@windyroad/itil 0.13.0-preview.152 → 0.14.0-preview.154

package/.claude-plugin/plugin.json

@@ -1,5 +1,5 @@
 {
   "name": "wr-itil",
-  "version": "0.13.0",
+  "version": "0.14.0",
   "description": "ITIL-aligned IT service management for Claude Code"
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@windyroad/itil",
-  "version": "0.13.0-preview.152",
+  "version": "0.14.0-preview.154",
   "description": "ITIL-aligned IT service management for Claude Code (problem, and future incident/change skills)",
   "bin": {
     "windyroad-itil": "./bin/install.mjs"

package/skills/work-problems/SKILL.md

@@ -171,6 +171,35 @@ notes: <one-line>
 
 Architect review (R2) requires the commit state fields (`committed` / `commit_sha` / `reason`) so **Step 6.75's Dirty-for-known-reason branch stays evaluable** from the summary alone. JTBD review requires `ticket_id` / `action` / `skip_reason_category` / `outstanding_questions` so Step 2.5 and the Output Format's Completed / Skipped / Outstanding Design Questions tables can be populated deterministically without the orchestrator having to re-parse ticket files.
 
+**Per-iteration cost metadata.** Alongside `.result`, the `claude -p --output-format json` response carries cost + usage fields in the same JSON blob. The orchestrator MUST extract these **named fields only** into per-iteration totals and session aggregates — nothing else from the JSON should be surfaced to the user or logged (PII guard: the response also carries `session_id`, `model`, `stop_reason`, and other envelope fields; the extraction is **scoped to the named fields** below so future contributors do not unconsciously broaden it).
+
+Extracted fields (explicit field list):
+
+- `.total_cost_usd` — dollar cost for the iteration.
+- `.duration_ms` — wall-clock duration of the iteration subprocess.
+- `.usage.input_tokens` — prompt tokens.
+- `.usage.output_tokens` — generated tokens.
+- `.usage.cache_creation_input_tokens` — tokens written to the prompt cache on this invocation.
+- `.usage.cache_read_input_tokens` — tokens read from the prompt cache on this invocation (cache-read is the signal for warm-cache reuse across subsequent subprocess invocations in the same Bash session; high values here indicate the iteration benefited from prior-invocation caching).
+
+Use `jq` (or an equivalent JSON parser) to extract them:
+
+```bash
+# $SUBPROCESS_OUTPUT holds the full JSON response body from claude -p.
+read -r ITER_COST ITER_DURATION_MS ITER_INPUT ITER_OUTPUT ITER_CACHE_WRITE ITER_CACHE_READ < <(
+  jq -r '[.total_cost_usd, .duration_ms, .usage.input_tokens, .usage.output_tokens, .usage.cache_creation_input_tokens, .usage.cache_read_input_tokens] | @tsv' <<<"$SUBPROCESS_OUTPUT"
+)
+# Accumulate into session totals for the ALL_DONE Session Cost section.
+SESSION_COST=$(awk "BEGIN { printf \"%.4f\", ${SESSION_COST:-0} + $ITER_COST }")
+SESSION_DURATION_MS=$(( ${SESSION_DURATION_MS:-0} + ITER_DURATION_MS ))
+SESSION_INPUT_TOKENS=$(( ${SESSION_INPUT_TOKENS:-0} + ITER_INPUT ))
+SESSION_OUTPUT_TOKENS=$(( ${SESSION_OUTPUT_TOKENS:-0} + ITER_OUTPUT ))
+SESSION_CACHE_WRITE_TOKENS=$(( ${SESSION_CACHE_WRITE_TOKENS:-0} + ITER_CACHE_WRITE ))
+SESSION_CACHE_READ_TOKENS=$(( ${SESSION_CACHE_READ_TOKENS:-0} + ITER_CACHE_READ ))
+```
+
+Do NOT extract `session_id`, `model`, `stop_reason`, `permission_denials`, `uuid`, or any other field from the JSON response. Those are subprocess-envelope fields that serve no user-visible purpose and risk leaking subprocess-internal identifiers into orchestrator output.
+
 **Exit-code semantics.** `claude -p` exits non-zero when the subprocess fails hard — subprocess crash, auth failure, unresolvable permission denial, API/quota exhaustion. The orchestrator reads the exit code BEFORE parsing `.result`:
 
 - Exit 0 → parse `ITERATION_SUMMARY` from `.result` field; proceed to Step 6.
@@ -197,14 +226,15 @@ After each iteration, report:
 - What was done (investigated, transitioned to known-error, fix implemented, etc.)
 - The outcome (success, partially progressed, skipped, scope expanded)
 - How many problems remain in the backlog
+- The iteration's cost metadata — format: `($<cost>, <duration_s>s, <total_tokens_K>K tokens)`. Cost comes from the `.total_cost_usd` field extracted in Step 5; duration from `.duration_ms`; total tokens is the sum of `.usage.input_tokens + .usage.output_tokens + .usage.cache_creation_input_tokens + .usage.cache_read_input_tokens`.
 
 Format as a brief status line, not a wall of text. The user will read these when they return.
 
 **Example:**
 ```
-[Iteration 1] Worked P029 (Edit gate overhead for governance docs) — implemented fix, closed. 8 problems remain.
-[Iteration 2] Worked P021 (Governance skill structured prompts) — investigated root cause, transitioned to known-error. 7 problems remain.
-[Iteration 3] Skipped P016 (Multi-concern ticket splitting) — fix released, awaiting user verification. Worked P024 (Risk scorer WIP flag) — implemented fix, closed. 6 problems remain.
+[Iteration 1] Worked P029 (Edit gate overhead for governance docs) — implemented fix, closed. 8 problems remain. ($0.32, 23s, 171K tokens)
+[Iteration 2] Worked P021 (Governance skill structured prompts) — investigated root cause, transitioned to known-error. 7 problems remain. ($0.85, 47s, 432K tokens)
+[Iteration 3] Skipped P016 (Multi-concern ticket splitting) — fix released, awaiting user verification. Worked P024 (Risk scorer WIP flag) — implemented fix, closed. 6 problems remain. ($1.12, 62s, 541K tokens)
 ```
 
 ### Step 6.5: Release-cadence check (per ADR-018)
@@ -313,10 +343,27 @@ The skill should produce a final summary when the loop ends:
 |------|---------|--------|
 | 9.0 | P012 (Skill testing harness) | Open |
 
+### Session Cost
+
+Extracted from each iteration subprocess's `claude -p --output-format json` response (source: measured-actual, not estimated — per ADR-026 grounding). Renders identically in interactive and AFK modes; no decision branch, so output-side only. Cache-read column surfaces the warm-cache-reuse signal observed across subsequent subprocess invocations in the same Bash session.
+
+| Metric | Value |
+|--------|-------|
+| Iterations run | 3 |
+| Successful (committed) | 2 |
+| Skipped | 1 |
+| Total cost (USD) | $2.29 |
+| Mean cost per iteration | $0.76 |
+| Total input tokens | 42 |
+| Total output tokens | 1,531 |
+| Cache-creation tokens | 78,000 |
+| Cache-read tokens (reuse) | 1,064,000 |
+| Total duration | 2m 12s |
+
 ALL_DONE
 ```
 
-When every skipped ticket is in the `upstream-blocked` category (stop-condition #3) or there are no skipped tickets (stop-condition #1), omit the Outstanding Design Questions section entirely rather than rendering an empty heading.
+When every skipped ticket is in the `upstream-blocked` category (stop-condition #3) or there are no skipped tickets (stop-condition #1), omit the Outstanding Design Questions section entirely rather than rendering an empty heading. The Session Cost section always renders when at least one iteration ran.
 
 ## Related
 

package/skills/work-problems/test/work-problems-cost-logging.bats

@@ -0,0 +1,109 @@
+#!/usr/bin/env bats
+# Doc-lint guard: work-problems SKILL.md must extract cost + usage metadata
+# from each iteration's `claude -p --output-format json` response, surface it
+# in the per-iteration Step 6 progress line, and aggregate it in the ALL_DONE
+# Output Format as a dedicated "Session Cost" section.
+#
+# Rationale: the subprocess-boundary dispatch lands per-iteration cost in the
+# JSON response alongside `.result`. Without an explicit extraction contract,
+# that data is invisible to the user even though it's already emitted. Cost
+# logging lets the user calibrate AFK loop sizing on return (e.g. "max out
+# the token usage" direction 2026-04-21 needs a feedback loop).
+#
+# Structural assertion — Permitted Exception under ADR-005 + ADR-037 (SKILL.md
+# is the contract document). A behavioural harness that exercises the `jq`
+# extraction against a fixture JSON is a potential follow-up; out of scope
+# for this doc-lint pass.
+#
+# @problem P084
+# @jtbd JTBD-006
+#
+# Cross-reference:
+#   P084 (iteration worker has no Agent tool) — parent ticket; cost logging
+#     is an additive observability overlay on P084's shipped subprocess
+#     dispatch.
+#   ADR-032 (governance skill invocation patterns) — subprocess-boundary
+#     sub-pattern; `--output-format json` parse shape is already pinned.
+#   ADR-026 (agent output grounding) — Session Cost section cites its source
+#     so downstream audits can distinguish measured-actual from estimated.
+#   ADR-037 (skill testing strategy) — contract-assertion pattern.
+#   JTBD-006 (Progress the Backlog While I'm Away) — "clear summary when I
+#     return" documented outcome includes cost/token traceability.
+
+setup() {
+  SKILL_DIR="$(cd "$(dirname "$BATS_TEST_FILENAME")/.." && pwd)"
+  SKILL_FILE="${SKILL_DIR}/SKILL.md"
+}
+
+@test "SKILL.md Step 5 extracts .total_cost_usd from iteration JSON response" {
+  # Cost per iteration lives in the same JSON blob as .result; parsing it
+  # costs nothing more than a jq call the orchestrator already needs for
+  # ITERATION_SUMMARY.
+  run grep -nE 'total_cost_usd' "$SKILL_FILE"
+  [ "$status" -eq 0 ]
+}
+
+@test "SKILL.md Step 5 extracts usage token fields from iteration JSON response" {
+  # input_tokens / output_tokens / cache_creation_input_tokens /
+  # cache_read_input_tokens are the four usage fields that give a full
+  # accounting. Cache-read is the key signal for reuse across subprocess
+  # invocations in the same Bash session.
+  run grep -nE 'input_tokens|output_tokens|cache_creation_input_tokens|cache_read_input_tokens' "$SKILL_FILE"
+  [ "$status" -eq 0 ]
+}
+
+@test "SKILL.md Step 5 names jq (or equivalent) as the extraction mechanism" {
+  # jq is already implicit in `--output-format json` consumption; naming it
+  # in the SKILL.md prevents bespoke sed/awk reimplementations.
+  run grep -nE '\\bjq\\b|JSON parser|JSON extraction' "$SKILL_FILE"
+  [ "$status" -eq 0 ]
+}
+
+@test "SKILL.md Step 5 scopes the extraction to named fields only (PII guard)" {
+  # Architect advisory 2026-04-21: the JSON response also carries session_id,
+  # model, stop_reason, etc. that should NOT be surfaced in user-visible
+  # output. The extraction list must be explicit so future contributors
+  # don't unconsciously broaden it.
+  run grep -niE 'extract only|only the fields|do not (surface|log|emit)|scoped to (the )?named fields|explicit field list' "$SKILL_FILE"
+  [ "$status" -eq 0 ]
+}
+
+@test "SKILL.md Step 6 per-iteration progress line includes cost marker" {
+  # Example progress line in Step 6 should show the (cost, duration, tokens)
+  # suffix so contributors see the target format.
+  run grep -nE '\$[0-9]+\.[0-9]+.{0,40}(tokens|iteration|s,)' "$SKILL_FILE"
+  [ "$status" -eq 0 ]
+}
+
+@test "SKILL.md Output Format includes a Session Cost section" {
+  # ALL_DONE summary aggregates per-iteration cost across the run. The
+  # section renders in every ALL_DONE — interactive OR AFK — because it's
+  # output-side, not a decision branch.
+  run grep -nE '^### Session Cost|## Session Cost|Session Cost.{0,40}Total' "$SKILL_FILE"
+  [ "$status" -eq 0 ]
+}
+
+@test "SKILL.md Output Format Session Cost table includes cache-read reuse signal" {
+  # Cache-read is the signal for "warm-cache savings across subprocess
+  # invocations in the same Bash session" — empirically observed 65-147K
+  # cache-read tokens on probes 2-4 during the P084 probe sequence. Making
+  # this visible to the user helps them reason about AFK loop cost dynamics.
+  run grep -niE 'cache.?read|cache reuse|reuse signal|cache hit' "$SKILL_FILE"
+  [ "$status" -eq 0 ]
+}
+
+@test "SKILL.md Output Format Session Cost section cites its data source (ADR-026)" {
+  # Architect advisory: the Session Cost numbers are measured-actual (from
+  # each iteration's claude -p JSON output), not estimates. Name the source
+  # so audit / downstream-tooling can trust the numbers.
+  run grep -niE 'extracted from.{0,80}(claude -p|--output-format json|iteration)|source:.{0,80}claude -p|measured.{0,40}(iteration|subprocess)' "$SKILL_FILE"
+  [ "$status" -eq 0 ]
+}
+
+@test "SKILL.md Session Cost section renders in both interactive and AFK modes" {
+  # JTBD-006 Rule 6 check: Session Cost is pure output, no AskUserQuestion,
+  # no policy-authorised action. Must render identically in both modes so
+  # AFK users see the same summary on return.
+  run grep -niE 'Session Cost.{0,160}(regardless|both|interactive.{0,40}AFK|AFK.{0,40}interactive)|output-side|no decision branch' "$SKILL_FILE"
+  [ "$status" -eq 0 ]
+}