/workflow-inspect

Fill a /workflow-retro record's pending ## Cost section with numbers computed from the raw session transcripts: tokens and wall-clock per sub-agent, sub-agents' share of the total, and the handoff tax (files the planner read that the main agent re-read). Run after /workflow-retro, within the transcript retention window.

Input

• .workflow-log/ records whose ## Cost is pending or pending re-inspection
• python3; inspect.py in this skill's directory (read-only: parses transcripts, prints markdown to stdout)

Ground rules

• Numbers come from the script, verbatim. Never estimate, adjust, or add a figure the script didn't print. Script fails or its output carries warnings → surface them to the user as-is (AGENTS.md Rule 2 applies)
• The only mutation is the log file: replacing the ## Cost section of the chosen record(s), after the user confirms. The one other edit allowed is adding session IDs to the record's Sessions: line (step 4), also after the user confirms. Never commit anything; .workflow-log/ stays local

Steps

Step 1 — Pick the record(s)

• Log directory, as in /workflow-retro: .workflow-log/ under the main worktree (parent of git rev-parse --path-format=absolute --git-common-dir)
• List records whose ## Cost is still pending or is marked pending re-inspection (sessions were added after the last inspect; Step 5 replaces the whole section, marker included)
  • One → proceed with it
  • Several → ask which (offer "all"; each is one script run)
  • None → report that every record is already inspected, and stop

Step 2 — Gather the session IDs

• Read the record's Sessions: line
• unknown, or Step 3 finds no transcript for any listed ID → cost is unrecoverable once transcripts are pruned. With the user's confirmation, close the section with unavailable — transcripts pruned or session IDs unknown instead of leaving pending forever

Step 3 — Run the inspector

One command, from the repository root:

python3 ${CLAUDE_SKILL_DIR}/inspect.py <session-id> [<session-id> …]

The script:

• Locates each session's transcript by global search of ~/.claude/projects/ (transcripts are filed under the session's launch directory, not necessarily the worktree; never derive the directory from a path)
• Resolves every sub-agent transcript via its spawn toolUseId
• Deduplicates records shared by resumed sessions
• Prints a complete ## Cost section
• Fails soft: what it can't parse or find becomes a warning line in the output, and the numbers are then lower bounds

Step 4 — Reconcile

Before writing, act on the output:

• Warning that sub-agent transcripts belong to sessions not listed in the record → the feature spanned more sessions than the retro knew. Offer to add those IDs to the record's Sessions: line and re-run the script once with the full list
• Cost data contradicts the record's outcome half (e.g. more sub-agent runs than the Steps table's review rounds) → point out the discrepancy to the user. The outcome sections are theirs to amend; never edit them yourself

Step 5 — Confirm and write

• Show the user the script's output
• Confirmed → replace the record's entire ## Cost section (heading included) with it, leave every other section untouched, report the file path
• Declined → leave the record as it was

Output

• The record's ## Cost section replaced (or closed as unavailable)