Skills

DivineSWE exposes its orchestration capabilities as skills — named behaviors with a clear purpose, trigger condition, and output. Claude Code exposes them as slash commands; Codex exposes them as $skill-name mentions (or through /skills).

The default path runs through two of them: avatar-forge builds a change in the live session, and oracle-envision — the optional first stage — designs an involved one into a forge spec of Mandates for it to build. The Book cycle — oracle-consult book, then prophet-fulfill — is deprecated, along with its helpers prophet-resume and prophet-revelation; they still run only for a Book already in flight.

For Codex, use the bare skill name after $, for example $avatar-forge or $oracle-envision, or type /skills and select it. The /divineswe: namespace is Claude Code-specific and does not invoke a Codex skill.

This is the complete skill reference, parallel to the Agents reference for the agent council.

Skill Reference Table

Skill	Claude Code command	When to Invoke	What It Produces
oracle-envision	`/divineswe:oracle-envision`	Before forging a new project, feature or involved change you want designed with you first	A forge spec of Mandates — researched, brainstormed, swept for gaps and opportunities, doubted once and triaged — for `/divineswe:avatar-forge` to build; or, in report mode, a fact-checked research report
oracle-consult	`/divineswe:oracle-consult`	Deprecated. Kept for Books already in flight; design new work with `oracle-envision`	The Book door: a Book (`books/book-NN.md`) with ordered prophecies, a decision record and acceptance criteria, for `prophet-fulfill`
prophet-fulfill	`/divineswe:prophet-fulfill`	Deprecated. Kept for Books already in flight; build new work with `avatar-forge`	Completed prophecies — tests written, implementation merged, verified, and sealed
root-cause-analysis	`/divineswe:root-cause-analysis`	Bug, crash, exception, unexpected behavior, or recurring failure	A root cause diagnosis, for `avatar-forge` to fix
canon-drift	`/divineswe:canon-drift`	Deprecated. Audited whether the deprecated Canon still agreed with the code; no replacement	A read-only drift report whose findings become vows
hermes-dispatch	`/divineswe:hermes-dispatch`	Summon a foreign model outside the Book cycle	A guarded ad-hoc carrier dispatch and its testimony
avatar-forge	`/divineswe:avatar-forge`	The default build path: forge an ad-hoc ask, or a spec from `oracle-envision`, in the live session	Commits per unit, each major one audited by 4eyes; Fulfilment lines in the spec; a scroll and paired `decision` ledger events per summons
offering	`/divineswe:offering`	Stage a screened finding for explicit upstream feedback	A locally audited Offering and a human-opened issue URL
orpheus-sing	`/divineswe:orpheus-sing`	Compose a consultation seed from an idea and repository evidence	A validated consult prompt or cross-repository handoff
prophet-full-audit	`/divineswe:prophet-full-audit`	Pre-release review, post-merge audit, or standalone adversarial inspection	One report at `.divineswe/audits/<slug>.md`: CRITICAL/HIGH/MEDIUM/LOW findings and a Forge Prompt for avatar-forge; no IMPLEMENT phase
prophet-resume	`/divineswe:prophet-resume`	Deprecated with the Book cycle: an interrupted Book already in flight; forge work resumes by re-running `avatar-forge` on its spec	Resumed orchestration from the last checkpoint; no lost work
prophet-revelation	`/divineswe:prophet-revelation`	Deprecated with the Book cycle: the single-invocation auto-chain into a Book	Fully chained cycle: a feature routes to `oracle-consult book`, then `prophet-fulfill`; a bug routes to `root-cause-analysis`, whose chain now hands the fix to `avatar-forge`
hermes-setup	`/divineswe:hermes-setup`	Once per machine, offered by `divineswe-init`, and again whenever you want to change or switch off the second opinion	`second_opinion_model` written to `~/.divineswe/config.json`, after it says where prompts go: the model of another family avatar-forge’s closing audit runs as a second 4eyes
divineswe-init	`/divineswe:divineswe-init`	First time in a project, to pre-configure before your first forge	Bootstrapped `.divineswe/` directory structure (state files, ledger, session state, config)
update-plugin	`/divineswe:update-plugin`	When a new DivineSWE version is available	Updated plugin from marketplace + re-bootstrap if version changed
atlas-explore	`/divineswe:atlas-explore`	Before planning, when you need a read-only survey/map of the codebase	A navigable map of the repository — invokes the Atlas explorer agent across multiple angles and synthesizes their findings
scry-project	`/divineswe:scry-project`	Deprecated — use `vow surface` to read what is owed	One honest report of the owed work and the health of the pipeline, with `unknown` never rendered as zero
vow-capture	`/divineswe:vow-capture`	The moment an obligation surfaces mid-session and would otherwise be lost	One line on the Vow register, defaulted for the fast path, surfacing in the next `vow surface`

Skill Details

canon-drift (deprecated)

Deprecated with the Canon, and there is no replacement. It was a read-only audit of the deprecated docs/canon/ declarations against the code: it relayed drift as vows and never authored or repaired the deprecated Canon.

hermes-dispatch

Use $hermes-dispatch for an ad-hoc foreign-model summons outside the Book cycle. Hermes carries either a codex/<model> or openai/<model> request, requires an explicit endpoint for the OpenAI-compatible channel, and warns that prompt files are delivered verbatim to a model with write access. Main worktrees, guarded or integration seats, and unverified endpoints require itemized human-confirmed risk acknowledgments. The carrier returns testimony; it does not promote edits into project state.

avatar-forge

Use $avatar-forge when the god should come down and do the work itself — no Book above it, no phase machine beneath it, just the model in the conversation descending into the tree and forging a change by hand, tests first, while the pantheon attends. It is Hermes’s nearer sibling among the ad-hoc doors, not the cycle’s: it opens a real scroll and marks decision events at the summon and at the close, and unlike a Book that is the whole of what it writes — no phase, no status, no session state, no checkbox; unlike Hermes, which carries a foreign summons and adjudicates nothing, the Avatar can call thoth, monkey, 4eyes, occam, loki-break, loki-exploit, loki-doubt, styx, or mason one at a time, each attendant handed one brief, filing one document, and done. Judgement never leaves the Avatar’s hands: it reads each seat’s testimony itself and decides what happens next.

Each 4eyes loop runs in audit rounds, and the hammer falls on less each round. Only a core finding is fixed — one that leaves a Mandate’s Accept: line unmet, or breaks what worked, on an input the system can actually meet; a nice-to-have or an unreachable what-if becomes a vow instead. 4eyes types each finding; the Avatar may override a type when the work needs it, and every override is named in the record you check. Round one fixes every core finding, round two MEDIUM and up, and from round three only HIGH and CRITICAL. A re-audit reads the fix rather than the whole unit again. There is no cap: a core HIGH still arriving from round three comes to you, and you rule whether to narrow the Mandate, change the design, or carry on.

At the end, after the native closing 4eyes, it runs the second opinion: a second 4eyes on a model from another family than the implementer’s, chosen once with /divineswe:hermes-setup, one after the other and never at the same time. Their findings are one set, both re-audit after any fix, and the closing record names both verdicts and the model. When no model is configured, or the second opinion is refused or fails, the closing goes ahead on the native verdict and records that the second opinion did not run. Only the claude-code package carries the Hermes carrier: copilot-cli will never ship it, and the codex-cli and opencode packages do not carry it whole, so on those hosts avatar-forge’s closing record says the second opinion did not run.

A forge spec whose work changes something you use — a skill you invoke, a command you run, a page you open — carries a human testing sequence: the numbered steps you take using the built work the way you would, each saying what to do, what you should see happen, and which Mandates it exercises. It is a functional walk, not a second proof: what a script or a test can prove stays proven in the spec’s ## Fulfilment and is never walked again, so not every Mandate needs a step, and a spec whose work you never use directly — a refactor, a test, a CI job — carries none. oracle-envision writes it, and the Avatar keeps it current as it builds: a commit that changes what you would do or see for a Mandate rewrites the steps that cite it, and scripts/verify-testing-sequence.mjs refuses a section with no step or a step citing a Mandate the spec does not declare. At the end you walk it beside the evidence table. When every step passes, the spec becomes human-verified; when you accept the evidence without walking it, or the spec has no sequence, it is fulfilled, and bringing a fulfilled spec back to avatar-forge takes it straight to the walk — writing one first when it has none — or, when its work is nothing you use, tells you there is nothing to walk. A step that fails is a finding the Avatar fixes and has re-audited before you walk it again.

offering

Use $offering to stage a locally discovered law or script opportunity for maintainer review. The rite is opt-in and deny-all by default: pick, prepare, display, redact, confirm, lint, and build a URL in order. No network request is made automatically; a human reviews the exact payload and opens the resulting link explicitly.

orpheus-sing

Use $orpheus-sing to turn a raw idea into a plan-time consultation seed. Orpheus interviews when needed, offers divergent framings, gathers repository evidence, scaffolds a consult prompt or handoff, fills its judgment sections, and validates the result. It writes to the resolved consultation dwelling and never substitutes for the run-time dispatch briefs produced by compile-prompt.mjs.

oracle-envision

Before the forge is lit, the Oracle looks ahead. oracle-envision designs a new project, a feature or an involved change with you: it frames the problem, its appetite and its hard constraints, grounds itself in the code (and, for outside problems, in prior art from the web, every source cited), runs a throwaway probe you confirm when nothing written can settle a fact a decision rests on, offers up to three genuinely different, plausible concepts with one recommended pick, sweeps the chosen one through eleven lenses for gaps, asks five questions for opportunities, and summons one fresh-context Loki-Doubt pass over the draft, whose findings Occam weighs before you see them. Every candidate lands in the triage, which the Oracle sets out in three groups by the ruling it counsels — Now, then Later, then No. Every row takes the ruling it counsels unless you mark it otherwise, naming only the rows you want ruled differently; Now becomes a Mandate, Later a vow, No a denial. It writes the draft spec as soon as Frame settles and records each answer in it as it settles, so nothing you agree lives only in the conversation. It never writes code, tests or a Book.

It is optional. For a small, clear change, describe it to /divineswe:avatar-forge directly; envision earns its place when the work is involved enough that what to build is itself the question. The craft it writes by — grounding, deciding, the forge spec’s shape and the Mandate grammar — lives in skills/oracle-envision/spec-craft.md, which the Avatar also reads before it writes a spec of its own.

It researches at a depth you confirm at Frame, beside the appetite. Quick is a web search with every source cited and anything unverified marked [SUSPECT]. Deep — proposed when a decision rests on outside facts: an architecture choice, an unfamiliar domain, or a library or service to choose — splits the question into angles, sends a researcher per angle and then an independent refuter per angle that sees the claims and their sources but never the researcher’s reasoning, and writes a research report to docs/research/<slug>.md, or to a directory you name, which the spec’s ## Design cites. A claim a refuter marks REFUTED never becomes a design fact. Ask for a report instead of a design and envision runs in report mode: the same deep method, ending at the report — no concepts, no doubt pass, no spec — and committing the report is left to you. The method lives in skills/oracle-envision/research-craft.md.

Trigger keywords (any of these in a message): envision, brainstorm, design, plan a feature, spec, make spec, create spec, define spec, update spec, envision a report, research report for a design, refute the research behind a design.

Output: A finalized forge spec at docs/specs/forge-<slug>.md — the Goal in your words, a ## Design section recording the chosen concept and its sources, with an architecture diagram or a logic flow where one makes the design clearer, Mandates each with an Accept: line (at least one a human check), a ## Human Testing Sequence when the work changes something you use — the steps you take using the built work, the way its user would —, and the deferred ideas under ## Later with their vow ids — committed together with the research report when a deep run wrote one, and with nothing else, ending with Next: /divineswe:avatar-forge <spec>. In report mode the output is the research report alone, left for you to commit.

Typical invocation:

/divineswe:oracle-envision

Then describe what you want to build. The Oracle will frame it with you.

oracle-consult (deprecated)

Deprecated. New work is designed with oracle-envision and built with avatar-forge; this skill remains for Books already in flight.

The deprecated oracle-consult is the Book door. The consultation is an iterative dialogue — Oracle asks clarifying questions one at a time and writes your intent as Mandates, each with an Accept: line, then decomposes the work into prophecies with file targets and test strategies. Loki-Doubt shadows the plan to challenge it, and Preacher returns GO / CAUTION / NO_GO before the Book is finalized. When the ask is really a feature or change to design and forge, the consultation says once that oracle-envision and avatar-forge are the default path, and carries on unless you switch.

Output: A Book file at books/book-NN.md with ordered prophecies and a decision record; the Book is the contract between design-time and run-time for the deprecated prophet-fulfill.

Typical invocation:

/divineswe:oracle-consult book   # deprecated: composes a Book for prophet-fulfill

Then describe the work. Oracle will guide the dialogue.

prophet-fulfill (deprecated)

Deprecated. New work is built with avatar-forge; Prophet still walks a Book already in flight.

Prophet is the immortal conductor of the deprecated cycle. The deprecated prophet-fulfill reads your Book and drives the state machine through eight phases in strict order: INIT → TEST → IMPLEMENT → AUDIT → MERGE → VERIFY → WRAP UP → SEAL.

Plan-time agents fire at design time — Oracle with Preacher and Loki-Doubt inside a Book consultation, and Oracle with one Loki-Doubt pass and one Occam pass over its findings inside oracle-envision — and are not part of the deprecated prophet-fulfill pipeline. Run-time agents are: Thoth (TEST), Monkey (IMPLEMENT), 4eyes + Loki-Exploit (AUDIT), Mason (MERGE), Styx (VERIFY).

Optional Book number arguments: Restrict execution to specific Books.

/divineswe:prophet-fulfill 52      # deprecated
/divineswe:prophet-fulfill 52 53   # deprecated

Without arguments, all incomplete Books are processed.

Output: Completed, merged, and verified prophecies. Sealed Books moved to TABLETS.md.

root-cause-analysis

Systematic diagnosis for bugs and failures. Describe the symptom — Oracle traces it to its origin: the specific code path, state condition, or invariant violation that caused the failure. The investigation produces a diagnosis document; describe the fix to avatar-forge, after oracle-envision when the fix needs designing.

When NOT to use: Do not invoke for new features or architectural changes. Use oracle-envision or avatar-forge for those.

The bug workflow:

/divineswe:root-cause-analysis    # diagnose
/divineswe:avatar-forge           # describe the fix and forge it

When the fix itself needs designing, envision it first with /divineswe:oracle-envision.

Output: A root cause diagnosis. In chain mode, once you accept, it hands the fix to avatar-forge as an ad-hoc ask marked already diagnosed, and each fix Mandate cites the root cause it remedies.

prophet-full-audit

The weighing of what already stands — a whole codebase, a directory, a file, a glob, or the work since a commit. Two attendants walk the scope side by side, each summoned through avatar summon on the model it resolves for that seat: 4eyes, in a single-pass correctness and security review, and Loki-Exploit, hunting injection vectors, privilege escalation and adversarial misuse. Then Occam weighs Loki’s testimony, so that no remedy outgrows the premise that called for it.

No Book is required. No IMPLEMENT phase. No TEST phase. Nothing in the source is touched: the audit’s only marks are each summons’ scroll, brief and decision events, the seats’ own documents and ledger events, and the one report.

When to use: Pre-release quality gate, post-major-merge sanity check, independent security review, or whenever you want a second opinion that actively tries to find what you missed.

Output: one report at .divineswe/audits/<slug>.md, laid down once through audit write — an executive summary, findings at CRITICAL/HIGH/MEDIUM/LOW with each one’s location, Occam’s reading and its remedy, the seats’ perspectives, recommendations, and a Forge Prompt: a ready-to-run /divineswe:avatar-forge ask naming each finding to fix. Hand the Forge Prompt to the Avatar, and it forges the fixes.

prophet-resume (deprecated)

Deprecated with the Book cycle. Forge work resumes by re-running avatar-forge on its spec.

Session recovery. Reads .divineswe/session-state.json and re-enters the Prophet state machine at the exact phase where it stopped. Restores scroll context, compares Book file state against the baseline SHA, and continues orchestration from the last checkpoint.

What is preserved: Phase, prophecy index, task progress, Ledger, Scrolls, agent models snapshot, book baseline SHA.

When to invoke: Any time the deprecated prophet-fulfill exits before WRAP UP. No diagnosis required — the deprecated prophet-resume reads the state and re-enters correctly.

prophet-revelation (deprecated)

Deprecated with the Book cycle. Describe the change to avatar-forge, after oracle-envision when it needs designing.

Auto-chained single-invocation workflow. Routes your prompt:

Feature/change → chains into the deprecated oracle-consult book, waits for the Book, then calls the deprecated prophet-fulfill with the new Book number
Bug/failure → chains into root-cause-analysis, whose chain now hands the fix to avatar-forge rather than to a deprecated Book, so this route no longer produces one

Trade-off: No review step between consultation and execution. The Book is produced and immediately fulfilled. It was meant for experienced DivineSWE users who had calibrated their consultation style and trusted oracle’s decomposition.

Output: Same as the chained manual flow — a completed, merged, verified cycle — with less ceremony.

hermes-setup

The guided setup for the second opinion — the second 4eyes avatar-forge’s closing audit runs after the native one, on a model from another family than the one that implemented the work. It reads the offer from scripts/hermes-setup.mjs: every channel the Hermes carrier carries, the top-rung models offered on each, and where each channel’s prompts go. It says all of that before asking for a choice, then writes second_opinion_model — and for an openai/ endpoint, its endpoint keys — into ~/.divineswe/config.json, keeping every other key there.

When to invoke: Once per machine; divineswe-init offers it at the end of setup. Run it again to change the model or to switch the second opinion off with none. A project overrides the machine-wide value in its own .divineswe/config.json.

Output: The machine-wide config, written. It writes no credential, and it refuses a symlinked or unparseable config file without writing anything. Only claude-code carries the Hermes carrier, so only claude-code runs the second opinion.

divineswe-init

Bootstraps the .divineswe/ directory structure for a project. Creates state files (status.json, session-state.json), the Ledger (ledger.jsonl), and the default config. Injects a DivineSWE context block into .claude/CLAUDE.md and configures the statusline if supported.

Idempotent: Safe to run multiple times. Does not overwrite existing state files.

Run it once to pre-configure a project before your first forge.

update-plugin

Checks for a new DivineSWE version in the marketplace and updates the plugin if one is available. Re-bootstraps (runs divineswe-init) if the version changed, so any new state schema fields or config defaults are applied.

When to invoke: When you see a notification that a new version is available, or periodically to stay current.

atlas-explore

Charts the repository before planning begins. The skill invokes Atlas, the read-only explorer/cartographer agent, across multiple complementary angles (structure, lexicon, dependencies, history) and synthesizes their findings into a single navigable map of the codebase. Atlas never writes — it only surveys and reports — so an atlas-explore invocation is always side-effect-free.

When to invoke: Before an oracle-envision session (or a deprecated oracle-consult one), or any time you need an up-to-date map of an unfamiliar codebase.

Output: A synthesized codebase map (structure, hotspots, and dependencies) feeding plan-time decisions.

scry-project (deprecated)

Deprecated — read what the project owes with vow surface. The deprecated scry facade still runs: a pure read across the Vow register, the deprecated Decrees, the Books, the state files and the git tree, in which an unreadable source reports unknown, never zero.

vow-capture

The fast path onto the Vow register. One line records an obligation the moment it surfaces, with owner and source defaulted for the in-session shower-thought case, so capturing a debt costs less than typing a TODO comment.

When to invoke: The instant someone says “we should remember to…”. A debt noticed and not recorded is a debt forgotten.

Output: A vow id printed by the facade, and a row that appears in the next vow surface reading.

Skill Variants

Skills may have platform-specific variants. The generator produces SKILL.<platform>.md variants for non-canonical platforms (copilot-cli, codex-cli, opencode). Variants adjust self-location logic and env var references for each platform’s runtime model. Canonical SKILL.md files are authoritative for claude-code.