Jira and Confluence Research

Jira and Confluence Research

atlcli chat answers ordinary scoped questions without a research graph. atlcli research produces one cited Markdown report from read-only Jira and Confluence evidence. The CLI and browser extension use the same request, scope-provenance, progress-event, workspace, structured-report, and Markdown contracts.

On this page

• Prerequisites
• Use ordinary chat from the CLI
• Run from the CLI
• Run from the browser extension
• Ask from the action palette
• Options
• Output and session workspace
• Shared workflow core
• Security boundaries
• Troubleshooting
• Related topics

Prerequisites

• Configure an Atlassian Cloud profile with access to the selected Jira projects and Confluence spaces.
• Set ANTHROPIC_API_KEY in the CLI process environment. Do not put it on the command line.
• For the extension, open an Atlassian Cloud page and enter the key in the global settings screen. The extension stores it only in browser session storage.

Use ordinary chat from the CLI

Use chat for a direct answer that does not need planning, dynamic subagents, reconciliation, or a deep-research report:

ANTHROPIC_API_KEY=... atlcli chat \
  "Summarize the most important changes in DOCS." \
  --profile work \
  --space DOCS \
  --thinking auto \
  --language en \
  --json

The JSON result contains the retained conversation ID. Continue it without repeating the scope:

ANTHROPIC_API_KEY=... atlcli chat \
  --profile work \
  --session research-session:... \
  "Which change has the largest operational impact?"

Ordinary chat never imports Jira or Confluence defaults from the selected profile. It uses only explicit --project/--space values or an exact, unambiguous project, space, issue, page, or URL named in the question. This prevents a Confluence-only question from silently creating Jira work. Use research when the task needs broader discovery or a planned multi-source investigation.

Chat thinking is independent from Deep Research. --thinking auto enables adaptive thinking and lets the model decide how much reasoning the current turn needs. Use --thinking quick for the lowest-latency chat path or --thinking deep for a more thorough direct answer. All three remain ordinary chat: they create no research graph, subagents, plan review, or reconciliation. The extension exposes the same Automatic, Quick, and Think deeper choices inside Chat; Deep Research remains a separate top-level mode.

Run from the CLI

Minimal example using project and space defaults from the selected profile:

ANTHROPIC_API_KEY=... atlcli research \
  "Which Jira work is explicitly linked to our Confluence documentation?"

Realistic bounded example:

ANTHROPIC_API_KEY=... atlcli research \
  "Which work completed this week is documented, and which relationships are only inferred?" \
  --profile work \
  --project PLATFORM --project DELIVERY \
  --space ENGINEERING,DOCS \
  --from 2026-07-24 \
  --to 2026-07-31 \
  --as-of 2026-07-31T12:00:00+02:00 \
  --timezone Europe/Berlin \
  --max-run-minutes 10 \
  --output /absolute/path/report.md

Repeated and comma-separated project/space keys preserve input order and enter the shared request as locked scope seeds. If a product key is omitted, the CLI uses that product’s configured profile default as an approved seed.

Run from the browser extension

1. Open a Jira or Confluence page and select Research in the side panel.
2. Enter the Anthropic key and the research question.
3. Add explicit Jira project and Confluence space keys when needed.
4. Keep or clear Use detected current context. Manual keys remain locked and are never replaced by the detected context.
5. Confirm the disclosure and select Run research.
6. Follow Live activity for current phases, subagents, read-only tool calls, bounded result counts, stop reasons, durations, and validation decisions.
7. Review the formatted report or raw Markdown, then copy or download it.

Ask from the action palette

Use Ask AI about this page for one bounded ordinary-Chat turn without first opening the side panel:

1. Open the action palette on a Confluence page or Jira issue and select Ask AI about this page.
2. Enter a 3–2,000 character question.
3. Confirm the per-invocation disclosure that the current Atlassian context is sent to the configured provider.
4. Select Ask AI. No provider request occurs before this explicit submit.
5. Review the bounded streamed result, cancel/detach if needed, or continue the same conversation in Research.

The background host re-derives and revalidates the current tab, entity, scope, provider, and credential. If the page changes before execution, the request fails closed. The palette does not persist the prompt, provider response, page body, or credential in its catalog, diagnostics, or action receipt.

This Quick AI path is available only in the browser extension. The Forge action palette has no AI capability in this release.

Options

Both commands accept --profile, --project, --space, --language, --max-run-minutes, --max-cost-usd, --output, and --json. chat also accepts --session for a follow-up turn. Research-only planning, time-window, scope-expansion, and reconciliation flags are rejected by chat.

Option	Type	Default	Constraints
--profile	string	active profile	Existing atlcli profile
--project	string, repeatable	profile project	Jira keys; comma-separated values accepted
--space	string, repeatable	profile space	Confluence keys; comma-separated values accepted
--from, --to	date	none	YYYY-MM-DD; start must not follow end
--as-of	date or timestamp	none	Valid date or ISO 8601 timestamp with timezone
--timezone	string	none	Valid IANA timezone, such as Europe/Berlin
--max-run-minutes	integer	10	From 1 through 10
--output	path	none	Atomically written Markdown
--json	boolean	false	Structured report on stdout; progress remains on stderr
--keep-session	boolean	false	Retains the temporary CLI workspace

Durable-session and planning flags are deliberately rejected until their session-control phases are available.

Output and session workspace

Normal stdout contains only the canonical Markdown. With --output, the file contains exactly the same bytes. The CLI also copies successful reports into a timestamped ~/Documents/atlcli/artefacts/ directory.

Detailed activity streams immediately to CLI stderr and the browser’s bounded Live activity list. It includes stable task/call IDs, subagent roles, read-only capability names, start/completion/failure state, item counts, pagination termination, durations, and host validation decisions. It does not buffer diagnostics until the report is complete.

Each run writes its host-owned state through a virtual workspace. Retained CLI research sessions and ordinary chat conversations use the private SQLite-backed session store under ~/.atlcli/research-sessions/. A chat follow-up restores the same DeepAgentsJS checkpointer and approved scope. Browser chat uses the equivalent IndexedDB-backed workspace so a fresh extension worker can restore the conversation.

Shared workflow core

Chat and Deep Research use one host-neutral agentic workflow contract, but they have different completion objectives. Ordinary Chat completes a conversation answer. Deep Research completes a cited research report. A validated graph is compiled into immutable depth-one profiles for acquisition, analysis, reconciliation, and synthesis; mutable tenant, thread, scope, provider-cache, steering, cancellation, and credential state remains bound to the individual run.

Every delegated task passes the same host-owned admission bridge before a provider call. The bridge validates the registered subagent type and response schema, then applies authorization, any required human approval, budget reservation, cancellation, and the durable journal transition. QuickJS may compose admitted tasks, but it cannot create a role, widen scope, skip a gate, or act as the durable scheduler. Quick Chat constructs neither delegated subagent middleware nor this task bridge.

Compiled-root reuse is disabled by default. A host may enable it only after fresh-versus-reused trajectory tests prove cross-user, thread, scope, cache, steering, and cancellation isolation and measurements show a material latency benefit.

Security boundaries

• Atlassian and Anthropic credentials never enter QuickJS.
• QuickJS has no fetch, shell, Node, Chrome, raw JQL/CQL/GraphQL, or write capability.
• Explicit scope is host-bound. Lower-precedence context cannot replace a manually locked project or space.
• Activity events contain no prompts, source bodies, cursors, credentials, raw model responses, provider errors, or hidden chain-of-thought. Structured reason codes expose reviewable decisions without persisting private internal reasoning.
• Reports cite only fully retrieved, non-empty, non-truncated detail evidence.

Troubleshooting

ANTHROPIC_API_KEY is missing

Set the variable in the process environment that launches atlcli. A CLI flag for the key is intentionally unsupported.

Select at least one Jira project or Confluence space

Pass the missing key explicitly or configure it as a default on the selected profile.

The report says the search is incomplete

The provider or configured item/detail/call budget ended before exhaustive coverage. Treat negative conclusions as bounded to the retrieved evidence and refine the question or scope.

A temporary workspace remains

This is expected with --keep-session. The command prints the retained path to stderr; remove it when it is no longer needed.

Related topics

• CLI Commands
• Atlassian action palette — bounded Quick AI, shortcuts, and host differences
• Authentication
• Configuration
• Environment Variables
• Troubleshooting