v1.9.0
Panda Recon is split into two skills, one for each side of security work. The catalog grows from 28 to 29 skills.
Panda Red and Panda Blue
- Panda Red (
panda-red, 0.4.0) covers offensive and assessment work:scan,recon,simulateandbounty. Active testing still needs an engagement that names exact targets, techniques and budgets. Red makes no network calls. - Panda Blue (
panda-blue, 0.4.0) covers defensive work:hunt,incident,verifyandintel. It analyses supplied evidence only and never contacts assessment targets.intelis its only network-capable action, and only with--allow-network. - Each skill refuses the other's modes and names the right skill, for example "hunt belongs to panda-blue". Nothing is silently dispatched.
- Both skills work on their own. Case, asset, evidence and finding records use the
shared
panda-case.*/v1schemas, so a case Red starts can continue in Blue. The shared helper code and schemas are byte-identical copies in each skill, and a test fails if they drift apart. - Red-only records use
panda-red.*/v1(engagement, scan metadata). Blue-only records usepanda-blue.*/v1(intelligence and its cache). - Default output folders are
.panda-suite/red/and.panda-suite/blue/. The intelligence cache moves to.panda-suite/blue/intelligence-cache. - Red's
scan-reportno longer takes--intel. For current CVE, KEV or EPSS status, run Blueintelwith a components file and pass its dated output in. - Panda Lead routes scans, surface mapping, lab validation and bounty work to Red; hunting, incidents, fix verification and current vulnerability intelligence to Blue. An unqualified "scan this project for vulnerabilities" still goes to Panda Security.
Guides: Panda Red · Panda Blue.
Other changes since v1.8.0
- Update check in the npm CLI. In an interactive terminal, an older
panda-suiteCLI checks npm for a newer release and asks whether to run the command with it. CI, offline use, non-interactive runs andPANDA_SUITE_SKIP_UPDATE_CHECK=1skip the check. If the newer version cannot start (for example, Node on Windows refuses to runnpx.cmdwithout a shell), the CLI says so and runs the command with the current version. - Installer fix. Fresh installs no longer fail with
EEXISTon Node 26.13, which rejects copying into an already reserved destination folder. - In-depth skill guides. Each skill has a page under
docs/skills/with when to use it, inputs, workflow, boundaries and example prompts. The documentation site publishes them as the skill pages. - Panda Evaris documents Evaris 0.4.0 campaign verification with Invaris.
- Documentation site: real dark mode and deployment from the
devbranch. - CI runs the Python tests, the skill validator and the npm CLI tests.
- New routing scenario.
RECON-SCAN-ROUTINGchecks that Lead sends a plain vulnerability scan request to Panda Security.
Upgrading
You need to act on two things.
- Remove the old skill. The managed updater keeps skills that the package no
longer ships, so delete any installed
panda-reconfolder after installingpanda-redandpanda-blue. Then restart or refresh your host. - Saved Recon records no longer validate. Records under
.panda-suite/recon/withpanda-recon.*/v1schema versions are rejected by both new helpers. Change theschemaVersionvalues to the new names if you need to keep using them.
Prompts change from $panda-recon MODE … to $panda-red MODE … or
$panda-blue MODE … (plugins: /panda-suite:panda-red, /panda-suite:panda-blue).
Verification
- All 130 Python unit tests pass, including a new shared-contract test. That test
checks the shared files are byte-identical (confirmed by changing one copy and
watching it fail), that neither skill references the other, that a case written
by Red validates in Blue, and that old
panda-reconrecords are rejected. - The skill validator reports 29 skills with valid references, manifests and scenario coverage.
- The documentation build produces 75 pages, including 29 skill guides and these notes, with no links to the removed Recon pages.
- All 13 npm CLI tests pass, including one where the newer version cannot start
and the install still runs with the current version. That test fails without
the fix. The package preview (
npm pack --dry-run) is version 1.9.0 with 156 files: 28 underpanda-red, 30 underpanda-blueand none forpanda-recon. - Each skill installs on its own with
panda-suite install --skilland its helper'shelpandvalidate casesucceed with no other skill present. - Four routing scenarios were run once each through Panda Lead in fresh
subagents. All picked the expected skill: surface mapping → Red
recon, incident review → Blueincident, a plain scan request → Panda Security, and a CVE question → Blueintel.
Known limits
- Windows: the update offer cannot switch versions. Node 20.12.2 and later
refuses to run
npx.cmdwithout a shell, so answering yes falls back to the current version instead of the newer one. Runnpx panda-suite@latest …directly to use the newer version. Windows itself is untested; CI runs on Linux. - Routing evidence is thin. Each routing scenario ran once, graded by the session that made the change. It shows the routes work, not how reliably.
- No behavioural gain was measured from the split itself. Earlier trials found the skills' main benefit is their output contracts and status labels; plain-model reasoning matched them on the existing fixtures.
- The Recon design record and earlier release notes keep their original text; only their links to the removed Recon guide were repointed.
Compatibility and recovery
Skill names changed: panda-recon is removed and panda-red and panda-blue are
added. Other skills are unchanged apart from routing text in Panda Lead and Panda
Security. To return to the previous behaviour, reinstall from
v1.8.0.