docs(#184): add review guidance for inconsistent backoff strategies

[fullsend-ai-coder]

Add a "Code pattern consistency" section to AGENTS.md instructing the review agent to flag inconsistent retry/backoff strategies within a single file as a low-severity informational finding. When a PR introduces or modifies a retry loop in a file that already has one, the reviewer should compare strategies (additive, exponential, constant, jitter-based) and ask the author to harmonize or justify the divergence.

Excludes from flagging: files with only one pattern, same-type strategies differing only in parameters, and cross-file divergence.

Note: pre-commit and make lint could not run in the sandbox due to shellcheck-py installation failure (network 403). Manual verification of lint is required.

---

Closes #184

Post-script verification

• Branch is not main/master (agent/184-review-backoff-consistency)
• Secret scan passed (gitleaks — f96750babbed5ada406a9ae04e8068449701d9c7..HEAD)
• Pre-commit hooks passed (authoritative run on runner)
• Tests ran inside sandbox

[fullsend-ai-review]

Review — PR #207

Verdict: approve

This PR adds a well-scoped "Code pattern consistency" section to AGENTS.md, directly implementing the guidance requested in #184. The change is documentation-only (10 lines added, 0 deleted), correctly authorized by the linked issue, and introduces no security concerns.

Dimension summary

Dimension	Result
Correctness	Two minor terminology/phrasing issues (see findings)
Security	Pass — no injection patterns, secrets, or control weakening
Intent & coherence	Authorized, correctly scoped, no scope creep
Style & conventions	Follows existing AGENTS.md formatting patterns
Docs currency	No staleness — docs-only change, no code identifiers affected
Cross-repo contracts	Skipped — no exported interfaces modified

Findings

1. Non-standard terminology: "additive backoff" (low)

The diff uses "additive backoff" as one of the example strategy types. This is not standard terminology in distributed systems literature — the conventional term is linear backoff (delay increases by a fixed increment each retry). The term "additive" comes from the source issue #184, which describes the pattern as delay += initial. While the meaning is clear from context, using the standard term would prevent confusion when reviewers encounter linear backoff patterns in code and need to match them against this guidance.

Suggestion: Replace "additive" with "linear" in both occurrences (the prose example and the strategy list).

2. Ambiguous severity phrasing: "low-severity informational finding" (low)

The guidance instructs reviewers to flag inconsistencies as a "low-severity informational finding." In the review system, low and info are distinct severity levels with different filtering behavior — info-level findings are suppressed by default when the severity threshold is low. The phrase "low-severity informational" could be read as either:

• A low-severity finding that is informational in nature (intended reading), or
• An info-severity finding (which would be filtered out by default, defeating the purpose)

The existing AGENTS.md convention (line 63) uses the simpler pattern: "medium-severity or higher finding." Following this pattern would remove the ambiguity.

Suggestion: Use "low-severity finding" (drop "informational") to match the existing phrasing pattern and avoid confusion with the info severity level.

Notes

• The section placement (between "Forge abstraction" and "ADRs") is consistent with the existing structure — both neighboring sections also contain "When reviewing PRs" guidance.
• The "do not flag when" exclusions are reasonable and well-scoped. The guidance already covers intentional divergence via "ask the author to... explain why they differ."
• The motivating example from issue Review agent: flag inconsistent retry/backoff strategies within a file #184 (dispatchRepoMaintenanceWithRetry in enrollment.go) no longer exists in the codebase, but this does not affect the validity of the guidance itself.

---

Protected paths detected — this PR modifies files under one or more
protected paths. The review agent cannot approve PRs that touch these paths.
A human reviewer must approve this PR.

Protected files in this PR:

• AGENTS.md

---

Labels: PR adds review guidance documentation to AGENTS.md

Previous run

Review — Approve

Scope: 1 file changed (AGENTS.md), 10 additions — adds a "Code pattern consistency" section with review guidance for flagging inconsistent retry/backoff strategies within a single file.

Dimension summary

Dimension	Result
Correctness	✅ No issues — backoff strategy types (additive, exponential, constant, jitter-based) are technically accurate; exclusion criteria are logically consistent and non-overlapping; no contradictions with existing AGENTS.md content
Security	✅ No issues — no secrets, prompt injection patterns, Unicode steganography, or permission changes detected
Intent & coherence	✅ Implements exactly what issue #184 requested with no scope creep; placement between "Forge abstraction" and "ADRs" is architecturally coherent
Style & conventions	✅ Follows established AGENTS.md patterns — ## heading level, **When reviewing PRs:** bolded directive, hyphen bullet lists, description → directive → exceptions structure
Docs currency	✅ No staleness — the new guidance is self-contained and doesn't conflict with or make stale any other documentation
Cross-repo contracts	⏭ Skipped — no exported interfaces or APIs modified

Verdict

Clean documentation addition that faithfully implements the guidance proposed in #184. The new section is well-scoped (per-file only, low-severity, informational), has clear exclusion criteria to avoid false positives, and follows the established conventions of AGENTS.md.

---

Protected paths detected — this PR modifies files under one or more
protected paths. The review agent cannot approve PRs that touch these paths.
A human reviewer must approve this PR.

Protected files in this PR:

• AGENTS.md

---

Labels: PR modifies only AGENTS.md with documentation guidance

[guyoron1]

/fs-review

[fullsend-ai-review]

See the review comment for full details.

[fullsend-ai-review]

[low] terminology

The term "additive backoff" is non-standard. The conventional term in distributed systems literature is "linear backoff" (delay increases by a fixed increment each retry). Using non-standard terminology in reviewer guidance may cause reviewers to miss pattern matches when encountering linear backoff in code.

Suggested fix: Replace "additive" with "linear" in both the prose example and the strategy list.

[fullsend-ai-review]

[low] consistency

The phrase "low-severity informational finding" is ambiguous because low and info are distinct severity levels in the review system. The info level is suppressed by default (threshold=low), so if a reviewer reads this as info severity, the finding would be filtered out — defeating the stated intent. The existing AGENTS.md convention (line 63) uses the simpler "[severity]-severity ... finding" pattern.

Suggested fix: Use "low-severity finding" (drop "informational") to match the existing phrasing pattern and avoid confusion with the info severity level.