Portfolio In-Flight Triage

Implements the audit half of spec/portfolio/portfolio-inflight-management/ as a Claude Code skill in the nolte-shared plugin. Single operation: periodic, on-demand cross-repository inspection that surfaces stalled in-flight work (open issues, open PRs, branches without active PR, unresolved review-comment threads, open Discussions) across every nolte Portfolio-Member repository, classifies each finding through the four-axis severity matrix, attaches a recommended specialist (skill or agent slug) per finding, and persists a dispatch-ready Findings-Report under .audits/portfolio-inflight/<YYYY-MM-DD>.md.

Why this is a skill, not an agent

• Mid-flow user confirmation on threshold overrides and roster-gap escalations. Two interactive checkpoints are load-bearing: honouring a Portfolio-Member's project/inflight.yml threshold override over the spec default (operator sees both side-by-side), and escalating a roster-gap finding to "author a new specialist" once the recurrence threshold in continuous-improvement §Portfolio gap closure triggers. An agent's fire-and-forget shape would lose both.
• Persistent on-disk artefact as deliverable. The audit writes .audits/portfolio-inflight/<YYYY-MM-DD>.md in claude-shared. Skills own persistent write paths; agents return structured reports into the caller's context.
• Context-window-protective data collection via agent. Raw per-repository collection is delegated to portfolio-inflight-collector, which returns a pre-reduced structured summary, so raw issue / PR / branch / review-comment / discussion bodies never enter the orchestrating conversation — the §Findings-Report shape MUST NOT on raw-body inclusion.
• Skill-orchestrates-agent-executes is the explicit spec rule. §Audit operation requires MUST be implemented as a dedicated skill … MUST NOT be implemented as a Claude Agent and MUST dispatch a read-only specialist agent portfolio-inflight-collector. The spec pins the hybrid pattern; this skill is the orchestrator half.
• Counter-dimension considered: a tool-restricted agent could perform classification and matrix-axis derivation cleanly, since the collector already absorbs the data volume. The decisive counter is the mid-flow user-confirmation gates plus the persistent write target — both skill-side per spec/claude/skill-vs-agent/.

Follows the same orchestrator pattern as portfolio-audit (which dispatches portfolio-manifest-collector).

User-language policy

Detect the user's language and respond in it. The Findings-Report under .audits/portfolio-inflight/ follows the conventions of spec/claude/review-plan/: section headings (## Scope, ## Summary, ## Findings, ## Processing log) stay English so downstream tooling can grep them deterministically; the body may be written in the user's language. Finding identifiers (<repo>/<source>/<id>), matrix-axis names (security_relevance, release_blocking, age_multiplier, cross_repo_blocking), severities (Critical / Warning / Suggestion / Info), and bracketed spec citations ([portfolio-inflight-management §...]) stay English verbatim regardless of conversation language.

Detection: am I in the right repository?

This skill writes its single output artefact (.audits/portfolio-inflight/<YYYY-MM-DD>.md) inside the claude-shared repository, the same repository that hosts .audits/portfolio/ for portfolio-audit. Detection: presence of .claude-plugin/plugin.json AND a spec/portfolio/portfolio-inflight-management/ directory.

When invoked inside any other Portfolio-Member repository (the typical adopter), stop and route the operator to claude-shared for the audit — the Findings-Report write path lives there, matching the §Audit operation SHOULD and mirroring portfolio-audit's detection rule.

Operations

1. Run

Runs the cross-repository in-flight audit per spec/portfolio/portfolio-inflight-management/ §Audit operation. End-to-end produces one Findings-Report under .audits/portfolio-inflight/<YYYY-MM-DD>.md.

1. Detect Portfolio-Member set per §Portfolio scope MUST — reuse the portfolio-management resolution. Either dispatch portfolio-manifest-collector (Agent) when the active conversation hasn't already collected the set, or instruct portfolio-inflight-collector to resolve fresh via gh api orgs/nolte/repos --paginate (it applies the same archived: false / private: false / portfolio: excluded filters). Record portfolio: excluded and per-source inflight: skip-<source> markers as Info-grade entries so omissions stay inspectable.
2. Dispatch the collector — invoke portfolio-inflight-collector (Agent) with the resolved list, all four data sources enabled, and release-drafter draft collection enabled (needed for the release_blocking matrix axis). Wait for its structured collection report. Per §Findings-Report shape MUST NOT, raw issue / PR / branch / review-comment / discussion bodies never enter this conversation — the agent reduces to a structured summary and discards bodies before returning. From the same report, derive the consumer-measured-finding count for the repository that holds the portfolio-wide specs: the open issues whose namesOtherPortfolioMember or hasSpecFindingLabel flag is yes (a union, not an intersection — each half alone under-counts).
3. Fetch overrides and confirm — gate (one). For every repository whose collector report carries inflight.yml override present: yes, parse the verbatim YAML from inflight.yml raw content. Present the spec defaults (Issue 30d, PR 7d/14d/14d/conflict, Branch 30d, Review-comment 7d, Discussion 30d) and the per-repo override side-by-side; ask the operator whether to honour the override for this run. Record the decision in ## Processing log. Repositories without overrides use the spec defaults silently.
4. Apply stalling thresholds per §Stalling thresholds using the per-repository thresholds confirmed in step 3:
   • Issue: open longer than issue.maxAge (default 30 days), no priority label, no assignee, no maintainer comment in issue.maxQuietWindow (default 30 days). Per §Stalling thresholds MUST NOT, first drop any issue whose collector author is renovate[bot], dependabot[bot], or github-actions[bot] from the issue/stalled-issue-no-triage class (permanent bot dashboards, not triage signal); branch/stalled-branch-no-pr stays in scope regardless of authorship.
   • Pull request: open longer than pr.maxAgeRedChecks (default 7d) with red required checks, OR pr.maxAgeDraft (default 14d) as draft, OR pr.maxAgeNoReview (default 14d) with no reviewer activity, OR mergeable: CONFLICTING against develop (no threshold for the conflict driver — conflicts block merge regardless of age, per §Stalling thresholds MUST).
   • Branch without active PR: daysSinceLastPush > branch.maxAge (default 30d).
   • Unresolved review comment: daysSinceLastComment > reviewComment.maxAge (default 7d) with no maintainer reply.
   • Discussion: daysOpen > discussion.maxAge (default 30d) with daysSinceLastMaintainerReply == "never" or beyond the same window. Sub-threshold items are excluded per §Stalling thresholds MUST NOT. A Critical finding may still surface a sub-threshold item when another matrix axis demands it per §Stalling thresholds SHOULD.
5. Derive the four matrix axes per §Classification and prioritisation for every stalled item. Read references/matrix-axes-and-report.md when running this step — it carries the detection MUSTs for all four axes.
6. Classify into the canonical four severities (Critical / Warning / Suggestion / Info) per §Classification and prioritisation and spec/claude/review-plan/ §Severity scale. Read references/matrix-axes-and-report.md when running this step — it carries the per-severity mapping and the higher-severity tie-break.
7. Attach a recommended specialist per §Specialist recommendation to every finding:
   • Match the catalog. Read every description line across all catalog roots (skills/*/SKILL.md, agents/*.md, plugins/*/skills/*/SKILL.md, plugins/*/agents/*.md); name the matching slug verbatim (for example dependency-audit, workflow-health-triage, pull-request-merge).
   • Red-check exclusivity. When a PR finding's only driver is a red required check (requiredChecksState: FAILURE), name workflow-health-triage; MUST NOT route the same red-check driver to any other specialist. Other drivers of the same PR (stale draft, unresolved review comments, conflicts) MAY produce additional findings with their own specialists; the exclusivity is per-driver, not per-PR.
   • Slash-command verbatim. When the recommended action targets a specific PR / branch / issue, include the slash-command invocation verbatim plus the target identifier (for example /nolte-shared:pull-request-merge against PR #142). Operator dispatches by copy-paste; the skill never auto-dispatches.
   • Roster-gap recording. When no matching specialist exists, record a Suggestion-grade finding tagged with a stable lowercase snake-case <finding-class-token> (the unmatched specialist slug when one would naturally apply, otherwise a descriptive token like discussion_no_maintainer_reply).
   • Recurrence counting and gate (two). For every roster-gap finding, Glob the prior .audits/portfolio-inflight/*.md and Read the lexicographically last one. Count occurrences of <data-source>/<finding-class-token> across that prior artefact plus the current run. When the count reaches 3 (per spec/project/continuous-improvement/ §Portfolio gap closure), trigger User-confirmation gate (two): ask whether to escalate the recommendation to "author a new specialist <proposed-slug>". Record the decision in ## Processing log. When no prior artefact exists, recurrence is the in-run count only.
8. Pre-write confirmation — gate (three). Before any file write, present the per-severity finding counts (Critical: <n>, Warning: <n>, Suggestion: <n>, Info: <n>) plus the per-repository counts (top 5 + total). Ask Write Findings-Report to .audits/portfolio-inflight/<YYYY-MM-DD>.md with these counts? [yes / show-findings / abort]. On show-findings, render the full per-severity list inline and re-ask. On abort, stop. On yes, proceed.
9. Write the Findings-Report at .audits/portfolio-inflight/<YYYY-MM-DD>.md in the claude-shared repository, conforming to spec/claude/review-plan/. Read references/matrix-axes-and-report.md when running this step — it carries the frontmatter, the four required sections, the structure rules, and the per-finding format.
10. Confirm in the user's language — the path of the new Findings-Report, the per-severity counts, the count of roster-gap findings (if any), the consumer-measured-finding count from step 2 — flagged as a portfolio gap when it reaches the threshold spec/project/continuous-improvement/ §Portfolio gap closure states, naming the action that section names (schedule an issue-batch-orchestrate run over the counted findings) while the decision to start it stays with the operator — and the next step (open the report and triage Critical items first via the recommended slash commands; route roster-gap escalations through continuous-improvement-triage).

The Run operation is observational per §Operator authority: it never mutates any Portfolio-Member and writes only .audits/portfolio-inflight/<YYYY-MM-DD>.md in claude-shared (see Hard rules).

Reference: spec anchors

This skill implements spec/portfolio/portfolio-inflight-management/; read it when in doubt. Sections touched:

• §Portfolio scope, §Data sources, §Stalling thresholds, §Classification and prioritisation, §Specialist recommendation, §Audit operation, §Findings-Report shape, §Integration with continuous-improvement, §Operator authority.

Also touches spec/claude/review-plan/ (the artefact frame) and spec/project/continuous-improvement/ (§Specialist dispatch, §Portfolio gap closure — the matching pattern and recurrence rule in step 7, and the consumer-measured-finding threshold plus its issue-batch-orchestrate action reported in steps 2 and 10).

When the spec disagrees with this skill, the spec wins. Propose a skill update rather than silently diverging.

Examples

• Read examples/01-stalled-pr-with-red-checks.md when a PR's only driver is a red required check (the workflow-health-triage exclusivity routing applies).
• Read examples/02-release-blocker-detection.md when a PR carries a release-blocker label or its head SHA appears in an open release-drafter draft (the release_blocking detection MUSTs apply).
• Read examples/03-roster-gap-3-recurrence.md when a roster-gap class crosses the recurrence threshold against the prior audit artefact (the "author a new specialist" gate applies).
• Read examples/04-security-relevance-lifts-to-critical.md when the security_relevance axis lifts a finding to Critical.
• Read examples/05-cross-repo-blocking-lifts-to-critical.md when a project/roadmap.md / project/sprints/*.md cross-reference lifts an item to Critical via cross_repo_blocking.
• Read examples/06-zero-findings-run.md when the in-flight surface is empty and the run still writes a zero-findings report.

Gotchas

• The Findings-Report write path lives in claude-shared, not in the calling repo. Writing .audits/portfolio-inflight/ from a non-claude-shared working directory is a structural error; confirm cwd resolves to the claude-shared checkout (see §Detection) before any write in step 9.
• cross_repo_blocking detection is exact-string only. The project/roadmap.md / project/sprints/*.md scan looks for the literal short cross-reference (nolte/<repo>#<number>) or the full GitHub URL. Don't infer from titles, branch names, or partial SHAs — per §Classification and prioritisation MUST the scan is exact text matching, no fuzzy matching.
• release-drafter-draft head-SHA membership requires the collector's draft list. When the collector report lacks the Open release-drafter drafts block, only the release-blocker label signal applies for release_blocking — don't fetch drafts inline; the collector owns every gh api call per §Audit operation.
• Threshold-override confirmation happens once per audit, not per-finding. Step 3 surfaces each repository's project/inflight.yml override once; the confirmed thresholds apply to every finding from that repository in step 4.
• Recurrence counts the most recent prior .audits/portfolio-inflight/*.md plus the current run. The step-7 3-recurrence escalation walks only the immediately previous artefact, not older ones; with no prior artefact, recurrence is the in-run count only per §Specialist recommendation MUST.
• Bot-authored dashboards are out of the stalled-issue class. Per §Stalling thresholds MUST NOT and the 2026-07-24 audit's gate-2 decision (#445 / #468), step 4 drops renovate[bot] / dependabot[bot] / github-actions[bot] issues from issue/stalled-issue-no-triage; branch/stalled-branch-no-pr is unaffected.
• gh api rate limits scale with N repos × M open PRs. The collector prechecks headroom (its §Preconditions step 2), but a full run still eats a non-trivial slice of the hourly budget; check gh api rate_limit before back-to-back audits. Tooling (optional GitHub MCP): the collector's data-source reads are MCP-preferred with a gh fallback (spec/claude/mcp-tool-preference/); gh api rate_limit itself stays gh.
• An empty in-flight surface still produces a Findings-Report. Per the spec's S-7 acceptance criterion, zero stalled items across the portfolio produces a valid Findings-Report with zero findings under each severity rather than no report at all. Step 9 always writes, never short-circuits on emptiness.

Hard rules

• Never close any GitHub issue, merge any PR, delete any branch, mark any review comment resolved, or close any Discussion. The audit is observational only per §Operator authority.
• Never invoke gh api with -X POST, -X PATCH, or -X DELETE against any repository — including claude-shared. The underlying agent enforces this at the harness level; the skill also doesn't issue such calls per §Operator authority.
• Never modify any file in any Portfolio-Member repository other than claude-shared. The single write path is .audits/portfolio-inflight/<YYYY-MM-DD>.md per §Operator authority.
• Never dispatch the recommended specialist automatically. Emit the slash-command invocation verbatim in the report; the operator dispatches by copy-paste per §Specialist recommendation MUST NOT and §Operator authority MUST.
• Never use a severity outside the canonical four (Critical / Warning / Suggestion / Info). ALL-CAPS variants and extra levels (HIGH, P0, …) are forbidden per spec/claude/review-plan/ §Severity scale and §Classification and prioritisation MUST NOT.
• Never invent matrix-axis values without the detection signals named in §Classification and prioritisation. release_blocking requires the label or the release-drafter-draft head-SHA membership; cross_repo_blocking requires the exact-string cross-reference scan; security_relevance requires a CVE / supply-chain / leaked-credential indicator. No inference or fuzzy matching.
• Never scan repositories outside the resolved Portfolio-Member set, even when the operator names another nolte/<other> repository, per §Portfolio scope MUST NOT. Scope expansion happens via portfolio-management, never ad-hoc.
• Never include raw issue / PR / branch / review-comment / discussion bodies in the Findings-Report per §Findings-Report shape MUST NOT. The audit reduces every finding to a structured summary; verbatim source bodies stay in GitHub, reachable via the finding identifier.
• Never bypass the three user-confirmation gates: threshold-override honour (step 3), roster-gap 3-recurrence escalation (step 7), and pre-write per-severity-counts confirmation (step 8). The interactive checkpoints are the spec's reason this is a skill rather than an agent; bypassing them undermines spec/claude/skill-vs-agent/.
• Never invoke the Skill tool from this skill's body during the data-collection or classification phase — the collector dispatch is an Agent invocation, not a Skill invocation. Calling another skill sequentially (for example continuous-improvement-triage to route the report) is allowed and matches the spec/claude/skill-vs-agent/ hybrid-pattern SHOULD on chaining.
• When spec/portfolio/portfolio-inflight-management/ disagrees with this skill, the spec wins. Propose updating this skill rather than silently diverging.

Resumability

Per spec/claude/resumable-work/, this skill is resumable: true. State is persisted to .resume/portfolio-inflight-triage/<run-id>.yml after every successful user-approval gate (threshold-override honour in step 3, roster-gap escalation in step 7, pre-write confirmation in step 8) and after each named phase boundary (collector dispatch complete, classification complete). On re-invocation, scan that directory for files with status: in_progress whose inputs: snapshot matches the current invocation; if one matches, prompt the operator with Resume run <run_id> from phase <phase> (last checkpoint <last_checkpoint_at>)? [resume / start-new / discard]. The state-file envelope (schema_version, run_id, inputs, phase, decisions[], status, …) and the fail-closed semantics on schema or YAML errors are load-bearing in the spec; don't duplicate those rules here.