odoo-deep-survey

Odoo Deep Survey - broad→narrow→deep, then synthesize

Consumer: the next execute AI agent (architect, coder, reviewer) - not a human. Every finding must be actionable without re-derivation: a file:line pointer or OSM citation, marked RESOLVED or UNRESOLVED, written so a downstream agent reads few tokens and understands at once.

Role

Read-only reconnaissance orchestrator. Loaded by the Skill tool from the main context, it fans out anonymous worker agents at three escalating tiers to map whether the codebase can meet the intent and where the work lands. Writes analysis artifacts under <SHARE_DIR>/survey/ (resolve <SHARE_DIR>/<ISOLATE_DIR> once per ${CLAUDE_PLUGIN_ROOT}/snippets/state-root-resolution.md; substitute the captured absolute path - never write the placeholder or a bare .odoo-ai/ into a Read/Write/Edit); never writes Odoo source.

When this runs (opt-in only)

odoo-intake asks its approach question from a light Phase R recon and offers a deep survey. The user replies refine: deep-survey - the PLAN reply set (${CLAUDE_PLUGIN_ROOT}/snippets/planning-gate-contract.md); deep-survey is the feedback, not a reply word of its own. ONLY THEN does intake invoke this skill via the Skill tool. This skill produces no routed deliverable, only analysis that sharpens the approach. Never run it automatically.

Inputs (passed by intake in the invocation)

• Intent / purpose / expected outcomes - the closed Phase 0 gate (what / why / done-looks-like).
• Odoo series + profile - from the intake version gate, else resolved per ${CLAUDE_PLUGIN_ROOT}/snippets/project-facts-resolution.md. Pin this concrete series on every OSM call (see § OSM grounding).
• The recommended approach - so synthesis can express its findings as a delta against it.
• Slug - reuse the feature slug intake already uses for brainstorm artifacts; the survey dir is <SHARE_DIR>/survey/<slug>-<date>/.

Hard rules

1. Main-context spawner. This skill launches subagents; it MUST be invoked via the Skill tool from the main context. A subagent MUST NOT invoke this skill - doing so creates uncontrolled nesting.
2. Read-only. Workers read Odoo source and call read-only OSM tools. The ONLY writes are this skill's own analysis artifacts under <SHARE_DIR>/survey/ and worklog entries under <ISOLATE_DIR>/worklog/. No edits to Odoo source, no routed deliverable.
3. Every claim is grounded. A finding without a file:line or an OSM citation is a guess - drop it or mark it UNRESOLVED. The worklog EVIDENCE field is mandatory (${CLAUDE_PLUGIN_ROOT}/snippets/worklog-contract.md).
4. Escalate by evidence, not by vibe. Phase 3 (opus) runs only when a measurable trigger fires (see § Phase 3). Default is to stop after Phase 2 and synthesize.
5. Zero-trust: descriptions are CLAIMS, source is TRUTH. Confirm every description (docstring, comment, manifest prose, prior report, OSM descriptive field) against the resolved source before marking a finding RESOLVED; when they disagree, source WINS and the description is flagged stale. OSM STRUCTURE stays the trusted primary. Full rule (inlined into every worker brief): ${CLAUDE_PLUGIN_ROOT}/snippets/zero-trust-code-survey.md.