triage-security-issues

Triage security issues

Contract and mandatory reuse

Execution covers inventory, complete triage/grouping, then remediation through verified source resolution, subject to narrower user limits. Authoring, reviewing, or testing this skill never starts a campaign or authorizes live alert mutations.

Invoke existing skills that implement the needed capability. Do not duplicate their natural-language procedures or copy their scripts. This rule applies to authors and executing agents. This wrapper owns only security-source acquisition, identity, private accounting, cross-source grouping and source-resolution policy. Inspect available skills/helpers before adding code for a demonstrated gap; do not build another GitHub client, browser-session manager, campaign engine or delivery script. Do not invoke triage-fix-all-bugs: its native Bug input contract differs.

Read repository/local instructions and LCM integration. Invoke lcm-memory for required project context without storing raw findings or credentials. Read procedural-development and its coordination/delivery references. Apply its root lifecycle admission before triage dispatch or unattended waiting. Every worker receives the worker execution brief and must supply its command completion/cleanup evidence.

Read sources, authentication, triage and accounting before collection. These references supply caller policy, not replacement workflows. Before collection or campaign dispatch, verify gh auth status --hostname <host>. If unauthenticated, ask the user to run gh auth login --hostname <host> and wait before continuing. Recheck after they authenticate; do not open a GitHub browser login or acquire another credential source.

Configuration

Invocation instructions, not CLI flags: REPOSITORY defaults to the checkout's canonical GitHub host/owner/repository; CODEX_CSV optionally supplies a snapshot; RESUME_RUN selects an existing private campaign. A repository override must map to the inspected checkout before assessment. Default sources are the four in the description; never silently add Issues, advisories/private reports or new scans.

Copy of the Alpha roster with mandatory security bindings:

ROLE	Model default	Reasoning default	Tier default
TRIAGE	gpt-daybreak-blue-latest	high	default
OWNER	gpt-5.6-sol	medium	default
IMPLEMENTER	gpt-daybreak-blue-latest	high	default
SECURITY_IMPLEMENTER	gpt-daybreak-blue-latest	high	default
ESCALATED_IMPLEMENTER	gpt-daybreak-blue-latest	high	default
REVIEWER_A	cortex-hq/zai-org-GLM-5.3	maximum supported	default
REVIEWER_B	xai/grok-4.6	medium	default
SYNTHESIS_REVIEWER	anthropic/claude-opus-5	medium	default

Resolve <ROLE>_MODEL, <ROLE>_REASONING, <ROLE>_TIER using shared precedence and route preflight. Owner/reviewer models are configurable. TRIAGE, IMPLEMENTER, SECURITY_IMPLEMENTER and ESCALATED_IMPLEMENTER must remain Daybreak Blue, including duplicate adjudication, validation and nested implementation calls. An incompatible model override is a configuration conflict, not a fallback. Inherit shared owner limits, candidate budgets and best-effort tier rules; do not redefine them. Record resolved values and pass them unchanged on resume/handoff.

Skill invocation boundaries

Discover installed skills by these names and read their entrypoints before invoking them. Missing required skills block that capability; do not recreate their workflow inline. External plugin paths are resolved from the active skill catalog, never hard-coded to this author's machine. The parent campaign owns dispatch and grouping; triage-finding runs inline inside its assigned Daybreak worker without nested triage dispatch, dynamic tests, deduplication or source mutation.

Invoke	Input	Required evidence / condition for continuing
lcm-memory	Target project and bounded context question	Applicable context or recorded empty/unavailable result under repository policy
codex-security:triage-finding	Frozen sanitized supplied findings, exact source identities, checkout/SHA and private output destination; reuse its GitHub intake reference through authenticated gh api before freeze	One static verdict per source ID with evidence/proof gaps; no unrequested advisory intake or secret persistence
codex-security:validation	Specific unresolved claims requiring bounded validation, original IDs/evidence, isolated fixtures and private artifact destination	Evidence-backed result per claim or explicit remaining uncertainty; never treat setup failure as disproof
playwright-cli or available Playwright/browser skill	Exact source/repository, collect or approved disposition action, campaign auth contract and private destinations	Verified account/filter, complete export or source readback; login budget and credential isolation preserved
procedural-development	Grouped actionable/uncertain inventory, full S0 ledger, current verified runtime root, same logical run ID and recovery record, spent rounds, resolved roles, private tracker, FOLLOWUP_CHANNEL=private-record, source-resolution and LCM integration contracts	Invoke only after complete triage barrier; accept delivery only against caller audit
codex-security:fix-finding	From procedural-development's assigned Daybreak implementer: approved group plan, constituent claims, isolated worktree, acceptance and evidence	Complete fix/no-change/blocked evidence; return to existing owner/review lifecycle, no second coordinator
codex-security:verify-fix	Only when the user explicitly requests security-fix verification: findings and exact candidate checkout	Read-only per-finding fixed/still-vulnerable/inconclusive result; otherwise use the existing fix/review verification
flock	Only when shared LCM integration requires it: declared resource, authorized coordinator and protected operation	Existing skill's live ownership through verification; no invented campaign locks

Use the existing skill's output contract, linking its private result from the ledger instead of rewriting reports. A delegated skill's narrower boundaries remain in force. Where it cannot meet this campaign's transport, privacy or model contract, report the specific incompatibility and continue independent work only.

Phase handoff

Freeze complete S0, triage every member, and finish centralized cross-source grouping before any implementation. Evidence-backed dismissals may happen during triage, after freeze and fresh source readback. Pass only actionable/uncertain groups to procedural-development, keeping the full denominator privately. The caller retains source dispositions and final audit; the invoked skill owns remediation execution. Neither an empty queue nor a merged PR proves completion.