deepseek-harness/packages/experimental/forge-github

description, kind

description	kind
Serve the experimental forge registry for GitHub hosts through the GitHub CLI on the host.	package-reference

@deepseek-ai/dsh-experimental-forge-github

English | 中文

Summary

Read GitHub and GitHub Enterprise pull requests for ctx.forge with the GitHub CLI that is already signed in on the host. Each poll batches pull requests by host to read state, changed lines, checks, review decision, mergeability, and unresolved review threads for every due pull request.

Table of Contents

• Use this package
• Understand the implementation
• Further Exploration
• Model Experience
• Known Limitations and Deferred Work
• Dev Note

---

Use this package

Mount the provider after the forge registry and list every GitHub Enterprise host explicitly.

- name: '@deepseek-ai/dsh-experimental-forge'
- name: '@deepseek-ai/dsh-experimental-forge-github'
  config:
    hosts: [github.com, ghe.example.com]

The provider runs gh as a host process, even when Sessions execute commands in another execution world. Sign in once per host with gh auth login --hostname <host>. The child receives the host environment without DSH_* variables and without variables whose names contain KEY, PASSWORD, SECRET, or TOKEN, so GH_TOKEN and GITHUB_TOKEN are not forwarded and gh uses its own credential store. Prompts, update notices, pagers, and colors are disabled. maxOutputBytes, maxReviewThreads, and maxCommits must be positive safe integers; only timeoutMs is bounded by the runtime timer limit. ghCommand, hosts, timeoutMs, maxOutputBytes, maxReviewThreads, and maxCommits are listed in the configuration catalog.

The provider serves a remote URL in https, http, ssh, or git form, or in scp form such as [email protected]:owner/name.git, when its host is configured. HTTP(S) remotes retain custom ports in the host identity; SSH and Git transport ports are excluded. It serves pull request URLs of the form https://<host>/<owner>/<name>/pull/<number> with any trailing path, query, or fragment, and builds that form without the suffix as the canonical URL. A readiness probe runs gh auth status --hostname <host>: exit status 0 is ready, another exit status is unauthenticated with the first diagnostic line, and a gh that cannot start is unavailable. A head lookup returns up to 20 open pull requests whose head is the branch, newest first, excluding pull requests from forks; a lookup of the default branch returns none.

Polling returns the first 100 head-commit check contexts without following pages. An incomplete preview sets checksTruncated and keeps GitHub’s aggregate state; its counts do not describe the full list. fetchChecks(ref, signal) reads complete details only for the requested PR, pinning continuation pages to the first response’s rollup. The combined UTF-8 JSON responses, review thread pages included, must fit maxOutputBytes; excess returns too-large. Failed or non-advancing pages fail only this detail read and leave polling results intact. Runs with the same application, workflow, event, and check name keep only the newest run id; a queued replacement supersedes an older completed run. Runs without a numeric database id are retained by node id because their relative age is unknown. Commit status contexts keep the newest result per context name. Complete lists derive their indicator from these results; unknown context kinds retain GitHub’s aggregate state.

Every status read counts review threads in pages of up to 100 until it has counted maxReviewThreads (default 1,000) of them; each page after the first is one more GraphQL query. A pull request with more threads sets threadsTruncated, and its unresolvedThreads counts only the threads read. A thread page that fails or does not advance fails that pull request's status read and leaves the other pull requests of the batch intact; a partial count is not published.

Unfinished check runs and completed runs with stale, missing, or unrecognized conclusions map to pending. Other completed checks map to success, failure (failure, timed out, startup failure, or action required), cancelled, skipped, or neutral. Cancelled checks keep the rollup non-passing. Check runs keep their start and completion times. Commit status contexts map SUCCESS to success, FAILURE and ERROR to failure, and every other state to pending, and carry no times. A GraphQL NOT_FOUND error becomes a not-found failure. A response without GraphQL data or errors, including an HTTP error JSON body, uses the first stderr diagnostic or the body's message. Authentication diagnostics become unauthenticated; oversized check pages become too-large; rate limits and other failures become unavailable. Head lookups preserve the same diagnostic in their rejection.

A pull request is mergeReady when it is open, not a draft, GitHub reports its merge state as CLEAN, HAS_HOOKS, or UNSTABLE, and the signed-in user has write, maintain, or admin permission. mergeMethods lists the squash, merge-commit, and rebase methods the repository allows, the user's default merge method first. Marking ready runs gh pr ready <number> --repo <host>/<owner>/<name>. Merging calls PUT repos/<owner>/<name>/pulls/<number>/merge through gh api with the method and the expected head SHA, so GitHub refuses once new commits arrived, which is reported as stale. Unlike gh pr merge, it never enables auto-merge, never joins a merge queue, and never deletes the head branch. A refused action reports GitHub's message, and a diagnostic that mentions authentication becomes unauthenticated.

diff reads headRefOid and baseRefOid in one GraphQL query and returns stale when the head is not the requested commit, which must be a full 40- or 64-digit hexadecimal commit. Otherwise it reads repos/<owner>/<name>/compare/<baseRefOid>...<headSha> through gh api --allow-escape-sequences with the diff media type, the patch from the merge base of both commits, so neither end follows a moving branch. Each status snapshot carries baseRefOid as baseSha, and leaves it out when GitHub reports no full commit there; given it, diff starts this comparison while the GraphQL query confirms the head, and a stale head or another base commit cancels the early comparison and waits for it to end before diff returns. For every open and merged pull request measured, this patch is byte-identical to gh pr diff, which takes no commit argument. Unlike the pull request diff, the compare endpoint returns diffs over 300 files, so only output over maxOutputBytes or GitHub's own refusal makes such a diff too-large. --allow-escape-sequences keeps terminal escape sequences that file contents hold, which gh 2.97 and newer otherwise refuse to print; an older gh rejects the option with unknown flag before sending a request and prints such sequences unchanged, so that rejection repeats the read without it. fileText reads repos/<owner>/<name>/contents/<path>?ref=<commit> through gh api with the raw media type and accepts only a full 40- or 64-digit hexadecimal commit. fileUrl returns https://<host>/<owner>/<name>/blob/<commit>/<path>. Both paths percent-encode each segment. commits reads a pull request's commits through GraphQL, 100 per page with the owner, name, number, and cursor as variables, until maxCommits; every page must report the head and commit count of the first one and a cursor no earlier page returned, else the listing is unavailable. compare reads the patch from the merge base of two full commits to the second through the same compare request as diff. A refused read becomes not-found for HTTP 404, unauthenticated for a diagnostic that mentions authentication, too-large for GitHub's HTTP 406 refusal or too_large error, which a diff or file over its size limit draws, and unavailable otherwise, with GitHub's message. A read whose output exceeds maxOutputBytes becomes too-large, and one that outlives timeoutMs becomes unavailable.

Review submission checks the current pull request head, then posts a single inline comment or one completed review through the REST API with an explicit commit_id. Multiline comments carry their side and inclusive start and end lines. Review decisions are Comment, Approve, and Request changes. JSON request bodies travel on stdin. The provider never retries a mutation: a timeout, unexpected response, or server error after dispatch returns an unknown outcome that the caller must reconcile on GitHub. A commit arriving between the check and the write can make the explicitly addressed review outdated; GitHub does not offer an atomic head precondition on these endpoints.

readReview reads review threads with one GraphQL query that names headRefOid and the signed-in viewer's login, and returns stale when the head is not the requested commit. It reads threads in pages of up to 100 until it has read maxReviewThreads of them, each thread page after the first naming the head again, so a push during the read returns stale; it reads every comment of each thread through comment pages, which do not name the head, and the commits and diffHunk of a thread's first comment once. All pages together, measured as JSON, must fit maxOutputBytes: a later page that would pass it is dropped with the threads after it, and the result is marked truncated, as when more than maxReviewThreads threads exist; a first page whose output alone passes it fails the read as too-large. A thread on lines gets a current position only when GitHub reports it not outdated, with a line, and its first comment's commit is the head: a comment written on an earlier commit after a force-push keeps a line that numbers that commit, and GitHub can mark a thread outdated while keeping its line. original comes from originalLine and originalStartLine at the first comment's originalCommit; subjectType: FILE threads have neither. A range whose start is on the other side of the diff keeps its last line. Threads without comments are left out. A pull request or repository GitHub does not have makes the read not-found, as it does the commit listing. REST's position does not tell outdated threads apart and is not read.

The same first page reads what the signed-in viewer may write: each thread's viewerCanReply and, for its current state, viewerCanResolve or viewerCanUnresolve, and each comment's viewerCanUpdate and viewerCanDelete. It also reads the viewer's pending review, the PENDING review the viewer authored, with its comment count; GitHub lists a pending review only to its author, and the read takes up to 100 so the viewer's is found even if another were listed first. The viewer may start a pending review unless the repository is archived or the conversation is locked without ADMIN, MAINTAIN, or WRITE permission.

writeReview first reads the pull request's headRefOid and the viewer's pending review in one GraphQL query and returns stale without writing when the head is not write.headSha. It then sends one mutation with its variables as JSON on stdin: addPullRequestReviewThreadReply, inside the pending review for a pending reply; resolveReviewThread or unresolveReviewThread; updatePullRequestReviewComment or deletePullRequestReviewComment; addPullRequestReviewThread into the pending review, or addPullRequestReview at write.headSha with the comment as its first thread when the viewer has none; and submitPullRequestReview or deletePullRequestReview for the pending review the write names, which the check must still find. A pending reply without a pending review first starts an empty one at the head, so it takes two mutations; when the reply then fails, the empty review stays. A write to a thread or comment also reads, in the same check, the pull request that holds it, and returns failed without a mutation when that is another pull request. GitHub adds every reply of a user with a pending review to that review, even one sent without its id, so a reply at once returns failed without a mutation while the viewer has one; a thread added to a pending review numbers the lines of the current head, whichever commit the review started at. A mutation's payload means GitHub applied it; GraphQL errors whose types all name a refusal (FORBIDDEN, INSUFFICIENT_SCOPES, NOT_FOUND, RATE_LIMITED, UNPROCESSABLE) and HTTP 400, 401, 403, 404, 409, and 422 refusals return failed, as does a gh that cannot start; any other GraphQL error, such as GitHub's timeout message, a gh timeout, a server error, or an answer without the payload returns unknown. No mutation is retried.

---

Understand the implementation

Implementation internals — click to expand

A fetch groups references by host and sends one initial gh api graphql query per host. Continuation queries use the original rollup node id, so a changed pull request head cannot mix commits within one check list. The query inlines validated owner, name, and number values and gives each pull request an alias, so partial GraphQL errors fail only their own aliases. gh exits nonzero on partial errors but still prints the complete response, which the provider parses before it inspects the exit code.

Source	Responsibility
src/index.ts	Configuration validation and provider registration
src/provider.ts	URL mapping, host grouping, and failure classification
src/graphql.ts	Query construction and response validation and mapping
src/compare.ts	Patch between two commits through the compare API, shared by every diff read
src/threads.ts	Review thread read, pagination, and mapping onto head and original positions
src/review-writes.ts	Head check and review thread and pending review mutations
src/gh.ts	Host gh execution with environment scrubbing and limits
src/urls.ts	Remote URL and pull request URL parsing, and validation of inlined names and commits

---

Further Exploration

• Forge registry — provider registration and shared polling.
• Forge subsystem — provider contract and change request types.
• Subprocess environment — the shared environment scrub.

---

Model Experience

None, as the provider reads pull request status for host-side watchers without contributing model context.

KV Cache effect

None. Provider output reaches only host-side watchers.

Known Limitations and Deferred Work

• GitHub CLI on the host — gh must be installed and signed in for each configured host; tokens in environment variables are not forwarded. Hosts missing from hosts are never read, including references retained across configuration changes. A status fetch returns unavailable for those references; a head lookup rejects before starting gh.
• Review thread bound — a status read counts, and a review read reads, at most maxReviewThreads review threads per pull request; beyond it unresolvedThreads is a lower bound that threadsTruncated marks, and a review read is truncated.
• Batch deadline — continuation pages of a pull request with many review threads lengthen its host's polling batch; a batch that outlives ctx.forge's fetchTimeoutMs reports unavailable for every pull request in it, not only for the one with many threads.
• Rewritten base branch — the diff starts from the merge base of the head and the base commit GitHub recorded at the pull request's last update. After the base branch is force-pushed, that merge base can differ from the one GitHub's pull request view uses, so the patch can include or omit commits until the pull request is updated.
• No fork discovery — head lookups exclude pull requests whose head is in another repository.
• Rate limits — every poll costs one GraphQL query per host, plus one query per further page of 100 review threads, and a complete check read one more query per further page of 100 check records; every diff read costs one GraphQL query and one REST request, and a second REST request when the base commit changed after the latest status. All count against the signed-in account’s rate limit. A review read costs one GraphQL query per page of 100 threads and one more per further page of 100 comments in a thread. A review write costs one GraphQL query for the head check and one mutation, or two for a pending reply that starts the pending review.
• No merge queue — a base branch that requires a merge queue refuses a direct merge, and the provider never queues one.

Dev Note

Working context for maintainers — click to expand

None.