warm invocations stay a guest: output the verdict block, hand control back, and offer none of the above unless the user asks — a mid-session interjection does not push a next-step or capture decision.