1---2name: reliability-observability-gate3description: Use `analysis-agent` to analyze recovery, SLO, capacity, degradation, and observability; `task-agent` to change resilience or telemetry; and `review-agent` to assess evidence. Skip self-review and no-runtime-impact work.4---5
6# reliability-observability-gate
7
8## Role
9
10- **Analysis mode (`analysis-agent`):** Model failure, objective, and recovery behavior.
11- **Task mode (`task-agent`):** Apply accepted resilience or observability controls.
12- **Review mode (`review-agent`):** Judge controls against failure and recovery criteria.
13
14## When To Use
15
16- reliability or recovery risk; observability contract change
17
18## Do Not Use
19
20- no runtime behavior impact
21- self review request
22- unit test or local performance change with no runtime objective
23- logging field-only change with no reliability decision
24- release ordering only
25- data correctness only
26
27## Required Inputs
28
29- acceptance; failure behavior
30- **Analysis mode (`analysis-agent`):** affected runtime path, objective, telemetry, and recovery evidence.
31- **Task mode (`task-agent`):** accepted control decision with failure, capacity, and recovery checks.
32- **Review mode (`review-agent`):** changed runtime path with recovery and telemetry evidence.
33
34## Professional Decision Rules
35
36- Define affected failure modes, user impact, recovery owner, and any decision-relevant operating objective.
37- Require an SLI or SLO only when an owned objective has a decision consequence.
38- Apply timeouts, backpressure, retry budgets, circuit breaking, and degradation only where latency or load risk triggers them.
39- Select only actionable signals, alerts, and runbook links justified by current risk.
40- Validate triggered restart, failover, replay, rollback, and capacity assumptions proportionally.
41
42## High-Value Gotchas
43
44- Retries can worsen overload.
45- An alert without an operator action is noise.
46- Average latency hides tail failure.
47
48## Execution Checklist
49
501. Trace the failure mode through user impact, dependency pressure, telemetry, and recovery ownership.
512. Choose objectives, timeouts, retry budgets, degradation, and alerts only when current risk triggers them.
523. Verify restart, failover, replay, rollback, capacity, and operator-action assumptions where material.
534. **Analysis mode:** select objectives and recovery controls from failure evidence.
545. **Task mode:** apply accepted controls at the affected runtime boundary.
556. **Review mode:** judge restart, failover, capacity, and operator-action evidence.
567. Stop when a material objective or recovery action lacks evidence and ownership.
57
58## Stop / Escalation Conditions
59
60- Block an owned objective when required indicator, target, budget, rollback, or watch evidence lacks an owner.
61- Block required observability lacking current evidence, an owner and action, safe data handling, or bounded cardinality.
62- Block incident closure with unverified cause, missing customer impact, or unowned corrective action.
63- Refuse diagnostic, load, profile, cloud, deploy, rollback, migration, connector, or network-write tools without permission, sandbox, rollback, and redaction evidence.
64
65## Output Contract
66
67- **Analysis mode (`analysis-agent`):** failure model; objective and recovery decisions; telemetry gaps; current evidence and freshness; proof limits; residual risk.
68- **Task mode (`task-agent`):** resilience changes; rollout-watch signals; current post-edit evidence and freshness; proof limits; residual risk.
69- **Review mode (`review-agent`):** reliability verdict; failure findings; current reviewed evidence and freshness; proof limits; unproven recovery behavior; residual risk.
70
71## Targeted References
72
73| Path | Type | Load when | Do not load when | Required by | Required output |
74|---|---|---|---|---|---|
75| [checklist](references/checklist.md) | decision-checklist | A bounded L2 mode needs compact checks for the triggered objective, telemetry, alert, dashboard, recovery, or runbook risk | The root contract is enough or mode-specific closure and targeted proof are required | analysis-agent, task-agent, review-agent | checklist-result, residual-risk |
76| [evidence patterns](references/evidence-patterns.md) | evidence-pattern | Closure depends on command/report artifacts, exit code, dashboard/alert proof, incident evidence, freshness, or proof limits | No reliability claim depends on runtime evidence or the body evidence contract is sufficient | analysis-agent, task-agent, review-agent | evidence-record, proof-limit, residual-risk |
77| [index](references/index.md) | index | competing reliability observability gate references require dependency, conflict, or output-fragment selection | the reliability observability gate root or a task-named reference already resolves selection | analysis-agent, task-agent, review-agent | reference-selection |
78| [reliability output and gates](references/reliability-output-and-gates.md) | targeted | L3-L5 work needs mode-specific closure and targeted gates for a selected objective, alert, resilience, telemetry, capacity/cost, recovery, or incident risk | A compact L1/L2 result is sufficient and no selected risk needs the extended proof contract | analysis-agent, task-agent, review-agent | gate-decision, residual-risk |
79| [solution optimality](references/solution-optimality.md) | targeted | An owned reliability objective, alert strategy, capacity bound, telemetry design, or failure-control choice has a material alternative | No owned objective or operational decision is affected, or current platform policy and evidence already determine the bounded control | analysis-agent, task-agent, review-agent | failure-decision, residual-risk |
machenjie