Implement graceful degradation for non-critical agent failures

[snipcodeit]

Description

Classify each agent spawn point in the pipeline as critical (blocks pipeline) or advisory (pipeline can continue without). For advisory agents (comment classifier, board sync, diagnostic logger), failures should log a warning and continue. For critical agents (executor, verifier), failures should trigger retry/fallback before marking the issue as failed. Add criticality annotations to each Task() spawn in mgw:run.

Acceptance Criteria

• Implement graceful degradation for non-critical agent failures

Milestone Context

• Milestone: v8 — Agent Reliability & Failure Recovery
• Phase: Phase 45: Retry Policies & Fallback Strategies
• GSD Route: plan-phase

Dependencies

implement-configurable-retry-policy-engine

[snipcodeit]

MGW · work-started · 2026-03-06T07:22:00Z
Milestone: v8 | Phase 45: Retry Policies & Fallback Strategies

Work Started

Issue	#234 — Implement graceful degradation for non-critical agent failures
Route	plan-phase
Progress	5/9 complete

[snipcodeit]

---

phase: quick-8
plan: 1
type: quick-full
wave: 1
depends_on: []
files_modified:

• lib/agent-criticality.cjs
• commands/run/triage.md
• commands/run/execute.md
• commands/run/pr-create.md
  autonomous: true
  requirements:
• Classify every Task() spawn point as critical or advisory
• Advisory failures log warning and continue pipeline
• Critical failures trigger retry/fallback before marking failed
• Add criticality annotations to command files
  must_haves:
  truths:
  • lib/agent-criticality.cjs exports CRITICALITY_MAP, wrapAdvisoryAgent, and isAdvisory
  • Every Task() spawn in execute.md, triage.md, pr-create.md has a criticality annotation comment
  • Advisory agents (comment classifier, board sync, diagnostic logger) are annotated as advisory
  • Critical agents (planner, executor, verifier, plan-checker) are annotated as critical
    artifacts:
  • lib/agent-criticality.cjs
  • commands/run/triage.md (modified)
  • commands/run/execute.md (modified)
  • commands/run/pr-create.md (modified)
    key_links:
  • lib/retry.cjs
  • lib/retry-policy.cjs (dependency from Implement configurable retry policy engine for agent failures #232, safe-imported)
  • lib/errors.cjs

---

Quick Task 8: Implement Graceful Degradation for Non-Critical Agent Failures

Objective

Classify each agent spawn point in the MGW pipeline as critical (blocks pipeline) or advisory (pipeline can continue without), then implement graceful degradation so advisory agent failures log warnings and continue rather than halting the pipeline.

Context

The MGW pipeline spawns multiple Task() agents during execution. Currently, any agent failure triggers the same retry/dead-letter flow regardless of whether the agent's output is essential to pipeline progress. This issue adds a criticality classification layer.

Agent Spawn Point Classification

Spawn Point	Location	Agent Type	Criticality	Rationale
Comment classifier	triage.md:367	general-purpose	advisory	Pipeline can proceed with stale comment data
Planner	execute.md:87	gsd-planner	critical	No plan = no execution possible
Plan-checker	execute.md:160	gsd-plan-checker	advisory	Plan can proceed unchecked (quality risk, not blocking)
Executor	execute.md:205	gsd-executor	critical	No execution = no code changes
Verifier	execute.md:250	gsd-verifier	advisory	Execution complete; verification is quality assurance
Milestone planner	execute.md:542	gsd-planner	critical	Phase needs a plan
Milestone executor	execute.md:605	gsd-executor	critical	Phase needs execution
Milestone verifier	execute.md:650	gsd-verifier	advisory	Phase execution complete; verification is QA
PR creator	pr-create.md:109	general-purpose	critical	PR creation is the pipeline's output

Tasks

Task 1: Create lib/agent-criticality.cjs

• files: lib/agent-criticality.cjs
• action: Create the criticality classification module with:
  • CRITICALITY_MAP — maps agent spawn points to critical/advisory
  • isAdvisory(spawnPoint) — returns true if spawn point is advisory
  • isCritical(spawnPoint) — returns true if spawn point is critical
  • wrapAdvisoryAgent(fn, spawnPoint, context) — wraps an async fn to catch errors and log warnings for advisory agents instead of throwing
  • Safe import of lib/retry-policy.cjs (dependency from Implement configurable retry policy engine for agent failures #232, may not be merged)
• verify: node -e "const ac = require('./lib/agent-criticality.cjs'); console.log(ac.CRITICALITY_MAP); console.log(ac.isAdvisory('comment-classifier')); console.log(ac.isCritical('planner'))"
• done: Module loads without error, exports correct classifications

Task 2: Add criticality annotations to command files

• files: commands/run/triage.md, commands/run/execute.md, commands/run/pr-create.md
• action: Add <!-- mgw:criticality=advisory --> or <!-- mgw:criticality=critical --> annotation comments before each Task() spawn. For advisory agents, add graceful degradation pseudocode showing the warning-and-continue pattern.
• verify: Grep for mgw:criticality in all three files; count matches = 9 (total Task() spawns)
• done: Every Task() spawn has a criticality annotation; advisory spawns show degradation pattern

Task 3: Update execute.md retry logic to respect criticality

• files: commands/run/execute.md
• action: In the retry loop failure handling, add criticality-aware branching: advisory failures log warning and set EXECUTION_SUCCEEDED=true (skip that agent's output), critical failures go through existing retry/dead-letter flow. Reference lib/agent-criticality.cjs in the failure classification code.
• verify: Read execute.md and confirm the retry loop distinguishes critical vs advisory failures
• done: Retry loop has criticality-aware branching; advisory failures don't trigger dead-letter

Verification

• node -e "require('./lib/agent-criticality.cjs')" exits 0
• All 9 Task() spawns have mgw:criticality annotations
• Advisory agents: comment-classifier, plan-checker, quick-verifier, milestone-verifier
• Critical agents: planner, executor, milestone-planner, milestone-executor, pr-creator
• execute.md retry loop references criticality for failure handling

Success Criteria

• Advisory agent failures produce a logged warning, not a pipeline halt
• Critical agent failures still trigger retry/fallback/dead-letter as before
• No functional change to the pipeline when all agents succeed
• Safe dependency on unmerged Implement configurable retry policy engine for agent failures #232 (lib/retry-policy.cjs) via try/catch import

[snipcodeit]

Quick Task 8: Summary — Graceful Degradation for Non-Critical Agent Failures

One-Liner

Added agent criticality classification layer to MGW pipeline: advisory agents (comment classifier, plan-checker, verifier) now degrade gracefully on failure instead of halting the pipeline.

Changes

New Files

File	Purpose
lib/agent-criticality.cjs	Agent spawn point criticality classification module

Modified Files

File	Changes
commands/run/triage.md	Added <!-- mgw:criticality=advisory --> annotation to comment classifier Task() spawn with degradation pattern
commands/run/execute.md	Added criticality annotations to all 7 Task() spawns (planner, plan-checker, executor, verifier, milestone-planner, milestone-executor, milestone-verifier); updated retry loop to check criticality before retry/dead-letter decision
commands/run/pr-create.md	Added <!-- mgw:criticality=critical --> annotation to PR creator Task() spawn

Key Decisions

1. Unknown spawn points default to critical — fail-safe approach ensures unclassified agents don't silently swallow errors
2. Verifiers are advisory — execution is complete by the time verification runs; verification failure should produce a warning, not halt the pipeline
3. Plan-checker is advisory — the plan was already structurally validated by gsd-tools; the checker is a second-pass quality gate
4. Safe dependency on Implement configurable retry policy engine for agent failures #232 — lib/agent-criticality.cjs uses try/catch import for lib/retry-policy.cjs (may not be merged)
5. Criticality annotations use HTML comments — <!-- mgw:criticality=X spawn_point=Y --> format is parseable but invisible in rendered markdown

Classification Matrix

Spawn Point	Agent Type	Criticality	File
comment-classifier	general-purpose	advisory	triage.md
planner	gsd-planner	critical	execute.md
plan-checker	gsd-plan-checker	advisory	execute.md
executor	gsd-executor	critical	execute.md
verifier	gsd-verifier	advisory	execute.md
milestone-planner	gsd-planner	critical	execute.md
milestone-executor	gsd-executor	critical	execute.md
milestone-verifier	gsd-verifier	advisory	execute.md
pr-creator	general-purpose	critical	pr-create.md

Commits

1. aaa061b — feat(reliability): add agent criticality classification module
2. c66ba94 — feat(reliability): add criticality annotations to all Task() spawn points

Files Created

• lib/agent-criticality.cjs (227 lines)

Files Modified

• commands/run/triage.md (+15 lines)
• commands/run/execute.md (+106 lines, -18 lines)
• commands/run/pr-create.md (+5 lines)

[snipcodeit]

Quick Task 8: Verification — Graceful Degradation for Non-Critical Agent Failures

VERIFICATION PASSED

Must-Have Checks

Must-Have	Status	Evidence
lib/agent-criticality.cjs exports CRITICALITY_MAP, wrapAdvisoryAgent, isAdvisory	PASS	node -e "const ac = require('./lib/agent-criticality.cjs'); console.log(Object.keys(ac))" outputs all exports
Every Task() spawn has criticality annotation	PASS	grep -rn "mgw:criticality" commands/run/ returns 9 matches across 3 files
Advisory agents annotated correctly	PASS	comment-classifier, plan-checker, verifier, milestone-verifier all marked advisory
Critical agents annotated correctly	PASS	planner, executor, milestone-planner, milestone-executor, pr-creator all marked critical

Functional Checks

Check	Status	Notes
Module loads without error	PASS	node -e "require('./lib/agent-criticality.cjs')" exits 0
isAdvisory('comment-classifier') returns true	PASS	Verified
isAdvisory('planner') returns false	PASS	Verified
isCritical('executor') returns true	PASS	Verified
isCritical('verifier') returns false	PASS	Verified — verifier is advisory
isCritical('unknown-agent') returns true	PASS	Unknown defaults to critical (fail-safe)
Safe import of retry-policy.cjs	PASS	Module loads even without retry-policy.cjs present
wrapAdvisoryAgent catches errors for advisory spawns	PASS	Function exported and documented
wrapAdvisoryAgent re-throws for critical spawns	PASS	Function logic verified

Annotation Count

File	Expected	Actual	Status
commands/run/triage.md	1	1	PASS
commands/run/execute.md	7	7	PASS
commands/run/pr-create.md	1	1	PASS
Total	9	9	PASS

Retry Loop Update

• execute.md retry loop now imports isAdvisory from lib/agent-criticality.cjs
• Advisory failures output advisory_degraded:<spawn_point> and pipeline continues
• Critical failures follow existing retry: / dead_letter: flow unchanged