Life Manager Affiliate Skill

status: LOCAL_RUNTIME_READY
legacy_migration: MIGRATION_ONLY
execution: MACOS_LOCAL_ONLY

The canonical source is this skills/affiliate directory in the Life Manager
repository. legacy/ contains byte-preserved evidence only; archived files are
never executed by this skill.

Mutable state lives at
${LIFE_MANAGER_STATE_HOME:-$HOME/.local/state/life-manager}/affiliate.
Installed data lives at
${LIFE_MANAGER_DATA_HOME:-$HOME/.local/share/life-manager}/affiliate.

The runtime follows the proven Coconala boundary: an immutable release, mutable
append-only receipts outside Git, an isolated browser profile, and launchd-owned
local wakes. It does not use Railway or an Anicca API redirect. Public content
uses an authenticated provider tracking link directly; clicks are not revenue.

The model boundary uses the repository-global runtime/agent-runner/
commit d150e4b1dba4c207a12fa4be974356124d425919. SOURCE.md, the preserved MIT
license, and SHA256SUMS are authoritative; an installed Gig release is evidence,
not source. The runner is not admitted into loop wake until the Affiliate
routing, privacy, schema, and budget gates in the SSOT are closed.

config/agent-runner.json exposes only two model routes: explicit Terra-high
strategy and a fallback-free, one-use Sol-high repair escalation. Both routes
are live-proven read-only, but remain disconnected from loop wake until the
remaining binary, privacy, schema, context, and budget gates close.

Production callers use scripts/agent_runner.py, never the vendored runner
directly. The gate requires AFFILIATE_CODEX_CAPABILITY_RECEIPT, rechecks the
canonical Codex path, --version, and SHA-256, writes the verified pin beside
the model-call receipts, and refuses to start a provider after any mismatch.
machine_capability_inventory.py creates the private receipt from an explicitly
requested codex_cli; it does not admit arbitrary executables or commands.
Before importing the vendored runner, the gate replaces the parent environment
with an explicit allowlist. Codex receives a fixed system PATH, an Affiliate-
owned HOME and CODEX_HOME, the configured auth-file path, locale/timezone,
and only the named budget variables. Parent API keys, database URLs, browser
routes, and every other ambient variable are absent. Never add a wildcard or
copy os.environ into this boundary.

The same wrapper owns model evidence. It applies 0700 to directories and
0600 to files, rejects symlinks, and removes a stale seal before each run.
After the vendored runner writes its atomic summary, the wrapper validates the
attempt JSONL and atomically writes evidence-seal.json with the summary,
attempt, exact result, source-set, and execution hashes/data. Consumers call
verify_evidence_seal with the expected source set; raw output without a valid
seal is incomplete evidence, never a model result.

F0 provides four deterministic, non-publishing primitives:

• bootstrap/install.sh verifies a reviewed pinned manifest and writes an
  atomic machine-capability receipt;
• scripts/authority_inventory.py binds a Keychain reference to one explicit
  intent and records human-only external challenges;
• scripts/profile_provisioner.py creates isolated EN/JA browser roots without
  launching or copying a browser session.
• scripts/machine_capability_inventory.py receipts an explicitly requested
  macOS browser app from held no-follow file descriptors without launching it;
  generic executable admission remains fail-closed.

The first provider CLI slice is live:

skills/affiliate/affiliate provider inspect \
  --provider hubspot-impact --cdp-port 9327 --receipt "$RECEIPT"

It attaches read-only to the dedicated impact-en:9327 browser, selects exactly one
origin/title/path-bound tab, maps rendered text through the versioned provider
playbook, and writes an atomic sanitized receipt. Unknown UI never becomes an
approval. The current slice observes status only; it cannot submit or publish.

Use provider poll with the same arguments in a loop. The first observation or
a real state change returns changed=true and a deterministic transition_id;
an unchanged retry returns next_action=NO_STATE_CHANGE. Downstream actions
must deduplicate on transition_id.

Use credential-first resume only on a supported semantic playbook:

skills/affiliate/affiliate provider resume \
  --provider elevenlabs --cdp-port 9324 --receipt "$RECEIPT"

It reads the mode-0600 Git-external Markdown, clears and fills the named controls
through CDP, submits at most once per invocation, and requires rendered readback.
Credentials never enter stdout, receipts, Git, selectors, or command arguments.

Impact device verification is also a provider command, not an owner task:

skills/affiliate/affiliate provider verify-device \
  --provider hubspot-impact --cdp-port 9327 \
  --receipt ~/.local/state/life-manager/affiliate/providers/hubspot-impact-device.json

It reads only inbound messages from the playbook-bound sender in the local
macOS Messages database, requires exactly one six-digit code newer than ten
minutes, submits it once, and retains neither the code nor message contents.
Missing, stale, or ambiguous codes fail closed.

Recover an expired Impact account only in its dedicated browser. Save the new
credential before the provider mutation, then require a fresh authenticated
application readback:

python3 -c 'import secrets; print("A!" + secrets.token_urlsafe(36))' | \
  skills/affiliate/affiliate programs store-credential \
  --id hubspot-impact --label Impact --verification SAVED_BEFORE_SUBMIT
skills/affiliate/affiliate provider reset-password \
  --provider hubspot-impact --cdp-port 9327 \
  --receipt ~/.local/state/life-manager/affiliate/providers/hubspot-impact-password-reset.json
skills/affiliate/affiliate provider resume \
  --provider hubspot-impact --cdp-port 9327 \
  --receipt ~/.local/state/life-manager/affiliate/providers/hubspot-impact.json

reset-password requires the exact official reset page, exactly two password
fields, and the versioned submit control. It journals the effect before submit
and never prints the password. Redirect to a different provider page proves only
reset acceptance; resume must still prove the authenticated application state.

Install the shared Life Manager runtime and reconcile its registry-owned loops:

scripts/bootstrap.sh
~/loops/current/bin/lm-loop reconcile shared-agent-runner --include-running --loop-id affiliate-browser
skills/affiliate/affiliate loop wake
skills/affiliate/affiliate loop placement --placement article-1 --locale en

The shared lm-loop control plane owns every launchd job. No Affiliate-specific
installer or external Python environment exists. ai.anicca.affiliate-browser
owns the isolated provider profile on CDP 9324;
ai.anicca.affiliate-x-browser owns the English X profile on CDP 9326.
ai.anicca.affiliate-composition consumes at most one due credential-free
source bundle per wake and writes a sealed terminal composition receipt.
ai.anicca.affiliate-loop wakes every 10 minutes. Receipts live under
~/.local/state/life-manager/affiliate; provider passwords and the executable
ElevenLabs link remain only in the mode-0600 private Markdown. The current wake
polls the rendered ElevenLabs login state, records only a deterministic provider
transition ID, and requires AUTHENTICATED before publication readiness. It
still proves readiness only: publication, provider click readback, commission,
and payout stay separate later gates.

Publish an already approved English X artifact only through the dedicated
Affiliate browser:

skills/affiliate/affiliate x post publish \
  --content "$ARTIFACT" --placement elevenlabs-en-1

The command verifies @selawmqt on CDP 9326, requires an explicit affiliate
disclosure and one owned aniccaai.com/blog/ CTA, persists an effect-possible
fence before clicking, reconciles an identical timeline post on retry, and only
returns LIVE after an exact post-page readback. It never prints the post body
or a provider tracking link. A content policy decision and live owned article
remain prerequisites; the command does not generate or approve copy.

Refresh the versioned ElevenLabs evidence plan without an LLM-dependent parser:

skills/affiliate/affiliate sources capture --plan elevenlabs-en

The admitted routes are installed CRWL for official web pages and gh api for
the official SDK. Raw artifacts and deduplicated receipts stay under the
Git-external Affiliate state root. Each receipt binds locale, evidence class,
license, parser version, body hash, observation time, and expiry; adapter failure
is fail-closed rather than converted into an empty source.

Build the first source-bound English article into private runtime state:

skills/affiliate/affiliate content build

The builder requires fresh official support markers and the executable private
ElevenLabs link, places disclosure before the first CTA, and prints only the
artifact identity and hash. The article body and tracking link remain mode-0600
Git-external state until the later policy and owned-publication boundary.

Before publication, issue the deterministic policy receipt and exact-once
placement intent:

skills/affiliate/affiliate content policy
skills/affiliate/affiliate loop placement \
  --placement elevenlabs-plans-for-solo-creators --locale en

The policy command fails closed unless the artifact hash and fresh source hashes
match, the disclosure precedes the first CTA, exactly one owned HTTPS
try.elevenlabs.io link exists, and forbidden income guarantees are absent. It
stores no tracking URL in its receipt. owned publish independently requires the
matching PASS receipt and later reads both disclosure markers and the exact link
back from production HTML.

After the owned article has a LIVE receipt, build and publish its disclosed X
artifact:

skills/affiliate/affiliate content build-x
skills/affiliate/affiliate x post publish \
  --content ~/.local/state/life-manager/affiliate/x-content/elevenlabs-en-1.txt \
  --placement elevenlabs-en-1

Both commands require the exact owned publication receipt to be LIVE; the X
publisher rechecks it before opening the composer.

After a generic campaign reaches owned/X LIVE, the same money owner may
syndicate one full English guide per 24 hours to the configured DEV account:

skills/affiliate/affiliate distribution publish-devto \
  --plan elevenlabs-discovered-audio-to-text-en

This adapter is a copy+tweak of Writer's proven Forem transport. It finds an
existing stable placement marker before any POST, uses the official
canonical_url field to identify the owned article as the SEO source, preserves
the article's affiliate disclosure, and requires API plus anonymous public
readback before writing a LIVE receipt. The shared external-effect journal
prevents a blind repeat after an ambiguous response. A daily cooldown prevents
the ten-minute money wake from turning a valid distribution lane into bulk spam.
The current host reads DEVTO_API_KEY from the process environment,
~/.config/anicca/affiliate.env, or the already-provisioned Writer private env;
the key never enters a receipt, log, prompt, or Git.

The existing ten-minute money owner also reads the authenticated Forem article
list at most once per hour and stores Affiliate-only page views, reactions, and
comments under distribution-metrics/devto.json. These are acquisition
diagnostics, not revenue. The daily Telegram summary reports the real DEV view
count beside provider clicks and approved commission so a zero-click result can
be separated into a reach problem or a conversion problem. The receipt marks a
publication's 24-hour reach baseline READY; the choice of what to improve next
remains an agent decision rather than a deterministic content rule. The first
eligible observation is frozen once under distribution-baselines/; later
hourly polls cannot rewrite the evidence used by that decision.

The same wake passes each new immutable baseline to the bounded acquisition
Agent exactly once. With no eligible receipt it returns WAITING_FOR_BASELINE
without invoking a model. The Agent chooses one acquisition variable, records a
falsifiable hypothesis, one next-campaign instruction, and one success metric
under acquisition-decisions/; it does not publish, edit, or infer revenue.
Telegram reports the resulting decision in natural language from that receipt.
The next discovered plan consumes one unused decision and carries its immutable
experiment envelope through source-set hashing, sealed composition, policy,
owned/X publication, and DEV/Substack receipts. Composition receives the prior
sealed campaign as the control and may change only the Agent-selected variable;
missing control evidence fails closed before any public effect.

Every future generic campaign acquires its own PartnerStack custom link before
publication by reusing the verified ElevenLabs link adapter. The raw URL stays in
the mode-0600 private Markdown; public receipts retain only provider identity and
fingerprints. Each revenue reconciliation rebuilds placement-ledger.json from
real campaign, DEV exposure, provider-link click, and commission receipts.
Unavailable denominators and costs remain null/UNKNOWN, never inferred zero.

The same distribution command exposes the current Writer Substack API shape
without importing Writer state or its retired manual sentinel path:

skills/affiliate/affiliate distribution publish-substack \
  --plan elevenlabs-discovered-audio-to-text-en

The adapter verifies profile/self ownership of aniccabuddha.substack.com,
extracts the already-public owned <article> HTML instead of adding a Markdown
renderer dependency, preserves disclosure and the exact tracking link, creates a
newsletter draft with should_send_email=false, publishes with send=false, and
requires authenticated draft readback plus anonymous public title/disclosure
readback. It uses a separate one-per-24-hour receipt and the shared external-
effect journal. SUBSTACK_SESSION_COOKIE is read from the process environment or
the same private env locations as DEV and never enters receipts or Git.

Build the second source-bound product campaign with the same publisher and X
adapter:

skills/affiliate/affiliate sources capture --plan elevenagents-en
skills/affiliate/affiliate content build-agents
skills/affiliate/affiliate content policy-agents
skills/affiliate/affiliate loop placement \
  --placement elevenagents-for-customer-support --locale en
skills/affiliate/affiliate owned publish \
  --slug elevenagents-for-customer-support --landing-root "$CLEAN_PRODUCTION_WORKTREE"
skills/affiliate/affiliate content build-x-agents
skills/affiliate/affiliate x post publish \
  --content ~/.local/state/life-manager/affiliate/x-content/elevenagents-en-1.txt \
  --placement elevenagents-en-1

This campaign requires the four fresh official ElevenAgents captures and the
private product-specific link. It does not create a new CMS or X publisher.

Capture the official PartnerStack overview after its one-time account,
email-verification, partnership, and program-terms bootstrap:

skills/affiliate/affiliate revenue observe
skills/affiliate/affiliate revenue capture
skills/affiliate/affiliate revenue reconcile

The observer extracts rendered bilingual metric cards, preserves the immutable
initial BASELINE_ONLY values and timestamp, reports later deltas, keeps
unavailable approved/reversed amounts as null, and returns the browser to
ElevenLabs home so the existing provider wake continues to work. The initial
total is never retroactively assigned to a placement.

revenue capture reads the rendered PartnerStack Commission Report and Payouts
surfaces, verifies their real field schemas, and writes a mode-0600 raw rendered
artifact plus a sanitized hash receipt under Git-external state. An empty report
is recorded as EMPTY; it is never replaced by fixtures or counted as money.
The authenticated report JSON is the row-count authority. Its provider-native
reward_key, raw status, amount, and attribution fields are normalized inside
the private artifact; customer name and email never enter the normalized ledger.
revenue reconcile verifies the source artifact hash and appends one stable
transition per provider key, raw status, amount, and source hash. Replaying the
same report is a no-op; a later provider status creates a new append-only event.
It joins a row to one LIVE owned placement by sub-ID/shared-ID or tracking-link
fingerprint. Zero or multiple candidates remain explicit UNMATCHED or
AMBIGUOUS; raw tracking links and customer PII never enter the ledger.

The 10-minute local wake invokes observe → capture → reconcile only when the
provider is authenticated and its independent one-hour revenue cooldown is due.
The cooldown receipt is written after all three commands succeed; a partial
failure is retried by the next wake rather than hidden for an hour.

Each wake also derives at most one owner-readable semantic transition. It first
appends that transition to telegram-outbox.jsonl, then sends it through Life
Manager's repository-owned shared Telegram client. A successful delivery is
recorded in telegram-sent.jsonl with the provider messageId; an acknowledged
transport failure remains quarantined without a blind retry. Stable event
UUIDs suppress repeated empty-report/status notifications across later wakes;
tracking links and customer PII are never included.

Every supported external mutation is wrapped by scripts/job_journal.py.
Immediately before provider login submit, X profile save, X post publish, owned
Git push, or Telegram send, the adapter persists a mode-0600 job containing
run_id, job_id, state, attempt, action_fingerprint, cooldown, and the
last verified external object. Only a semantic readback changes it from
EFFECT_STARTED to VERIFIED; an unresolved effect refuses a blind retry.
On a later process, a fresh provider/profile/post/publication/Telegram readback
may reconcile exactly one unresolved target. It keeps the original run_id and
job_id, increments attempt, and records resumed=true; zero or multiple
unresolved effects cannot be silently retried.

Build and deliver the non-affiliate English foundation article through the same
installed skill:

skills/affiliate/affiliate content build-foundation
skills/affiliate/affiliate owned publish \
  --slug how-to-test-ai-voice-tools-before-you-pay \
  --landing-root "$CLEAN_PRODUCTION_WORKTREE"

owned publish accepts only a hash-valid READY_FOR_PUBLICATION artifact,
writes one deterministic blog JSON target, refuses unrelated worktree/index
changes, commits and pushes only that target, then records DELIVERED until a
later tick reads the title and three immutable markers from the public page.

Use the versioned English program research before any application:

skills/affiliate/affiliate programs list
skills/affiliate/affiliate programs next --decision READY_NO_REVIEW
skills/affiliate/affiliate programs credential --id hubspot-impact
python3 -c 'import secrets; print("A!" + secrets.token_hex(24))' | \
  skills/affiliate/affiliate programs store-credential \
  --id elevenlabs --label ElevenLabs

The registry stores only official-source eligibility and the latest receipted
application decision. Execute its next_action, then require authenticated
rendered readback before changing application or tracking-link state. Never
bulk-apply past an audience, content, fit, or traffic gate.

Provider passwords are never committed. The mode-0600 Git-external local
~/.config/anicca/affiliate-credentials.md is the recovery SSOT and is written
before any signup/reset submit. Each program may also bind to a fixed
keychain://service/account mirror. programs credential checks that mirror:
only a non-empty value is VERIFIED_NONEMPTY; a present but empty item is
MISSING_OR_EMPTY and login stays disabled. Impact uses
keychain://ai.anicca.affiliate.provider.impact/primary. After official recovery,
write the new value there and prove a fresh-tab login before resuming Grammarly.
store-credential reads the secret only from stdin, writes the Git-external
mode-0600 private Markdown first, then mirrors it to the fixed program Keychain
reference, and returns status only. Run it before every signup/reset submission;
after fresh login repeat with --verification VERIFIED_LOGIN. Never pass a
password on the affiliate CLI command line.

If an imported account has a stale or descriptive Login field, replace it from
an authenticated owner source through stdin without rewriting its password:

AUTHORIZED_LOGIN_SOURCE | skills/affiliate/affiliate programs store-login \
  --id hubspot-impact --label Impact

store-login atomically updates only the named mode-0600 Markdown section and
returns state, never the login value. Browser automation must not guess an email
or reuse a password as a username.

Store a provider-generated product link without exposing it on the command line
or in command output:

PRODUCT_LINK_SOURCE | skills/affiliate/affiliate programs store-link \
  --id elevenlabs --label ElevenLabs --field "ElevenAgents affiliate link"

store-link accepts one HTTPS URL through stdin, atomically adds or replaces the
named affiliate-link field in the mode-0600 private Markdown, and returns status
only. Never commit or print the referral URL.

When an approved program delegates reporting to a separate network dashboard,
bootstrap a separate login section without reusing the program password:

PASSWORD_GENERATOR | skills/affiliate/affiliate programs store-credential \
  --id elevenlabs --label PartnerStack --source-label ElevenLabs \
  --credential-ref keychain://ai.anicca.affiliate.provider.partnerstack/elevenlabs

The source label contributes only the existing login identifier. The new password
is read from stdin and saved to the new private Markdown section before Keychain.

The committed bootstrap manifest pins PBS CPython 3.14.7+20260814 for macOS
arm64 by immutable URL and SHA-256. The installer verifies and extracts the same
held artifact, validates the full runtime tree, and atomically activates it
without changing PATH or the system Python. The current Mac also has a receipted
CloakBrowser app. Affiliate-only Keychain readback is live-proven; an unverified
reference is never AUTHORIZED. Authenticated X account-handle verification,
publication, provider commission readback, and revenue remain gated; launchd and
the isolated local browser are owned by this installed skill.