Agentic os
🧩 Curated, auditable agent skills for knowledge, mastery, career, and cross-system orchestration.
npx -y skills add giacomoguidotto/skills --skill agentic-osAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- 17 days oldThe repository was created 17 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Coordinate versioned cross-System capabilities. Use when the user invokes /agentic-os scout, /agentic-os pursue, or /agentic-os upskill.
SKILL.md
23.0 KB, ~4.7k tokens by cl100k_base, as published. Nobody here has run it
Agentic OS
Coordinate only the requested cross-System capability. The public surface currently supports:
/agentic-os scout --target <positive integer>
/agentic-os pursue
/agentic-os pursue <opportunity-ref>...
/agentic-os upskill <knowledge-project-ref>...
For scout, reject missing, repeated, non-integer, or non-positive targets. For
pursue, accept no arguments for automatic selection or one or more non-empty
opaque opportunity references. Reject flags, duplicate references, and every
other argument shape. For upskill, require one or more non-empty opaque
Knowledge project references and reject flags or duplicates. Repository roots,
System bindings, profile data, throughput targets, and native configuration are
derived from installed Systems, never accepted as public caller inputs.
Scout contract
Scout builds one fresh, ephemeral Career profile for this invocation, reconciles
only its native delta, and invokes Career discovery and evaluation through the
canonical gateway. Load the bundled Knowledge Request from
resources/scout-knowledge-request.json
and return a document matching
resources/scout-result.schema.json.
Agentic OS coordinates the calls. It does not read Knowledge provider bindings, inspect Career implementation scripts, edit Career files, or reproduce either System's domain rules.
1. Establish fresh readiness
Resolve the installed setup-knowledge-system and lookup capabilities and the
configured Career System root. Do not fetch or silently upgrade either System.
Run a fresh read-only /setup-knowledge-system check for agentic-os.scout.
Require the installed Knowledge interface, lookup, Snapshot Token validation, and
every required role in the bundled request to be ready. Record only the structured
status, capability evidence, and observation time, never personal values or
provider details.
From the Career root, invoke only:
node main.mjs career-system.check/v1 --input -
Supply the versioned capabilities career.profile.check/v1,
career.profile.reconcile/v1, and career.opportunity.discover/v1. Record the
structured readiness evidence and observation time. A missing root, malformed
result, unsafe repository state, unavailable required capability, or blocked
readiness returns a capability-scoped blocked result. Do not guess a fallback or
invoke an internal Career script.
2. Build fresh profile input
Pass the bundled request unchanged to the installed automation-mode /lookup.
Require a valid knowledge-system-interface/v1 snapshot whose caller and
capability are agentic-os and agentic-os.scout.
Map its current role results into a complete career.profile.snapshot/v1 with
exactly these sections:
identityapplication_defaultsopportunity_preferencespositioning_and_proofcommunication_strategy
Preserve every field's value, absent, or unresolved state, visibility,
restrictions, evidence, and provenance. Preserve uncertainty and weighted
preferences. Only explicit hard rejects become gates. Never infer a value from a
search miss, strip a qualifier, broaden the lookup, or manufacture a field.
Keep the Knowledge snapshot and Career profile only in memory or in a permission-restricted temporary file. Remove any temporary file before returning, including on errors. Never write a profile, snapshot, token, cache, receipt, ledger, or resumption record in Agentic OS.
3. Validate drift and reconcile the native delta
Immediately before a dependent write, ask the installed Knowledge interface to validate the opaque Snapshot Token. Never parse, compare, log, or derive meaning from the token.
If the covered Knowledge state changed, discard the complete Knowledge and Career
snapshots. Repeat fresh Knowledge readiness, fresh Career readiness, lookup,
mapping, and token validation once. If it changes again, return blocked with a
scoped persistent_drift action. Do not combine revisions or ask a broader query.
When the token is unchanged, pass the complete Career profile through stdin to:
node main.mjs career.profile.check/v1 --input -
node main.mjs career.profile.reconcile/v1 --input -
Use the same opaque source revision in snapshot.revision and
expected_revision. Treat check as read-only. Let the Career gateway own field
mapping, managed projection isolation, validation, and writes. An unresolved field
may block only its dependent capability; independent safe deltas may remain
reconciled. Never roll back a native delta or touch applications, reports,
attempts, outcomes, follow-ups, offers, observations, queues, or generated
artifacts.
After reconciliation, run a new Career career-system.check/v1 for
career.opportunity.discover/v1. This post-reconcile result, not the earlier
check, authorizes scouting.
4. Invoke native scouting
Invoke only the canonical gateway:
node main.mjs career.opportunity.discover/v1 --input -
Pass a versioned request containing the requested target and native continuation references returned by the Career gateway, if any. Do not pass Knowledge provider details, raw private profile values, or Agentic OS state.
The Career capability owns discovery, evaluation, replacement attempts, locking, native verification, and artifact writes. Continue until exactly the requested number of evaluations succeed or the native capability returns a terminal result. Never invoke scan scripts, batch runners, modes, tracker writers, or other Career internals directly.
Preserve every successful native evaluation and artifact when later work blocks or fails. Do not delete, reset, rewrite, or roll back partial Career work. Resume only through continuation references returned by the Career capability, never through an Agentic OS cache.
5. Return the scoped result
Return agentic-os.scout.result/v1 with capability agentic-os.scout. Include
fresh Knowledge and Career readiness evidence, requested and successful counts,
replacements, failures, native report and tracker references, verification,
profile reconciliation, proposed captures, and blocked actions. Reference native
artifacts instead of copying them and redact private profile values.
Use terminal statuses exactly:
completed: exactly the requested evaluations succeeded and native verification passed;blocked: a recoverable prerequisite, required field, readiness, busy state, or persistent drift prevented safe progress;incomplete: native recovery paths ended after preserving partial work;failed: a schema, protocol, safety, or post-check invariant was violated.
Return one result even when the run stops before native scouting. Unresolved data
and blockers remain scoped to agentic-os.scout; unrelated System capabilities
remain usable. Never claim completion from counts alone without native
verification.
Pursue contract
Pursue selects safe existing Career work, produces native plans and draft packs,
and reviews externally owned waits. Load the bundled Knowledge Request from
resources/pursue-knowledge-request.json
and return a document matching
resources/pursue-result.schema.json.
Agentic OS coordinates only versioned contracts. Every read or update of Career state goes through the canonical gateway. Never inspect Career trackers, reports, pipelines, modes, templates, action files, or implementation scripts directly.
1. Establish fresh readiness
Resolve the installed setup-knowledge-system and lookup capabilities and the
configured Career System root. Do not fetch or silently upgrade either System.
Run a fresh read-only /setup-knowledge-system check for
agentic-os.pursue. Require the installed Knowledge interface, lookup, Snapshot
Token validation, and every required role in the bundled request to be ready.
Optional roles may be absent or unresolved without blocking the run. Record only
structured readiness evidence and observation time, never personal values or
provider details.
From the Career root, invoke only:
node main.mjs career-system.check/v1 --input -
Require fresh readiness for career.profile.check/v1,
career.profile.reconcile/v1, career.opportunity.select-related/v1,
career.opportunity.advance/v1, and
career.opportunity.review-waiting/v1. A missing root, malformed result, unsafe
repository state, unavailable required capability, or blocked readiness returns a
capability-scoped blocked result. Do not guess a fallback or invoke a Career
internal.
If Career reports active Scout, batch, pipeline, or other conflicting work,
return blocked with its native busy evidence. Never infer idleness from files or
processes.
2. Build fresh inputs and validate drift
Pass the bundled request unchanged to the installed automation-mode /lookup.
Require a valid knowledge-system-interface/v1 snapshot whose caller and
capability are agentic-os and agentic-os.pursue.
Map the current profile roles into the complete ephemeral
career.profile.snapshot/v1 used by the Career profile capabilities. Preserve
field state, visibility, restrictions, evidence, provenance, uncertainty, and
weighted preferences. Resolve the current application-throughput target from
job-search-strategy; when it is genuinely absent, request the Career-owned
conservative small-batch default. Never invent a target.
Treat communication-strategy as optional personalization. When it is absent,
unresolved, or unavailable, pass an explicit generic-defaults selection to the
Career capabilities and continue. Never substitute remembered guidance, copy
personalization into Agentic OS, or treat missing optional personalization as a
blocker. Career owns complete generic planning defaults.
Keep the Knowledge snapshot and Career profile only in memory or in a permission-restricted temporary file. Remove any temporary file before returning, including on errors. Never write a profile, snapshot, token, cache, receipt, ledger, or resumption record in Agentic OS.
Immediately before any dependent Career write, ask the installed Knowledge
interface to validate the opaque Snapshot Token. Never parse, compare, log, or
derive meaning from the token. If covered Knowledge changed, discard all derived
inputs and repeat fresh Knowledge readiness, Career readiness, lookup, mapping,
and validation once. Persistent drift returns blocked; never combine revisions.
When unchanged, invoke only:
node main.mjs career.profile.check/v1 --input -
node main.mjs career.profile.reconcile/v1 --input -
Use the same opaque source revision in snapshot.revision and
expected_revision. Treat check as read-only and let Career own the safe native
delta. Never touch applications, reports, attempts, outcomes, follow-ups, offers,
observations, queues, or generated artifacts directly. Run a new
career-system.check/v1 for all three pursue capabilities after reconciliation;
only this post-reconcile evidence authorizes pursuit.
3. Select eligible work
Invoke only:
node main.mjs career.opportunity.select-related/v1 --input -
Pass a versioned request containing the optional caller-supplied opportunity references and the resolved throughput selection. Explicit references narrow scope but never override native eligibility, lifecycle, ownership, or related-opportunity suppression.
Treat the native eligible set as exclusive. Preserve suppressed alternatives and research-blocked groups exactly as references and summaries returned by Career. Never reconstruct candidates from Career files, add suppressed work back, or use an unattended override. Research-blocked groups remain blocked while independent eligible groups may continue.
4. Advance and review through native capabilities
For eligible Agent-owned work, invoke only:
node main.mjs career.opportunity.advance/v1 --input -
Career owns selection order, lifecycle routing, communication planning, artifact generation, evidence checks, and safe projection writes. Agent-owned planning, draft packs, and the internal projection that a draft now exists may proceed without approval when the native result explicitly classifies them as safe.
For due or cold externally owned waits, invoke only:
node main.mjs career.opportunity.review-waiting/v1 --input -
Use only confirmed native attempt and outcome evidence. The capability may recommend waiting, draft a next route, or recommend deprioritizing or discarding. It must not invent a reply, record an attempt, or change a factual lifecycle state. Wait reviews do not count toward the application-throughput target.
Keep every real-world boundary approval-gated. Never submit an application, send a message, click a final action, contact a person, record a follow-up as sent, assert a reply or external outcome, or mark a factual real-world lifecycle change without explicit user approval and the evidence required by Career. A draft does not prove that an external event happened. Return required actions as human approvals, not completed events.
If the native result surfaces a durable Knowledge signal, return it only as a
proposed capture. Any Knowledge write remains behind the installed /capture
approval contract and requires a fresh read before application.
Preserve every native plan, draft, safe advance, and recommendation when later work blocks or fails. Do not delete, reset, rewrite, or roll back partial Career work. Use only native continuation references; never create an Agentic OS cache.
5. Return the scoped result
Return agentic-os.pursue.result/v1 with capability agentic-os.pursue. Include
fresh readiness evidence; selected, suppressed, and research-blocked references;
plans and draft packs; wait recommendations; safe Agent-owned advances; evidence
sufficiency; capacity shortfall; required human approvals; reconciliation;
proposed captures; and blocked actions. Reference native artifacts instead of
copying them and redact private profile or personalization values.
Use terminal statuses exactly:
completed: every usable selected opportunity advanced safely, or a checked genuinely idle run found no useful work; a throughput shortfall may coexist with completion when every usable opportunity advanced;blocked: a recoverable prerequisite, busy state, required field, required capability, or persistent drift prevented safe progress;incomplete: native recovery paths ended after preserving partial work;failed: a schema, protocol, approval-boundary, or post-check invariant was violated.
Return one result even when pursuit stops before selection. Unresolved data blocks only its dependent opportunity or capability. Optional personalization always degrades to Career-owned generic defaults. Unrelated System capabilities remain usable.
Upskill contract
Upskill reconciles current Career demand and Knowledge-owned Upskill Mappings into
Mastery-owned Cycle Proposal lifecycles. Return a document matching
resources/upskill-result.schema.json.
Agentic OS coordinates only the three canonical versioned System interfaces. It
does not inspect a Knowledge provider, Career reports or trackers, Mastery issues
or storage, or any System implementation module.
1. Establish fresh readiness
Resolve the installed Knowledge interface, configured Career System root, and configured Mastery System root. Do not fetch, upgrade, or repair a System.
Run a fresh read-only /setup-knowledge-system check for
agentic-os.upskill. Require knowledge.project.snapshot/v1 to be ready. From
the Career root invoke only:
node main.mjs career-system.check/v1 --input -
Require career.requisite.snapshot/v1. From the Mastery root invoke only:
node main.mjs setup-mastery-system check
Require the standalone Mastery setup result to be ready. Establish
mastery.cycles.snapshot availability from the valid snapshot acquired in the
next step, and establish mastery.cycles.reconcile availability only through its
canonical versioned result when reconciliation is needed. A missing root, unsafe
repository state, malformed result, or unavailable capability returns a blocker
scoped to agentic-os.upskill. Do not guess a fallback, import provider structure,
or invoke an internal System script. When readiness blocks before snapshot
acquisition completes, report summaries only for snapshots actually acquired.
An empty snapshots object is valid when none were acquired; never fabricate
snapshot evidence to fill an unobserved System slot. Report
snapshot_acquisition as not_started, partial, or completed to match zero,
one or two, or all three acquired summaries.
2. Acquire exactly three fresh snapshots
Acquire one complete input set with exactly these revisioned snapshots:
node main.mjs career.requisite.snapshot/v1 --input -
knowledge.project.snapshot/v1
node main.mjs mastery.cycles.snapshot
Send career.requisite.snapshot.request/v1 through the Career gateway. Send the
caller-supplied project references to the installed Knowledge interface in a
knowledge.project.snapshot/v1 request whose caller and capability are
agentic-os and agentic-os.upskill. Never resolve a provider or infer a
project, mapping, Requisite, Capability, or project seam. Let the Knowledge
System establish those values. Treat disabled mappings as resolvable input.
Require each snapshot's exact schema, revision token, and observation time.
Require Career's native status, Knowledge's native capability_status, and
Mastery's native capability identifier mastery.cycles.snapshot; do not require
a status field that a System snapshot contract does not declare. Keep tokens
opaque. Do not parse, log, persist, or derive ordering from them. Keep all three
snapshots only in memory or in permission-restricted temporary files and remove
every temporary file before returning.
An unresolved Knowledge project or mapping blocks only the affected reference. Partial Career coverage blocks mappings whose Requisite is not established, but does not block mappings backed by a returned Requisite. An unknown Mastery Capability blocks only its mapping. Continue independent mappings.
3. Build one deterministic reconciliation plan
Use mapping_key as the sole managed identity. Never synthesize or normalize it,
and never use a title, Capability label, project label, provider identifier, or
array position as identity. Match a mapping to Career demand only when its
Knowledge-owned Requisite key exactly equals a returned Career requisite_key.
For each fully resolved mapping:
- request
proposalonly when the project hasmastery_enabledvaluetrue, the mapping is enabled, and the exact Career Requisite is current; - request
withdrawnwhen Mastery is explicitly disabled, the mapping is disabled, or complete Career coverage establishes that the Requisite is not current; - preserve the existing Mastery lifecycle when absence, ambiguity, partial coverage, a missing Capability, or an invalid prerequisite prevents a safe decision.
Use only the Knowledge-owned Capability key as primary_capability_id. Use the
lexicographically first canonical project reference as project_ref. Render the
display-only title as Cycle: <Capability label> in <Project name> and pass the
Knowledge rationale unchanged.
Prerequisite edges come only from prerequisite_mapping_keys. Validate the full
returned mapping graph before writing. Missing keys and cycles block every
affected dependent mapping without changing its lifecycle. Order actionable
mappings with a stable topological sort: prerequisites first, then higher Career
weighted_score, higher opportunity_count, higher prevalence, lexical
Requisite key, and lexical mapping_key. Treat missing Career ranking values as
zero only when complete Career coverage safely establishes the Requisite as not
current. Never use input array order as a tiebreaker.
4. Revalidate and reconcile through Mastery
Immediately before a write, reacquire the Knowledge Project and Career Requisite snapshots through the same canonical versioned interfaces and require their opaque revision tokens to be unchanged. Equality-check tokens only; never interpret them. If either changed, discard the entire Career, Knowledge, and Mastery input set, rebuild all three fresh snapshots, and recompute once. Persistent drift returns scoped blockers. Never combine snapshot revisions.
Pass the deterministic plan and the Mastery snapshot's opaque revision token to:
node main.mjs mastery.cycles.reconcile --input -
Use mastery.cycles.reconcile.request/v1. Mastery alone creates or updates Cycle
Proposals, preserves active and terminal lifecycles, materializes native
dependencies, validates its revision, and writes operational state. Agentic OS
must not open, edit, close, label, rank, or link a Mastery issue directly.
Accept only mastery.cycles.reconcile.result/v1. Preserve every safe independent
native action if another mapping blocks. Do not retry a native write after an
ambiguous response.
5. Verify the full graph and return one terminal result
After reconciliation, acquire a fresh mastery.cycles.snapshot/v1 through
mastery.cycles.snapshot. Verify the full returned graph, not only changed
mappings: every mapping_key is unique, every safe managed lifecycle matches or
is natively preserved, every nonterminal prerequisite edge is present, no
unexpected managed edge was introduced, and every native claimed write is
observable. A post-check mismatch is failed.
Return agentic-os.upskill.result/v1 with a summary for every System snapshot
actually acquired, the deterministic rank, the native terminal result for every
mapping, write count, full-graph verification, and capability-scoped blockers.
After acquisition completes this is exactly three snapshot summaries. A result
that stops earlier contains only the acquired subset, including none when
readiness blocked before acquisition. Set snapshot_acquisition from the actual
summary count independently of terminal status. Reference native cycles instead
of copying their contents and never return revision tokens.
Use terminal statuses exactly:
completed: all safe mappings reached or preserved their native lifecycle and the full-graph post-check passed;blocked: no write occurred and recoverable scoped blockers prevented at least one mapping;incomplete: safe writes were preserved but at least one mapping remains blocked;failed: a schema, protocol, identity, authority, native result, or full-graph invariant was violated.
An identical rerun after completed must submit the same ordered plan, receive
only native unchanged or preserved results, pass the same full-graph post-check,
report writes as zero, and write nothing in any System. Knowledge writes remain
behind explicit /capture approval, and this capability never requests capture.
Real-world Career actions remain outside this capability.
What ships with it: 10 files
20.1 KB alongside SKILL.md
agents/
- openai.yaml223 B
resources/
- examples/pursue-result.json2.1 KB
- examples/scout-result.json1.3 KB
- examples/upskill-blocked-readiness-result.json1.1 KB
- examples/upskill-result.json1.7 KB
- pursue-knowledge-request.json1.4 KB
- pursue-result.schema.json3.0 KB
- scout-knowledge-request.json1.3 KB
- scout-result.schema.json2.7 KB
- upskill-result.schema.json5.4 KB