Repository navigation
Releases
Panda Suite v1.9.2
Panda Suite v1.9.2
- Shortens the main instructions for all 29 skills and moves conditional guidance into skill-local references.
- Reduces the risk of Codex truncating Panda Plan's main prompt. A native Codex skill-loading check in Orca reproduced the warning with v1.9.1 and completed without it using the revised standalone skill.
- Preserves the upstream Buddhist Method skill snapshot and attribution alongside the Suite adaptation.
See the installation guide. Actual token usage depends on the host and references loaded.
Panda Suite v1.9.1
Panda Suite v1.9.1
- The documentation homepage separates GitHub plugin installation from npm standalone skill installation and links directly to host-specific instructions.
- The homepage adds a copyable Panda Lead prompt and workflow links into the existing 29-skill directory.
- Site search, mobile navigation, and the saved light or dark theme remain available across the site.
This release keeps the v1.9.0 skill set. See the installation guide and the v1.9.0 release notes.
Panda Suite v1.9.0
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
sharedpanda-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 and
PANDA_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 installedpanda-reconfolder after installing
panda-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 oldpanda-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 → Redrecon,
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 runnpx.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.
Panda Suite v1.8.0
v1.8.0
Panda Recon 0.3.0 adds an on-demand threat intelligence layer. Security work can
now use current, dated vulnerability data instead of model memory, and correlate
it with a project's actual components.
- New
intelmode:latest,update,status,owasp,cve CVE-ID,kevand
incidents. Sources are NVD, the CISA KEV mirror, FIRST EPSS, OSV, OWASP (Top 10,
API Security, ASVS), MITRE ATT&CK and CISA advisories, from a fixed registry. scan,huntandincidentaccept--intel off|cache|live(default off).
Scan writesPROJECT_EXPOSURE.mdonly for components that match by exact
package and version. Hunt and incident keep external intelligence separate from
local evidence and never raise an incident classification from it.- Freshness is visible. Each source has a refresh interval, a private dated cache
and a status of Fresh, Stale or Unavailable. A failed or offline lookup is never
shown as fresh, and the helper exits 3 when any enabled source is not fresh. - Network use is opt-in. The helper contacts sources only with
--allow-network,
sends package names and versions to OSV only with--allow-package-queries,
refuses redirects and never sends source code. - KEV, EPSS and CVSS stay separate. KEV means exploited somewhere, not that this
project was compromised. EPSS is a dated probability. Missing NVD enrichment does
not mean unaffected. intel latestreads one day of NVD changes by default and ranks results:
KEV-listed first, then publisher severity, EPSS and recency. The report shows
the top 50;intelligence.jsonkeeps all. Live on 2026-11-17 it retrieved
1,445 changed records in about 10 seconds, andupdaterefreshed every source.intel latest --severity Skeeps records at or above publisher CVSS severity S.
Records without CVSS stay visible, because missing enrichment is not low
severity, and project exposure is computed before filtering.- An optional scheduled refresh guide shows cron
examples forintel update. Nothing is scheduled by default. - Panda Lead routes current vulnerability intelligence to Recon
intel. Dependency
triage inside a requested code audit stays with Panda Security.
See the threat intelligence reference,
command reference and
CVE lookup example.
Verification
All 122 local unit tests pass, including 32 intelligence tests that use mocked
source responses: aliases, withdrawn records, version ranges, cache expiry,
offline reads, partial failures, rate limits, invalid responses and prompt
injection inside advisories. The Suite validator, documentation build and
internal link check pass. Five new reviewer scenarios cover KEV/EPSS
interpretation, freshness without network, advisory prompt injection, project
exposure and routing.
Live checks on 2026-11-17 retrieved every source successfully. NVD, KEV, EPSS and
OSV: record.
OWASP (Top 10 2025, API Security 2023, ASVS 5.0.0), MITRE ATT&CK (version 20.1)
and CISA advisories: record.
All 9 npm installer tests pass on Node 26.8.2, and the package preview includes
panda-recon.
The five intel scenarios were run once per arm in fresh subagents
(comparison). With the
skill, all 14 checks passed. Without it, 9 passed, 3 were partial and 1 failed
(the 14th, choosing a skill, does not apply without one). The no-skill answers
treated record dates as proof of freshness and silently dropped an injected
advisory description. These are single observations graded by the skill author.
Evidence limits
NVD change volume. About 1,400 NVD records change per day. latest and
update read one day by default, in pages of 1,000, and refuse a window over
8,000 records with NVD_WINDOW_TOO_LARGE instead of truncating it. A live
--days 7 window (9,081 changes on 2026-11-17) is refused this way. A missed-run
backlog can be restarted with intel update --source nvd --days 1, which says
what was skipped (live record).
The OWASP version check matches page text and may break if the wording changes;
a failed match leaves the version unknown (null) rather than guessing. The MITRE
source reads the version index, not the technique catalog, so revoked or
deprecated technique IDs must be checked on attack.mitre.org. One
trial per arm of the new scenarios is recorded; the helper itself was not run
by an agent. Automatic dependency inventory
reads npm package-lock.json v2/v3 only; Go and PyPI need a reviewed component
list. There is no separate GitHub Security Advisories source; GHSA records arrive
through OSV. No scheduler or continuous monitoring is included.
Some topics return less than the full proposal. intel owasp returns each
project's current version, not its categories or what changed. intel incidents
lists dated CISA advisories as leads; when an incident happened and which facts
are confirmed are left for the agent to establish from primary sources. All
topics share one report layout. Container images and runtime versions are not
inventoried.
Compatibility and recovery
The catalog remains at 28 skills. Existing Recon modes, helper commands and v1
records are unchanged; new schema fields are optional. The intelligence cache
lives under .panda-suite/recon/. There are no new required runtime dependencies.
Update or reinstall from tag v1.8.0. To recover a previous installation, reinstall
from v1.7.2.
Panda Suite v1.7.2
v1.7.2
Panda Plan now makes requirement traceability, execution readiness, and future
release evidence explicit.
- Tasks link requirements, dependencies, interfaces, acceptance checks, and
expected evidence. Small changes can keep these links inline in one task. - Verification identifies plausible incorrect implementations and the checks that
reject them, including relevant failure and compatibility conditions. - Plans report execution-ready, partially ready, or blocked scope and identify
the first unblocked task. - A conditional reference defines release, deployment, migration, and
compatibility gates. Gates start with evidence pending. Scope revisions also
reconcile their gates and traceability. - The saved-plan handoff explicitly requests review and an execution mode in the
user-facing response when applicable.
Three scenarios, synthetic faulty/correct controls, a blind-preparation regression,
and frozen instruction/trial records accompany the refinement. The usage guide,
routing documentation, English/Thai release references, and GitHub Pages are
updated. See the refinement audit
and trial results.
Verification
Fresh local checks for this release cover the Suite validator and upstream
hashes, tooling tests, reviewer controls, saved execution artifacts, Codex
plugin/standalone discovery, and documentation generation and links. The site
build produced 50 pages with 25 skill guides; all 3,168 internal links passed
target and fragment checks. Diff whitespace checks passed.
The 32 tooling tests pass. All 25 skills are validated and discovered in both
Codex modes. The 14 recorded hashes and final instruction snapshot match.
The saved campaign and scope-execution suites each pass six tests, and the
saved resume probe passes 30 checks including its 15-test suite. The reviewer
probe rejects the intentionally faulty implementation with the expected
assertion and recovery failures.
Evidence limits
Recorded model trials have mixed results. Both original and candidate handled
core correctness well in paired cases. Early candidates exposed readiness and
handoff omissions, and the final matched handoff pair passed in both arms.
Conciseness and ambiguous-policy handling remain inconsistent. The trial report
preserves these observations and the conditions under which they were obtained.
Fresh model comparisons, fresh subagent execution, Python 3.9, and live-provider
checks were not run for this release. General behavioral superiority, token
savings, and deployment safety are unmeasured. Synthetic provider gates describe
the evaluation fixture.
Compatibility and recovery
The catalog remains at 25 skills. Default and explicit plan paths, Markdown task
checkboxes, chat-only output, native/subagent execution choices, and existing
Execute handoffs remain supported. There are no runtime migrations or new
required runtime dependencies.
Update or reinstall from tag v1.7.2. To recover a previous installation, reinstall
from v1.7.1. GitHub Pages publication is checked through the
Documentation workflow and the live homepage, Plan guide, and release page.
Panda Suite v1.7.1
v1.7.1
Panda Suite credits the original Buddhist Method project by nai0om and documents
the akkaraponph fork used as the Suite's pinned source. The attribution appears
in the English and Thai READMEs and the published source notices.
See the original project and
the fork.
The source notices also identify Udomsak as the copyright holder in the
original MIT license.
The plugin manifests and installation examples advance to 1.7.1. GitHub Pages
includes these release notes, the attribution, and an updated latest-release
section.
Verification
The release candidate based on attribution commit d77b0d4 passed these local
checks using Python 3 and the repository's existing development environment.
| Command | Result |
|---|---|
scripts/validate-skills.py |
Validated 25 skills, references, manifests, and scenario coverage |
scripts/vendor-upstreams.py |
All pinned upstream hashes matched |
python -m unittest discover -s tests -p 'test_tooling.py' -q |
31 tests passed |
scripts/build-docs.py |
Built 49 pages with 25 skill guides |
scripts/check-docs.py |
Checked 3,056 internal links without broken targets or fragments |
git diff --check |
Passed |
These checks verify packaging and documentation for this patch. Behavioral model
evaluations and interactive host invocation were not repeated. The existing
v1.7.0 evaluation limits still apply.
Compatibility and recovery
This patch updates attribution, documentation, and version metadata. The catalog
contains 25 skills. No migrations or additional runtime dependencies are required.
Existing installations can update to the 1.7.1 tag. To recover a previous local
installation, reinstall from v1.7.0. The preceding version remains
available as a GitHub release.
GitHub Pages deployment is verified through the Documentation workflow and the
published release and source-notices pages. A failed workflow or missing published
attribution requires correction before reporting the site update complete.
Panda Suite v1.7.0
Panda Suite adds two standalone specialists, bringing the catalog to 25 skills.
- Panda Setup initializes projects and prepares existing repositories for
agent-assisted development. It checks current tool versions, preserves
existing instructions, reviews ignore rules, and reports observed setup checks.
See the Panda Setup guide and evaluation. - Panda Logo explores, creates, refines, and audits visual identities. It
supports editable SVG, raster exports, favicons, banners, and brand-kit
archives when the host supplies the required tools. See the
Panda Logo usage guide.
Panda Lead can route requests to either skill. Both can also be installed and
invoked independently.
Verification and limits
The local release candidate was checked with the suite validator, tooling unit
tests, Codex plugin and standalone discovery, Claude Code manifest validation,
the documentation build and link checker, and the Logo reviewer controls.
Logo creation and the no-renderer fallback were exercised in blind Codex
trials. The creation trial produced requested files, but its visual verification
claim lacked a recorded viewing action. Additional behavioral scenarios and
interactive plugin invocation remain unrun. See the
Logo trial record and the
Setup trial record for scope and limits.
Compatibility and recovery
The source plugin manifests advance to 1.7.0. Existing installations may need
an update or reinstall to discover the new skills. There are no runtime
migrations or new required runtime dependencies. To revert a local installation,
use the source from v1.6.3.
v1.6.3
v1.6.3
This patch release strengthens panda-write for Thai technical, academic, and
professional documents. It adds focused writing guidance and three blind
scenarios for source-grounded Thai output.
Panda Write
- Convert source fragments into complete, natural Thai prose while preserving
evidence boundaries and exact technical identifiers. - State explicit model/table counts before listing exact names; keep model names
distinct from database tables. - Calculate percentages only from confirmed values and avoid unsupported causal,
statistical-significance, or performance claims. - For DOCX documents, use consistent heading styles, omit a TOC unless requested,
and indent Thai body paragraphs by one tab stop. - When a diagram is requested, create and verify Mermaid source before rendering
and embedding an image. Keep the source with the document deliverables; do not
install a renderer when none is available.
Three blind scenarios cover Thai academic transformations, exact table-count
writing, and the DOCX diagram/heading/TOC/indentation workflow. They complement
the existing Thai status-note and DOCX template cases.
Verification and limits
On the local 1.6.3 release candidate, the structural validator passed for all 23
skills, local references, manifests, and scenario coverage. The documentation
build created 45 pages; the link checker passed 2,616 internal links. The tooling
suite passed all 30 tests, and git diff --check passed.
These checks validate packaging, documentation, scenario preparation, and
repository tooling. Baseline-versus-skill model comparisons for the new Thai
writing scenarios are Not Run; their presence does not establish improved
generated-document quality. The SQL and writing fixtures are synthetic.
Compatibility and recovery
No application runtime dependencies or migrations are introduced. Plugin
manifests advance to 1.6.3. Existing local plugin installations may need an
update or reinstall to load the new instructions. The previous source version is
available at v1.6.2.
v1.6.2
v1.6.2
This release updates planning handoffs and document authoring across the 23-skill
Panda Suite. It includes the changes prepared on feature/1.6.2-rc through the
release commit.
Panda Plan and Execute
- When a user makes an authoritative scope decision, Plan updates the full saved
plan and checks for stale references to removed work. - Execute applies settled scope changes to the plan, removes descoped tasks and
checks, and proceeds with the remaining authorized work. It asks only when
scope is still unclear.
The refinement adds blind cases for Plan and Execute scope changes. Existing
planning and execution cases remain regression controls.
Panda Write
- New documents are grouped under
.panda-suite/docs/<story-slug>/; the skill
asks when the request does not identify a story folder clearly. - Supplied DOCX templates guide formatting, including styles and page layout.
Template content is not treated as factual evidence unless the user says so. - Writing guidance now targets clear, evidence-linked language for product
managers, project managers, and technical leads when that audience is relevant. - On macOS, the documented DOCX export path can use the already available
textutilcommand. The skill does not install conversion software. DOCX
template fidelity still depends on what the available conversion tools can
read and reproduce.
Blind cases cover DOCX template fidelity and the Thai PM/Tech Lead writing
register. These cases are prepared for evaluation; they do not establish a
behavioral improvement by themselves.
Verification and limits
On the local 1.6.2 release candidate, the structural validator passed for all 23
skills, their local references, manifests, and scenario coverage. The
documentation build created 44 pages; the link checker passed across 44 pages
and 2,514 internal links. The tooling suite passed all 30 tests. git diff --check passed.
These checks cover packaging, generated documentation, repository tooling, and
the whitespace state. Scenario definitions and fixture checks are not substitutes
for fresh baseline-versus-skill model evaluations. Behavioral comparisons for the
new refinements remain Not Run.
Compatibility and recovery
No application runtime dependencies or migrations are introduced. Plugin
manifests advance to 1.6.2. Existing local plugin installations may need an
update or reinstall to load the new instructions. The previous source version is
available at v1.6.1.
v1.6.1
This patch release tightens the commit boundaries in panda-git and documents
the updated skill in the generated GitHub Pages site.
Panda Git changes
- A commit request stays within Git operations. It does not authorize code fixes,
refactors, or amending an existing commit. A new commit is the default. - Files larger than 10 MiB require explicit confirmation naming the file before
inclusion in a commit. - Detected secrets and credentials stay out of commits. The skill reports their
paths and concern without exposing values. - Sensitive or credential-like configuration and files that appear ignored,
generated, local, or environment-specific must be listed with a reason and
confirmed path by path before inclusion. Existing staging is not consent. - Added blind scenarios for commit-only scope, large files, secret-bearing
.env
files, and files force-added despite.gitignorerules.
Verification and limits
./.venv/bin/python scripts/validate-skills.py validated 23 skills, their local
references, manifests and scenario coverage. ./.venv/bin/python scripts/build-docs.py built 43 pages; ./.venv/bin/python scripts/check-docs.py checked 2,413 internal links with no broken targets or
fragments. The new behavioral scenarios are present but Not Run; this release
makes no claim that a fresh model evaluation has verified their behavior. Secret
detection is best-effort and is not a comprehensive scan.
Compatibility and recovery
No runtime dependencies or application migrations are added. Plugin manifests
advance to 1.6.1. Installing this source release does not automatically refresh
an already installed plugin cache; update or reinstall the plugin to load the
new skill instructions. Use v1.6.0 to return to the previous
instructions.