Wiring The Citadel: Observatory Automation and Run-Scoped File Naming

The Problem: Silent Overwrite and a Passive Observatory

The Citadel launched with an Observatory directory containing five empty data folders — quality-reports/, cost/, latency/, evaluations/, metrics/ — and a README that said "keep this folder passive." The intent was right, but nothing actually filled it. No run wrote anything there. After two pipeline runs, neither had left a trace in the Observatory.

The same failure mode existed in content/. Every /publish-doc invocation wrote to the same three files: brief.md, article-draft.md, article.md. Run two overwrote run one completely. There was no history, no baseline, and no way to compare two runs side by side.

Both problems share one root cause: the pipeline had no concept of a run identity.

The Fix: Run-Scoped Naming Convention

The fix is a naming convention enforced at step 1 of every pipeline invocation.

At the start of each run, the pipeline derives a stable run ID from two components: a timestamp from date +%Y%m%d_%H%M%S and a slug generated from the brief's H1 title (lowercase, non-alphanumeric replaced with hyphens, consecutive hyphens collapsed):

TIMESTAMP=20260627_105134
SLUG=brief-wiring-the-citadel-observatory-and-run-naming-convention
RUN_ID=20260627_105134_brief-wiring-the-citadel-observatory-and-run-naming-convention

From that RUN_ID, three scoped paths are derived:

BRIEF = content/input/brf_<run-id>.md
WORK  = content/work/wrk_<run-id>/
PUB   = content/published/pub_<run-id>/

Every step in the pipeline reads and writes through $WORK and $PUB. Nothing writes to a flat filename anymore. The original brief is copied to its stamped path but never deleted.

The result is visible in the repository right now. content/published/ holds two intact directories:

pub_20260627_092341_lab-note-building-the-citadel/
pub_20260627_105134_brief-wiring-the-citadel-observatory-and-run-naming-convention/

The same pattern was applied to the House of Personal Writing. create-personal-piece.md runs the same step 1. The standalone review-personal-piece.md and publish-personal-piece.md commands accept the wrk_* path as $ARGUMENTS, falling back to the most recently modified directory if none is provided.

Migrating the existing flat files reconstructed timestamps from their last-modified times. No content was lost.

Observatory Automation: What the Run Record Captures

Step 10 of publish-doc.md fires unconditionally — whether the pipeline succeeded or failed at step 7. It writes a JSON record to:

observatory/quality-reports/<run-id>.json

The record is defined by observatory/schemas/run-record.schema.json. Fields written per run:

• house — which pipeline produced the record
• date, run_id, brief_slug, brief_path — run identity
• pipeline_status — "published" or "failed"
• output — paths to article.md, publish-manifest.json, release-notes.md; all null if the pipeline failed before step 9
• models — model ID per role (writer, illustrator, reviewer, editor, publisher)
• review.status, review.confidence, review.cycles, review.summary, review.required_changes_count — pulled from $WORK/review-result.json
• cost_usd, latency_seconds — both null; those fields require API instrumentation that does not exist yet

The models and output objects use additionalProperties in the schema, so both houses write to the same schema without house-specific variants. A failed run is still recorded. An incomplete record with pipeline_status: "failed" and null outputs is valid data — it tells you the run happened, which review cycle it reached, and how confident the reviewer was before stopping.

The schema_version field is present on every record so future dashboard tooling can handle schema evolution without guessing.

What Was Created and What Was Modified

Files created:

• observatory/schemas/run-record.schema.json — JSON Schema v2020-12 for observatory records
• ~/.claude/commands/brief.md — global /brief skill; saves a Why/How/What session summary to content/input/brf_<timestamp>_<slug>.md

Files modified:

• observatory/README.md — now documents automated population, the field list, and dashboard readiness
• houses/house-of-technical-publication/CLAUDE.md — "Standard File Paths" replaced with "File Naming Convention" documenting brf_/wrk_/pub_ prefixes
• houses/house-of-technical-publication/.claude/commands/publish-doc.md — step 1 (run ID) and step 10 (observatory write) added; all internal path references now use $WORK and $PUB
• houses/house-of-personal-writing/CLAUDE.md — same naming convention section added
• houses/house-of-personal-writing/.claude/commands/create-personal-piece.md — step 1 + step 11 observatory pattern
• houses/house-of-personal-writing/.claude/commands/publish-personal-piece.md — derives PUB from $ARGUMENTS or most recent wrk_*
• houses/house-of-personal-writing/.claude/commands/review-personal-piece.md — same fallback

Files renamed (migration, no data lost):

• content/input/brief.md → content/input/brf_20260627_092341_lab-note-building-the-citadel.md
• content/work/*.md, review-result.json → content/work/wrk_20260627_092341_lab-note-building-the-citadel/
• content/published/*.md, publish-manifest.json → content/published/pub_20260627_092341_lab-note-building-the-citadel/
• Personal writing: content/input/brief.md → content/input/brf_20260627_075710_brief.md

Takeaway

The Observatory was passive because the pipeline had no run identity. Once you add a stable RUN_ID at step 1, everything downstream — file isolation, artifact paths, observatory records — becomes a naming problem, not an architecture problem. The /brief skill closes the remaining gap: it captures the session before it disappears into chat history and drops a stamped brief file directly onto the /publish-doc input path.