chain-studio/assets/docs/approvals.md
flemming-it 28f6fe1a9a feat(approvals,runs): explain approvals in place + one-click hub update (0.81.0)
Approvals: the pending card now shows its full origin — flow, step,
run id (previously dropped at the Dart mapping layer), project, and
requested-at — under an ORIGIN heading, led by a one-line intro strip
that says what the inbox is and what Approve/Reject do. Approve/Reject
buttons carry tooltips; the history dialog gains project + run id.
Fixes the approvals doc drift (title/details/reviewer ->
prompt/show/timeout_seconds). Guard: approvals_origin_test renders the
card via the hermetic fake hub and pins every origin fact.

Runs: the "hub too old" state now leads with an in-place update button
(same `chain update apply` path as the Diagnose page), the Diagnose
deeplink demoted to secondary, with a CLI-absent fallback. Guard: two
new RunsLoadErrorView widget tests.

Signed-off-by: flemming-it <stefan.a.flemming@googlemail.com>
2026-07-26 15:50:41 +02:00

83 lines
2.5 KiB
Markdown

# Approvals
An approval is a deliberate human-in-the-loop checkpoint inside
a flow. When a step needs a human decision before it proceeds,
the flow pauses at that step until a reviewer approves or
rejects it.
## When to use one
Add an approval step before:
- writing to a system of record (cases, contracts, mailings)
- spending money (LLM tokens above a threshold, API quotas)
- destructive actions (delete, archive, send)
- anything the operator's compliance owner asked you to gate
The approval is a recorded event — every approval shows up in
the audit log with the reviewer, decision, and reason. The
`prompt` is the sentence the reviewer sees; `show` selects the
data put in front of them (a prior step's output).
## How a flow declares one
```yaml
steps:
- id: classify
use: text.classify@^1
- id: review
use: system.approval@^0
with:
prompt: "Review the classifier output"
show: "$classify.response"
timeout_seconds: 600
- id: write
use: erp.write@^1
with:
record: "$review.payload"
```
`system.approval` is a built-in capability — no module install
needed. It blocks the flow's run until the matching approval
record reaches a `decision` (approved / rejected) or the
configured `timeout_seconds` elapses.
## How an operator answers one
The Approvals page lists every pending approval. Each card
shows its origin (flow, step, run, project, requested-at), the
`prompt`, and the `show` data. Read them, then decide:
- **Approve** — the flow resumes from this step with the
approval's `payload` as the step output.
- **Reject** — the flow fails at this step with the reviewer's
reason; the audit log records the rejection.
## Timeouts
If no decision arrives within `timeout_seconds`, the engine
returns `ApprovalTimedOut` and the flow fails. Set a timeout
that reflects how long a reviewer can realistically take.
## What gets logged
Every approval write hits the audit log:
- approval requested (with the prompt + show-data snapshot)
- reviewer-decision recorded (with reason on reject)
- payload + signature on accept
The hash chain catches tampering — an entry that was approved
yesterday can't be quietly flipped to rejected today without
breaking the chain.
## Architecture trade-off
Approvals are stored alongside events in the SQLite audit DB.
For higher-assurance deployments, mirror the audit DB to a
WORM sink (Compliance Gap 3, option 2). The approval contract
is the same either way — the WORM sink just makes the audit
trail tamper-evident to outside reviewers, not just to the
hub itself.