Offline HTML reports¶
Create one shareable file from a recorded native trace:
Open report.html in a browser. It contains its own styles, uses no JavaScript or
CDN assets, and needs no network, server, database, or dashboard dependencies.
Expand the evidence sections to inspect estimator inputs and recorded spans.
Finding and parent links jump to the corresponding timeline row.
The report separates recorded runtime/tokens/cost, inferred or declared finding evidence, and uncalibrated predictions. It shows the longest recorded spans, compatible-selection metadata, estimator provenance, and trace/research metadata. Synthetic traces and incomplete analysis are labeled explicitly.
Include a baseline and quality evidence¶
The positional trace is the candidate when --baseline is supplied:
agentloop analyze candidate.json --baseline baseline.json \
--quality-fixtures fixtures.json --min-quality-score 0.9 \
--html comparison.html --json-out comparison.json
The comparison includes observed deltas, configured gate results, indeterminate
costs, and supplied quality evidence. --json-out includes the same replay object
in the combined analysis payload. Without a baseline, its original
trace/report/diagnosis/optimization shape remains unchanged.
analyze produces an artifact even when comparison gates fail; it is an analysis
command. Use agentloop replay or agentloop ci when failure should block CI.
Quality comparison options require a baseline. Report generation does not run
candidate application code.
Content and sharing¶
Raw event prompts, outputs, and full event metadata are omitted by default. Trace-level metadata and relevant finding evidence remain included, and supplied quality fixtures/results may contain raw task data. Review them before sharing.
To include the complete event content intentionally:
The file then displays a raw-content notice. Values from traces, findings, and quality results are escaped as text; they cannot create markup, event handlers, scripts, or external resource references. A restrictive content security policy blocks scripts and network assets as an additional boundary. Source span IDs are displayed as text and mapped to generated local anchors.
CI artifacts and research supplements¶
After your workflow has produced application traces and installed AgentLoop, add a report step and upload its single output file:
- name: Export analysis report
run: uv run --frozen agentloop analyze runs/candidate.json --baseline runs/baseline.json --html report.html
- name: Upload analysis report
uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
with:
name: agentloop-analysis
path: report.html
if-no-files-found: error
Keep original trace JSON, study manifests, intervention records, and quality fixtures alongside an HTML research supplement when reproducibility matters. The HTML is a readable artifact; it does not establish that a suggested change was implemented or that its predicted savings were measured. See CI evidence and research guidance.
Harness decisions¶
Traces with versioned harness evidence gain a decision section showing policy versions, proposal/application outcomes, dispatch status, timing, parent spans, and retained retry/conflict links. Shadow proposals are not shown as executed interventions, and individual policy benefits remain unverified. Shared hook timing is counted once even when several policies proposed actions.
Raw policy configuration remains omitted unless it was explicitly captured and
the report also enables --include-content. Invalid or unsupported decision
evidence produces a notice; its raw contents are not dumped into the generic
metadata section. Traces without harness evidence render as before.