CodeDoc for Confluence — Documentation

Other styles give other structures: Developer Documentation has 9 sections including architecture, API reference and troubleshooting; Management Overview is a non-technical summary with risks and strategic recommendations.

1. Select Anthropic, OpenAI or Google Gemini and paste your API key: Anthropic, OpenAI, Google AI Studio.
2. Click Connect. CodeDoc checks the key and loads the complete list of models from your provider.
3. CodeDoc preselects a balanced, fast model. Pick another one in the searchable Model list if you like; recommendations are at the top.

Fast models (Claude Haiku, GPT mini models, Gemini Flash) write a page for a medium repository for a few cents. Set a spending limit in your provider account.

Add by address works for all token types and is the only option for Bitbucket access tokens, which cannot list repositories. GitLab projects in subgroups (group/subgroup/project) are supported. For your own GitHub Enterprise or GitLab server, connect the server in Settings → Connections first, then add its repositories here. If adding fails, the message names the repository and the connection CodeDoc tried.

Key paths: keys separated by dots (image.tag), list items by index (containers[0].image), all entries with * (services.*.image.tag) and keys that contain dots in quotes (labels["app.kubernetes.io/version"]). Files with several YAML documents, such as Kubernetes manifests, are searched document by document. If a key path doesn't exist, the message lists the keys that do.

Snippet markers: besides codedoc:start / codedoc:end, AsciiDoc tags (tag::install[] … end::install[]) and [START install] … [END install] work too. Snippets are shown as code with syntax highlighting; Markdown files formatted.

Headings one level smaller: for Markdown content you can tick Make headings one level smaller, so a # heading from the file becomes a level-2 heading and fits under the page's own headings.

• JavaScript: package.json (dependencies and the Node engine; not devDependencies).
• Helm and containers: Chart.yaml (chart, app version, chart dependencies), values*.yaml (image repository and tag), docker-compose*.yml / compose*.yaml (service images), Dockerfile (base image of every stage).
• Python: requirements*.txt, pyproject.toml (PEP 621 dependencies, Poetry dependencies, required Python).
• Go, Java, Terraform: go.mod (direct requirements and the Go version), pom.xml (dependencies with a version, properties resolved), versions.tf, providers.tf, main.tf, terraform.tf (required providers and Terraform version).
• Runtimes: .nvmrc, .node-version, .python-version, .tool-versions.

Up to 25 files, top-level folders first; node_modules, vendor, build output and test data are skipped. In large repositories, enter a Folder (for example services/api) to list only what is declared there. The table shows up to 200 dependencies; a note under it says when something was left out or could not be read.

Open Fine-tune (optional) under the explanation. Only these files takes one path per line; CodeDoc then reads exactly these files and does not search. Search depth limits the folders below the chosen folder (only the folder itself, one or two levels of sub-folders). Kinds leaves out what the table should not show, for example everything except container images. Include development dependencies adds the devDependencies of package.json, marked “dev”. Settings you leave unchanged are not stored. If nothing is left to show, the block says why, for example which kinds were not found in which files.

Where to look: Automatically recognises image references (nginx:1.25.3, or a Helm image.tag next to image.repository), dependencies in package.json and name==version lines of requirements.txt. For any other value, choose the source and enter the name: a Docker Hub image, an npm or PyPI package, the releases of a GitHub project (owner/repository), or a product on endoflife.date.

Images are compared with tags of the same form: 1.25.3-alpine with the newest x.y.z-alpine, 20-alpine with the newest NN-alpine. Pre-releases are left out. Images outside Docker Hub (ghcr.io, quay.io, private registries) are not checked yet.

Privacy: only the public name of the image or package (for example nginx or react) goes to the public registry, never the file or the repository, and only for blocks where the check is switched on. For GitHub releases the request also carries your connected github.com token, which only raises GitHub's rate limit. Don't switch it on for internal package names. Answers are kept for 12 hours and the block shows no "checked on" date, so the page changes only when the answer changes. Blocks set to "On every push" with the check also run once a day, because new releases appear without a push.

• Settings saved in the block dialog count as approved: after you publish the page, the block starts syncing on the first view.
• Settings that reach a page another way (copied with a page or template, written through the API, changed later) show Waiting for approval until a Job Operator who can edit the page checks the repository, file, branch, content, update rule and version check, and clicks Approve and sync.
• A block reads only files of its own connected repository: paths and branches with ../ parts (also written as %2e%2e), backslashes or control characters are refused (Settings incomplete, with an explanation).
• Access tokens stay in CodeDoc. The block settings stored on the page contain only the repository, path and options.

• Files up to 1 MB. Snippets, line ranges and whole files are shown up to 20,000 characters; a Markdown section has no such cut, only the 1 MB file limit. Up to 20 key paths or 50 table rows per block; regular expressions up to 200 characters.
• Edits inside a block are replaced on the next update. Write your own text above or below the block.
• Every real update creates a page version, and Confluence notifies the page's watchers as for any other edit.
• Relative links in Markdown point to the file in your Git provider; images are shown as links, because Confluence cannot load images from private repositories. The source line under each block links to the file.
• Without an active license, 3 sync blocks are kept up to date; blocks approved before keep running (Editions).

Next names what is still missing in a step. Edit changes an existing AI page: Go to step jumps to the part you want to change, and Save changes works from any step (the page itself changes with the next run). Copy creates a new AI page with the same settings, including the target space.

1. Repository → Settings → Webhooks → Add webhook.
2. Paste the Webhook URL into "Payload URL" and set Content type to application/json. If you created a webhook secret in CodeDoc, paste it into Secret.
3. Choose Just the push event, then click Add webhook.

1. Project → Settings → Webhooks.
2. Paste the Webhook URL. If you created a webhook secret, paste it into Secret token.
3. Under "Trigger", check Push events, then click Add webhook.

1. Repository → Repository settings → Webhooks → Add webhook.
2. Enter a title (e.g. "CodeDoc") and paste the Webhook URL. If you created a webhook secret, paste it into Secret.
3. Under "Triggers", choose Repository → Push (repo:push) and click Save.

1. Project settings → Service hooks → Create subscription → Web Hooks.
2. Select the trigger event Code pushed.
3. Paste the Webhook URL. If you created a webhook secret, choose Basic authentication: any user name, the secret as the password. Click Finish.

A page is written from at most 50 files (adjustable in Settings) within about 120k tokens. If the scope has more than 1,500 code files, or the repository more than 10,000 files, step 1 says "Too large for one page" with the numbers and next steps, and a run started anyway stops right after reading the file list, before any AI call, so no tokens are used. Narrow the scope with the Folder field (one module or service per AI page), choose From my file list (then only the ticked files are read, whatever the repository size), or keep single facts current with a sync block.

The file list is read once per editor session: going back to the step or to an earlier folder does not ask Git again. Show the files / Choose files opens the list with a filter field; long lists show the first 60 files and Show all. If you change the folder, files outside it drop out of your list.

Expand File filters (advanced) at the bottom of the file panel to fine-tune the list with glob patterns, one per line:

• Leave out: matching files are removed, for example **/*.test.ts, **/__tests__/**, **/migrations/**.
• Keep anyway: rescues files a "Leave out" pattern would remove, for example **/api.test.ts. It is not a whitelist: a file that matches neither field is always kept. To read only one directory, use the Folder field.

**/ needs at least one folder in front of the name, so add *.test.ts as well to leave out test files in the repository root. With both fields empty, the defaults apply: build output, lock files, minified and binary files are filtered, and files that usually hold credentials (such as .env files and private keys) are never read.

In step 2 choose Short (about 5–7 pages), Standard (about 10–15 pages, recommended) or Long (about 20–30 pages). It is an upper limit, not a target: the AI stops when the page is complete. If pages end abruptly, choose a longer option. Options a model cannot write are not offered (for example Long with gpt-4o-mini). Admins can switch on the exact number of output tokens instead: Settings → AI limits → Advanced: exact page length in tokens.

Confluence page titles are unique per space. When you save an AI page, CodeDoc checks the space; on a conflict, step 3 warns you and you can rename the AI page or click Save anyway. The AI page then shows a ⚠ Title conflict badge, and Run asks for confirmation first.

• If someone changed the page after the run, CodeDoc does not undo, so no edits are lost, and names the version to restore from the page history instead. A run that created the page has no earlier version; delete the page in Confluence to remove it.
• A discarded draft can be restored from the space trash. If Confluence refuses to delete the draft, the message says why and what to do, and the draft stays linked so you can try again.
• If you run an AI page again while a draft is waiting, the old draft goes to the trash and a new one is written. Drafts waiting for review also appear under Needs attention on the Overview.
• A run published after review shows Published in the run history; click it to see who published the draft and when.

• The documentation type (one of the 8 styles, or Custom with your own instructions; your text is used as the complete instruction), output language, maximum page length and exclude patterns.
• Only Admins create, edit, copy or delete styles (+ New style, Edit, Copy, Delete); everyone with access can use them.
• A saved style fills in all matching fields at once; you can still adjust any of them. Custom instructions typed directly into one AI page apply to that page only.

• Every run, newest first: time, AI page, result (Published, In review, Discarded, Undone, Failed), AI provider and model, the files read (for example "42 of 128", with how they were chosen), tokens and a link to the page. Click Failed to see why and what to do.
• Next scheduled runs: for AI pages with a schedule, the next run with a countdown and the last run.
• Runs per period: successful runs per day, week or month.
• Clear history (admins) removes all run records; the pages in Confluence stay as they are.

Add connection opens the same form as the setup. Remove deletes the stored token (or the GitHub App's key) and the repositories of that connection; the last connection cannot be removed. Removing the last connection to an own server also withdraws its approval (Atlassian asks a site admin to confirm).

• The Model list contains every text model your provider offers, newest first, with Recommended models at the top (for example the newest Sonnet, Haiku and Opus, GPT and GPT mini, Gemini Pro and Flash). Type to search.
• Refresh list loads models released since you connected the provider. Audio, image, embedding and other non-text models are left out.
• If the selected model is no longer offered, a warning says so; choose another model, otherwise AI pages fail with "model not available".

Show secret and Replace secret are next to it; a secret cannot be removed. After replacing it, update the secret in every Git provider's webhook. The same URL and the same secret work for all Git providers: GitHub and GitHub Enterprise use the webhook's Secret, GitLab the Secret token, Bitbucket the Secret, Azure DevOps Basic authentication with the secret as the password. A push that arrives while an AI page still waits in the queue starts no second run: the waiting run reads the newest commit anyway.

CodeDoc runs on Atlassian's cloud, not inside your network. The server must be reachable from the internet over HTTPS, with a certificate from a public certificate authority. If a firewall protects it, allow Atlassian's outgoing IP addresses (IP addresses and domains for Atlassian cloud products).

The approval is listed in Atlassian Administration → Connected apps → CodeDoc → Data management ("Data storage, sharing, receiving and processing"), where an admin can remove it again. Repositories, file links, sync blocks, AI pages and webhooks then work as on the cloud services.

The App ID, the installation and the private key are stored encrypted in your Confluence site (Forge secret storage) and never shown again. For each read, CodeDoc signs a short-lived request with the key and gets a token from GitHub that is valid for one hour and may only read repository contents and metadata. CodeDoc uses only the installation in the organization that owns the app, checks at every token that the installation still belongs to that organization, and does not use a token with more than read access. No key is shared between customers: every Confluence site has its own app.

To replace the key, generate a new one on GitHub, connect the same app again in CodeDoc (the repositories stay), then delete the old key on GitHub.

The IPv4 ranges listed under "Outgoing connections for Atlassian apps" on Atlassian's page; GitHub's API is reached over IPv4. As of October 2026 they fit into 13 entries: 104.192.136.0/21, 185.166.140.0/22, 131.162.88.0/21, 13.52.5.96/28, 13.200.41.224/28, 13.236.8.224/28, 16.63.53.224/28, 18.136.214.96/28, 18.184.99.224/28, 18.234.32.224/28, 18.246.31.224/28, 43.202.69.96/28, 52.215.192.224/28. Check Atlassian's page for changes.

Atlassian's security requirements for Marketplace apps do not allow apps to collect or store the credentials of Atlassian accounts, and Bitbucket API tokens belong to an Atlassian account. Access tokens belong to a repository, project or workspace instead. A connection made with a personal token before CodeDoc 8 keeps working for now and is marked for replacement: add a connection with an access token, then remove the old one.

Create two dedicated groups, for example codeDocAI-Admins and codeDocAI-JobOperator, to keep CodeDoc access separate and easy to audit. Your own role is shown in the top-right corner of the app.

A user added to a group gets access within about 2 minutes; a user removed from a group loses it within about 30 minutes. There is no way to force a refresh; this keeps the number of Confluence API calls low.

• All data transfers are encrypted with TLS.
• CodeDoc calls only your Git providers (including own servers a site admin approved), for AI pages your AI provider or Atlassian AI, and, for blocks with the version check, the public registries named above.
• API keys and access tokens are stored in Forge encrypted storage within your Confluence site. They are never shown again, never written to pages and never logged.
• CodeDoc creates and updates pages as the CodeDoc app: AI pages in the spaces you choose, sync blocks only inside their own block. Page restrictions that exclude the app stop these updates.
• Usage statistics: CodeDoc counts how often features are used (for example "a sync block was updated"), summed over all sites, and writes technical logs that Atlassian makes available to app vendors. No page, space, repository, file or user information is part of the counts.

Check the file path and the branch in the block settings: paths are case-sensitive and start below the repository root. Copy the path from your Git provider (the part after the branch name). If the file was renamed or moved, update the path. Show preview checks the path before you save.

The message lists the keys that exist at that point of the file, for example Keys under "image": repository, pullPolicy, tag. Key paths are case-sensitive and work only with YAML and JSON files; for other files choose a regular expression, a line range or the whole file.

Settings incomplete lists the missing settings: edit the page, select the block and click the pencil icon. Waiting for approval means new or changed settings are not approved yet: a Job Operator or Administrator who can edit the page has to open it once (details).

The token stored in CodeDoc is invalid or expired. An administrator can replace it under Settings → Connections. The block keeps its last good content until the next successful update.

Your GitHub organization accepts requests only from the addresses on its IP allow list, and CodeDoc reads your repositories from Atlassian's cloud. Two ways, both done by an owner of the organization: connect a GitHub App of your organization (steps), whose IP allow list entries apply to that app only; or add Atlassian's outgoing IP ranges (IPv4 and IPv6) to the organization's list under Settings → Authentication security → IP allow list, and the token stays as it is. Those ranges are shared by Atlassian's whole cloud; GitHub still requires your access token for every request. Already connected with a GitHub App? Then the app's IP allow list or the tick “Enable IP allow list configuration for installed GitHub Apps” is missing.

Push updates need the CodeDoc webhook with its secret in the repository (Webhook setup) and a push to the branch the block follows (its Branch setting, or the repository's default branch). Update now checks the block in the meantime.

Page titles are unique per space. Rename the AI page, or rename or move the other page in Confluence. CodeDoc warns about conflicts when you save an AI page (Confluence target).

• Input token budget too low: raise it in Settings → AI limits.
• Page length too short: choose a longer Maximum page length in step 2.
• Model limits: try a more capable model (for example Haiku → Sonnet, or a GPT mini model → the full GPT model).

AI output is sometimes not valid Confluence markup. CodeDoc repairs broken tags and macros automatically; if that does not work, it falls back to a simplified format (plain text with headings), so the page is always created. Run the AI page again.

Your AI provider is rate-limiting requests. CodeDoc retries up to 3 times with increasing delays (3 s, 8 s, 15 s). If the error persists, wait a few minutes, check the limits of your provider plan, and reduce the number of files the AI reads.

Your AI provider account has no credits left. Top up at Anthropic or OpenAI; for Gemini, check the quota in AI Studio.

The code exceeds the model's maximum input. Let CodeDoc pick the files automatically or tick fewer files, lower the input token budget, add Leave out patterns or set a Folder (File selection), or switch to a model with a larger context window.

The run did not start, so nothing was used. The allowance resets on the first day of the next month (UTC). Meanwhile, choose Your own AI provider in Settings → Connections, or wait. If only a little is left, fewer files or a shorter page may still fit.

Atlassian checks every request like its own AI features. Leave out files with unusual content (AI page → Source → file filters), or use your own AI provider for this page.

Schedules are checked hourly, so a run starts up to about 1 hour after the configured time. If nothing ran after 2 hours or more, check that the update rule is On a schedule, that Next scheduled runs under AI pages → Runs is correct, and that your Git token (and AI key, if you use your own) is still valid.

Choose a longer Maximum page length, include more relevant files (the AI can only document code it receives, so add config files, the README and package.json), or try a more capable model. AI-written pages can contain errors: review them before sharing, and consider Review each run before it is published.

The webhook sends an event CodeDoc does not use. Send push events, or one of the merge events in the events table; others such as fork, issue or comment are ignored.

No AI page or block matches. Check that the repository is added in the Repositories tab with the same address as in the webhook, that the pushed branch is the branch the AI page or block follows, and that the update rule is On every push.

CodeDoc accepts only Bitbucket access tokens (Bitbucket tokens). Create a repository access token with "Repositories: Read" and paste only the token value.

Bitbucket does not let access tokens list repositories (its account-wide repository list was removed in 2026). Add each repository with Add by address; CodeDoc checks there that the token can read it. With Bitbucket Premium, one project or workspace access token covers several repositories.

• May not connect yet: click Allow access to the server; a site admin has to confirm Atlassian's dialog. Right after a new approval, wait a minute and click Connect again.
• Could not reach: Atlassian's cloud cannot open the address. Check HTTPS with a certificate from a public certificate authority, and that a firewall allows Atlassian's outgoing IP addresses.
• Reached, but the token was not accepted: the token is wrong, expired or lacks read permission (GitHub: Contents read; GitLab: read_api and read_repository).

No. A sync block copies exactly what is in the file: no AI calls, no token costs, nothing sent to an AI provider.

Yes. The synced content is stored in the page like normal text, so Confluence search, PDF and Word export and the page history include it.

No. CodeDoc writes to the page only when the content in Git changed, so every page version stands for a real change in the file.

No. Atlassian AI in CodeDoc Advanced has its own monthly allowance, included in the subscription. Nothing is taken from your Rovo or Atlassian AI credits.

Only if the server is reachable from Atlassian's cloud: CodeDoc runs on Forge, not inside your network. Allow Atlassian's outgoing IP addresses or make the server reachable over HTTPS. See Your own Git server.

With Atlassian AI (Advanced), AI pages are included up to the monthly allowance, with no extra charge. With your own key, CodeDoc adds no per-page fee: your AI provider bills the tokens, and the run history shows the token count of every run. A medium repository typically uses 10,000–50,000 input and 2,000–8,000 output tokens. Rough cost per page with your own key: Google Gemini's free tier costs nothing (Google may use free-tier data to improve its products); Claude Haiku or a GPT mini model usually $0.01–0.05; larger models (Claude Sonnet, the full GPT models) $0.10–0.50, depending on repository size.

Not yet: one model is used for all AI pages. Change it any time in Settings → Connections; the next run uses it.

Yes, up to 2 repositories per AI page. The AI receives the code of both and writes one combined page.

All text-based files; there is no language whitelist. How well the AI understands a file type depends on the model. CodeDoc leaves out dependencies and build output, lock files and minified files, binary and media files, and never reads files that usually hold credentials (.env files, private keys and certificates, Terraform state and variable files, credentials files), even if an include pattern names them.

English, German, French, Spanish, Italian, Portuguese, Japanese, Chinese and Korean are in the list; Other language… accepts any language you type (quality depends on the model).

As a regular Confluence page in the space you selected. Edit, move or delete it like any page. Running the AI page again updates the same page, no duplicate. With review switched on, each run is first saved as a [DRAFT] page.

Yes. Tick Review each run before it is published in step 3 of the AI page. The published page changes only when you publish the draft in CodeDoc. Discard keeps the published page unchanged; Undo last run brings back the version before a run. See Review and undo.

Yes. Each update reads the current code from Git, the AI writes the page again and CodeDoc updates the Confluence page in place.

Yes. CodeDoc uses your access token; with read access to a repository, public or private, it works.

Go to Confluence Settings → Apps → Manage apps, find CodeDoc and click Uninstall. From then on, the app can no longer read its configuration and stored data; Atlassian keeps it for 28 days and then deletes it (within 21 days, you can ask us to have it restored to a new installation). Generated Confluence pages remain in your spaces; delete them manually if needed.