session-summary-consistency

name	session-summary-consistency
description	Quote-level consistency check on scene_extractions_new/ or scene_extractions/ for VTT transcription errors, garbled phrases, proper-noun misspellings, and clarity issues in verbatim quote blocks. Use when the user asks for /session-summary-consistency [session-dir].
metadata	{"short-description":"Check scene quote transcription consistency"}

Session Summary Quote Consistency Check

Check the verbatim quote blocks in scene_extractions_new/ or scene_extractions/ for VTT transcription errors, garbled phrases, proper-noun misspellings, and pronoun or clarity issues. Propose fixes before touching anything; apply only fixes the user approves.

This skill is quote-level only. It does not check scene summaries for canon facts, mechanical accuracy, or prose quality, except for player-name scrubs described below. It restores what was said, meaning it fixes transcription errors; it does not smooth grammar, style, or voice. For higher-level campaign consistency, use consistency-check. For VTT cleanup before extraction, use vtt-spell-pass if available. Voice-aware sentence smoothing belongs to the narration layer, not here.

This is the Codex port of the Claude skill. Do not edit ~/src/mytools/dotfiles/claude/skills/session-summary-consistency/ when changing this skill.

Codex Compatibility

• Ask user questions in chat. Do not refer to Claude AskUserQuestion.
• Use update_plan for multi-step progress when useful.
• Use apply_patch for manual edits to tracked documents.
• Codex does not have the Claude Artifact review callback flow. If the user asks for batch review, write a normal Markdown or JSON review queue file in the session directory or present the queue in chat, then ask the user to provide decisions in chat.
• Do not apply edits merely because the workflow found likely fixes. User approval is the gate.

Inputs

• session-dir: the summaries/YYYYMMDD/ directory to check. Default to CWD only when it is already a session directory. If CWD is a campaign root, ask which session to use or choose the most recent summaries/*/ that has a scene extraction directory.
• Scene files are in <session-dir>/scene_extractions_new/ or <session-dir>/scene_extractions/. Detect whichever exists and call it <scene-dir>.
• Corrections glossary is at <campaign-root>/notes/vtt_transcription_corrections.md.

Workflow

1. Locate Files

Find the scene extraction directory:

ls <session-dir>/scene_extractions_new/ 2>/dev/null || ls <session-dir>/scene_extractions/

If neither exists, tell the user and stop. Scene files are expected to be named NN_<slug>.md and contain ## Verbatim moments sections.

Walk up from session-dir to find the campaign root, usually the directory that contains notes/vtt_transcription_corrections.md. Read that glossary; it is the primary reference for known proper nouns and mishearing patterns.

Also check for docs/entity_registry.yaml or docs/entity_inventory.md at the campaign root. The glossary documents mishearing patterns; the registry is the identity authority for canonical spelling, aliases, and known entities. Use both when a name-like garble's referent is uncertain.

2. Read Scene Files

Read every *.md in <scene-dir>. For each file:

• Extract every quote block under ## Verbatim moments, usually lines inside Markdown blockquotes.
• Note the scene filename, line number, and speaker label above each quote.
• Sweep speaker labels and ## Scene summary prose for player-name leaks.

Do not quote-correct or reword ## Scene summary. The only exception is the player-name scrub rule below. If a summary-level garble cannot be safely scrubbed as a name replacement, flag it as an upstream scene-extract issue rather than rewriting the summary.

3. Classify Quote Problems

For each quote problem, classify it:

Category	Meaning	Action
VTT mishear	Single word or name is phonetically garbled and the correction is clear	Replace with corrected word
Proper noun error	Name is misspelled and canonical form is known	Replace with canonical form
Garbled phrase	Multi-word phrase is scrambled but meaning is reconstructable	Replace with reconstruction and bracket uncertain words
Unrecoverable	No reconstruction is possible	Replace with [inaudible] or [inaudible -- probable "X"] if a guess exists
Grammar/pronoun error	Wrong pronoun, article, or verb changes meaning	Replace with correct form
Duplicate across scenes	Same garbled line appears in multiple scene files	Flag each instance for the same fix

Cross-reference rules:

• Check every proper noun in a quote against the glossary. If a name appears in a wrong-form column, consider the paired canonical correction.
• Context overrides the glossary for identity. The same garble can map to different names in different scenes, so confirm against the scene summary, surrounding dialogue, and entity registry before applying.
• Watch for glossary-substitution artifacts where an earlier replacement left doubled reads or false choices, such as the same referent appearing as both sides of an "X or Y" phrase.
• Do not assume a suspicious name variant is an authentic nickname. Check the registry, glossary nickname entries, and DO-NOT-CORRECT list first. If neither corroborates it, surface it for user ruling.
• Once a garble is confirmed, search the whole scene set for phonetically similar variants before calling the fix complete.
• Use the scene summary as ground-truth paraphrase for meaning, not exact wording.
• Use surrounding dialogue to resolve garbled phrases.

Speaker label and player-name rule:

• Speaker labels must use the character name or GM, never the player's real name.
• Use the campaign's own glossary ## Player names -> characters section for the mapping. Do not hardcode player names; they vary by campaign.
• Label format varies across scenes, including [GM], bold names, or other Markdown forms. Sweep every scene's labels before producing the report.
• Apply the same player-name scrub to ## Scene summary prose. This is a name replacement, not a summary rewrite.
• Player names inside verbatim quote content may be left as-is when they are actual out-of-character table speech, but flag them as observations for the user.

Do not change:

• Genuine speech disfluencies, repetitions, false starts, um, or uh.
• Grammar, run-ons, or phrasing that reflect how someone actually spoke.
• Profanity or table vocabulary.
• Clearly marked out-of-character crosstalk.
• Numbers, dice results, or mechanics unless the transcription is plainly wrong and evidence settles it.
• Speaker attribution without user confirmation. Flag suspected attribution errors as notes.

4. Produce a Proposal Report

Group findings by scene file. For each issue, show:

**Scene N - `filename.md`**

Line <N> - [Category] ([Speaker])
> Original: "quoted text with the problem"
> Proposed: "corrected text"
Reason: one sentence explaining the correction.

Include a summary table:

Category	Count
VTT mishear	N
Proper noun error	N
Garbled phrase	N
Unrecoverable -> [inaudible]	N
Grammar/pronoun	N
Duplicate across scenes	N

Present the report in chat unless it is too large. For a large report, write it to <session-dir>/quote_consistency_review.md, summarize counts in chat, and ask the user whether to approve all, approve selected item ids, reject all, or discuss specific items.

5. Wait for User Approval

After presenting the report, ask:

Want me to apply all of these, or go through them selectively?

Supported responses:

• apply all: apply every proposed fix.
• selective: walk through the report item by item and apply only confirmed fixes.
• Item ids or ranges: apply only those accepted ids, and leave the rest.
• none or just the report: stop without editing.

6. Apply Fixes

For each approved fix, use apply_patch with enough surrounding context to make the edit unique. If the exact string cannot be found because the file changed, search for the current line and retry with fresh context.

After edits, verify that original target strings no longer remain:

grep -rn "<original garbled string>" <scene-dir>/

Any remaining hits mean a fix was missed or a duplicate variant exists. Inspect and either apply the approved fix or report why it remains.

7. Feed Confirmed Garbles Back to the Glossary

After approved fixes are applied, identify new VTT patterns not already in notes/vtt_transcription_corrections.md and ask:

Found N new VTT patterns not yet in the glossary. Want me to add the safe ones?

Only add safe, non-word garbles. Real English words are context-dependent and must not become global replacements because the applier is case-insensitive across the whole transcript.

Safe examples:

• glabbagel -> Glabbagool
• Ragum -> Grygum
• Hulkrist -> Alkrist
• Demonor('s) -> Deneir('s)
• Fembrance -> Fembris
• graffled -> grappled
• stinge -> singe

Unsafe examples:

• teeth -> tea
• high -> ki
• fake -> point
• snake -> sneak
• allowed -> Avowed

Append only user-approved safe pairs to the appropriate glossary section. If the user confirms a suspicious variant is authentic table speech or a nickname, add it to the glossary's DO-NOT-CORRECT area instead of the garble table.

Batch Review Queue

Use this when the report is too large for comfortable chat adjudication or the user asks for a batch review artifact.

Create a review queue file such as <session-dir>/quote_consistency_review.json or <session-dir>/quote_consistency_review.md. Keep the shell/chat summary visible with per-scene counts and category totals.

Use stable ids like s03-04: scene number, then the finding number within that scene. Include:

{
  "id": "s03-04",
  "file": "scene_extractions_new/03_example.md",
  "line": 120,
  "category": "VTT mishear",
  "speaker": "Brewbarry",
  "original": "Brube's world",
  "proposed": "Brewbarry's world",
  "reason": "The alternate transcript and surrounding dialogue identify Brewbarry.",
  "evidence": "Summarize supporting evidence without long transcript quotes."
}

Then stop and ask the user to provide decisions in chat. Map decisions this way:

Decision	Action
approve	Apply with apply_patch, then run verification grep
reject	Leave the quote as transcribed and log the ruling
discuss + note	Follow the note; confirmed coinages go to DO-NOT-CORRECT
discuss without note	Bring the item back to chat for ruling
unmarked	Leave unchanged and mention it in the summary

Auto-apply only after explicit user approval. Even in a batch queue, do not silently edit files based on confidence.

Conventions

• Bracket notation for preserved artifacts: [Corrected form; VTT: "original garbled text"].
• Inaudible tag: [inaudible] for total loss; [inaudible -- probable "X"] when context suggests a likely word.
• Unclear tag: [unclear -- possibly "X" or "Y"] when two interpretations are equally plausible and the distinction matters.
• A quote can be garbled while the summary's paraphrase is accurate. Fix the quote to match the intended speech, not to copy the summary wording.
• This skill does not update session-summary or gm-assist docs. Flag any upstream errors as out-of-scope observations.