continue-claude-work

What the source does

• FilesystemDeclared4 observations
• NetworkDeclared10 observations
• ShellDeclared4 observations
• DependenciesNot pinnedSome versions float
• SecretsFoundSee the findings; values are never shown

Five audit layers

Gates applied: critical finding, instruction override, no behavioural pass, undeclared transfer.

- Private Keys (`-----BEGIN PRIVATE KEY-----`)

| Account + MCP OAuth tokens | macOS Keychain, entry named from the config dir | **Isolated by default**; set `CLAUDE_SECURESTORAGE_CONFIG_DIR=""` to make every profile share one entry, so an MCP serv

exec(compile(resolver_path.read_text(encoding="utf-8"), str(resolver_path), "exec"), owner.__dict__)

'~/.ssh',

'~/.aws',

'~/.gnupg',

### 1. Gather Credentials

- Before a native quota query, read [refresh ownership](references/oauth-guide.md#bind-refresh-ownership-before-native-queries); a nominally read-only query can refresh credentials.

3. Perform a **live authentication call** to `https://open.gangtise.com/application/auth/oauth/open/loginV2` to verify the credentials actually work.

This is empirical proof, captured in seconds, that beats any amount of "it should work". Do this yourself — never tell the user "you can check the log next time".

- **Not present:** continue silently; do not install a hook or tell the user about

- Do not tell the user to reinstall blindly. Verify first.

fixed duration cap or bypass safety checks. Preserve work already completed; use

# Shell alias (add to ~/.zshrc)

__import__(module)

print(f"…", file=sys.stderr)

print('   { "api"…"api_key": "…" } }')

print(api_key)  # Prints: SecretStr(7fb3...DPRR)

print(api_key.get())  # Get actual value when needed

logger.info(f"…")  # Safe! Automatically masked

return f"http://127.0.0.1:{PORT}/?{urlencode(params)}"

The server prints the actual URL. It starts at `http://127.0.0.1:8765/` and tries later ports if the default is occupied.

"id" "A bytes" "B bytes" "Δsize%" "A dur" "B dur" "Δdur%"

if not lines or lines[0].lstrip("").strip() != "---":

if lines and lines[0].lstrip("").strip() == "---":

Audit 0.4.1 · source sha 0f7f7e0023de · 29 Oct 2026Report an issue or request a re-scan

• Publisherdaymade GitHub userOwner of the repository this skill is read from
• Source folderdaymade/claude-code-skillsThe exact folder we audited
• LicenceMITAs the repository states it
• Commit pinned for this verdict0f7f7e0023deA new audit runs when the source changes
• Repository statusNot archivedPushed today
• Sandboxed runNot runSkills are read, not executed; layer L4 is skipped

Its description: Improves an existing application using its code, product contract, authorized work history and competitor evidence. Use for 取长补短、取其精华去其糟粕、参考竞品优化应用 or evolving an app from successful and failed work. Carries authorized…

---
name: app-evolver
description: >-
  Improves an existing application using its code, product contract, authorized
  work history and competitor evidence. Use for 取长补短、取其精华去其糟粕、参考竞品优化应用
  or evolving an app from successful and failed work. Carries authorized changes
  through implementation and user-result verification, including REA and reverse
  Skill composition; use skill-evolver for Skills.
---

# App Evolver

Improve the application's actual user result. Extract transferable mechanisms,
resolve their fit with the current product, implement authorized improvements,
and verify them through the application's consuming surface.

Use this as an orchestration Skill. Keep repository engineering, research,
history access and release operations with their existing owners. For a Skill
target use `skill-evolver`; for competitor intelligence alone use
`competitors-analysis`; for a new product specification use the available
requirements workflow. Do not replace those owners with another execution engine.

## 1. Bind the application and unfinished result

Before research or edits, establish:

- Exact application, repository/ref, platform and runtime or deployment identity.
- Original user request, current product contract and active stage; distinguish
  long-term positioning from the work selected now.
- User-visible improvement, falsifiable acceptance evidence, authorized
  deliverables, protected behavior/data and stopping condition.
- Analysis-only versus implementation authorization, and any unresolved
  product choice whose answer changes the outcome.
- Scope of the learning task. For a broad product or competitor learning
  effort, let the bound device and operator roles and the complete intended
  user result, with its impact judgments, constrain the research questions
  before selecting any single technical scenario for study. Keep a task about
  a single feature or a known defect narrow; do not expand it into a
  whole-product report.

Reuse earlier user decisions; ask in the current conversation only for missing
information or a real unresolved choice. Use the host's question tool when
available. State the evidence, alternatives, consequences and recommendation.
Do not delegate a user decision through another agent.

For multiple applications, bind each contract and acceptance path separately.
A successful design, permission or measurement in one app is an unverified
proposal in another. Do not transfer a desktop contract to its mobile sibling
without checking the sibling's current contract.

Read repository instructions and relevant domain Skills. Pin the code used for
judgment to an immutable ref; inspect current runtime/config separately. Preserve
other writers' work and coordinate overlapping ownership before editing.
On resume, reconcile the latest request, actual artifacts and unfinished evidence;
continue from the first unmet result rather than replaying the original plan.
Apply the same reconciliation when the user corrects a premise, redirects
attention or explicitly changes scope mid-task: update the binding to match,
retract only the conclusions that depended on the superseded premise, and
reuse unrelated already-verified facts. Select the next research question
under the current authorized task. Correcting one example does not promote
that example into the whole task, and a genuine scope narrowing takes effect
rather than being treated as a passing remark.

## 2. Ground opportunities and constraints

Read [evidence and fit](references/evidence-and-fit.md) before selecting changes.
For criticism of interface hierarchy or a fragmented whole-product experience,
use its [structure comparison](references/evidence-and-fit.md#reconsider-interface-structure)
before selecting local visual fixes.

Inspect the existing user journey, assets, services, tests and implementation
owners first. Search available prior work and personal collections when requested
or required by the user's environment. Perform explicitly requested research
lanes; do not silently replace deep research or history mining with a quick read.
Stop a lane once its evidence resolves the decision it was meant to inform.

Use history only within explicit authorization. Retrieve through the owning index
and exact session reader, preserving user/assistant/tool roles and child-agent
prompt/result links. Exclude hidden reasoning. Keep unavailable, stale or
incomplete coverage explicit; do not fall back to an unbounded raw-history scan.
Treat assistant conclusions as hypotheses, not user endorsement or current truth.

For competitors, inspect implementation and current primary evidence beyond a
feature list. Use `competitors-analysis` for repository evidence and the available
research owner for non-repository products. Resolve revisions and distinguish
observed behavior from claims. Treat source access, architectural inspiration,
code reuse and distribution permission as separate questions; reading a repository
does not authorize copying its implementation.

### Compose reverse-analysis Skills when the target needs them

Read [Skill composition](references/skill-composition.md) when investigating a
released app, package, binary or runtime behavior. Select a specialist by the
missing evidence, return its findings to this workflow, and retain the current
product contract, mechanism decision and engineering/acceptance owners.

Use the public Skill routes directly when no personal collection is available.
If an authorized personal favorites integration exists, use it to find additional
mechanisms or framework candidates; verify their actual Skill sources before
adoption. Its absence does not block public routes. This is reference composition:
it bundles no third-party Skill or engine and grants no installation, MCP setup
or target-execution permission.

## 3. Decide what transfers

For each mechanism that could change the next action, record:

| Decision input | Required answer |
|---|---|
| Evidence | Exact source/ref and o

Excerpt, first 6,000 characters of SKILL.md, verbatim from the audited commit. It is text from someone else’s repository, quoted so you can judge it before your agent reads it.

1. 2026-10-29
   ✕BLOCK F 25 / 100

   first audit · source 0f7f7e0023de

Questions about continue-claude-work

• github.com/daymade/claude-code-skills — the audited source at commit 0f7f7e0023de and every finding aboveObserved · 29 Oct 2026

Spotted something wrong? Report a correction · Methodology