@harness-engineering/cli 1.23.0 → 1.23.2

package/dist/agents/skills/claude-code/add-harness-component/SKILL.md

@@ -29,11 +29,11 @@
    - **Component:** Name, which layer it belongs to, what it depends on, what will depend on it
    - **Skill:** Name, purpose, type (rigid or flexible), triggers
 
-3. **Check prerequisites.** The project must already be initialized with harness. If `harness.yaml` does not exist, stop and run initialize-harness-project first.
+3. **Check prerequisites.** The project must already be initialized with harness. If `harness.config.json` does not exist, stop and run initialize-harness-project first.
 
 ### Phase 2: VALIDATE — Check Against Existing Constraints
 
-1. **Read the current configuration.** Load `harness.yaml` and `AGENTS.md` to understand existing layers, constraints, and architecture.
+1. **Read the current configuration.** Load `harness.config.json` and `AGENTS.md` to understand existing layers, constraints, and architecture.
 
 2. **Verify the new component does not conflict:**
    - Does the layer name already exist?
@@ -53,7 +53,7 @@
    - Component: `harness add component <name> --layer <layer-name>`
    - Skill: `harness add skill <name> --type <rigid|flexible>`
 
-2. **Review generated files and configuration changes.** `harness add` modifies `harness.yaml` and may generate template files. Check that the changes look correct.
+2. **Review generated files and configuration changes.** `harness add` modifies `harness.config.json` and may generate template files. Check that the changes look correct.
 
 3. **Create the actual code or content.** `harness add` creates the configuration entry but not necessarily the implementation. Create the directories, files, and initial code as needed.
 
@@ -63,7 +63,7 @@
 
 2. **Update `AGENTS.md`.** Add the new component to the architecture section. Document its purpose, boundaries, and relationships to other components. This keeps agent instructions accurate.
 
-3. **Update layer configuration** if the new component changes dependency relationships. Ensure `harness.yaml` reflects the actual import graph.
+3. **Update layer configuration** if the new component changes dependency relationships. Ensure `harness.config.json` reflects the actual import graph.
 
 4. **For new skills:** Write the `skill.yaml` and `SKILL.md` files following the harness skill format. Use harness-skill-authoring for guidance on writing good skill content.
 
@@ -84,7 +84,7 @@ harness scan [path]
 Skipping this step means subsequent graph queries (impact analysis, dependency health, test advisor) may return stale results.
 
 3. **If validation fails,** fix the issues before committing. Common causes:
-   - New layer not properly registered in `harness.yaml`
+   - New layer not properly registered in `harness.config.json`
    - Component placed in wrong directory for its declared layer
    - Imports from forbidden layers
    - `AGENTS.md` references outdated architecture
@@ -102,7 +102,7 @@ Skipping this step means subsequent graph queries (impact analysis, dependency h
 
 ## Success Criteria
 
-- The new component is properly registered in `harness.yaml`
+- The new component is properly registered in `harness.config.json`
 - The component's files exist in the correct directories for its declared layer
 - `AGENTS.md` is updated to reflect the new component
 - `harness validate` passes after the addition
@@ -110,6 +110,15 @@ Skipping this step means subsequent graph queries (impact analysis, dependency h
 - No circular dependencies were introduced
 - The addition is committed as a single atomic commit
 
+## Rationalizations to Reject
+
+| Rationalization                                                                | Why It Is Wrong                                                                                                                                     |
+| ------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------- |
+| "I will add the component and fix any constraint violations later"             | Phase 2 requires running harness check-deps for a clean baseline BEFORE adding. Phase 5 requires it to pass AFTER adding.                           |
+| "AGENTS.md does not need updating for a small internal component"              | Phase 4 explicitly requires updating AGENTS.md for every new component. Without it, AI agents have no context.                                      |
+| "The new layer imports are obvious, so I do not need to check for circularity" | Phase 2 checks whether new dependency relationships create circular imports. Circular dependencies are invisible until they cause runtime failures. |
+| "I will commit the config change and the code separately for cleaner history"  | The success criteria require a single atomic commit. Splitting creates a window where config references a nonexistent component.                    |
+
 ## Examples
 
 ### Example: Adding a New Layer
@@ -122,7 +131,7 @@ DETERMINE: Adding a layer. Name: infrastructure. Dirs: src/infrastructure/.
   Imported by: business layer (services call external APIs through infrastructure).
 
 VALIDATE:
-  Read harness.yaml — existing layers: presentation, business, data.
+  Read harness.config.json — existing layers: presentation, business, data.
   No conflict with "infrastructure" name.
   Run: harness check-deps — passes (clean baseline).
 
@@ -131,13 +140,13 @@ ADD:
   mkdir -p src/infrastructure
 
 WIRE:
-  Update harness.yaml: allow business → infrastructure imports.
+  Update harness.config.json: allow business → infrastructure imports.
   Update AGENTS.md: document infrastructure layer purpose and boundaries.
 
 VERIFY:
   harness validate    # Pass
   harness check-deps  # Pass
-  git add harness.yaml AGENTS.md src/infrastructure/
+  git add harness.config.json AGENTS.md src/infrastructure/
   git commit -m "feat: add infrastructure layer for external API clients"
 ```
 
@@ -157,7 +166,7 @@ WIRE:
 
 VERIFY:
   harness validate  # Pass
-  git add harness.yaml AGENTS.md
+  git add harness.config.json AGENTS.md
   git commit -m "feat: track API spec for documentation drift detection"
 ```
 
@@ -170,7 +179,7 @@ DETERMINE: Adding a component. Name: notification-service. Layer: business.
   Depends on: data layer (notification repository). Depended on by: presentation layer (routes).
 
 VALIDATE:
-  Read harness.yaml — business layer exists, maps to src/services/.
+  Read harness.config.json — business layer exists, maps to src/services/.
   No existing notification-service directory.
   business → data is an allowed import. Presentation → business is allowed.
   Run: harness check-deps — passes.
@@ -187,6 +196,6 @@ WIRE:
 VERIFY:
   harness validate    # Pass
   harness check-deps  # Pass
-  git add harness.yaml AGENTS.md src/services/notification-service.*
+  git add harness.config.json AGENTS.md src/services/notification-service.*
   git commit -m "feat: add notification service to business layer"
 ```

package/dist/agents/skills/claude-code/align-documentation/SKILL.md

@@ -160,6 +160,15 @@ Skipping this step means subsequent graph queries (impact analysis, dependency h
 - Documentation updates are committed with clear references to the triggering code change
 - No orphaned references to old names, paths, or deleted features remain
 
+## Rationalizations to Reject
+
+| Rationalization                                                               | Why It Is Wrong                                                                                                                                        |
+| ----------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| "The code change is small, so docs probably do not need updating"             | A single function signature change changes the contract for every consumer. Map every code change to its documentation impact regardless of diff size. |
+| "I updated the main reference -- the other mentions can wait"                 | Orphaned references in AGENTS.md and tutorials actively mislead agents and developers. Phase 4 requires verifying no orphaned references remain.       |
+| "I will match the new documentation to what I think the code does"            | Phase 4 requires cross-checking each update against the actual code. Documentation that is wrong is worse than documentation that is missing.          |
+| "The existing doc style is inconsistent, so I will improve it while aligning" | Follow the existing style. Documentation alignment is about accuracy, not style improvement. Style changes should be a separate effort.                |
+
 ## Examples
 
 ### Example: Syncing docs after a module rename

package/dist/agents/skills/claude-code/check-mechanical-constraints/SKILL.md

@@ -116,6 +116,15 @@ For each violation that was not auto-fixed, report:
 - Auto-fixed changes are verified by running the test suite
 - No violations are suppressed without explicit team approval and a documented reason
 
+## Rationalizations to Reject
+
+| Rationalization                                                                       | Why It Is Wrong                                                                                                                         |
+| ------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------- |
+| "This forbidden import is just for one utility function -- I will suppress it inline" | The gate says no suppressing rules without documentation. Undocumented suppressions accumulate and erode the constraint system.         |
+| "The auto-fix looks right, so I do not need to re-run tests"                          | The gate says no auto-fix without test verification. Even import reordering can break code that depends on module initialization order. |
+| "This is just a Tier 2 warning -- it can wait until after merge"                      | Tier 2 violations must be resolved before merge to main. Warnings that accumulate on main become the new baseline.                      |
+| "The linter rule does not make sense for this project, so I will just disable it"     | Propose a config change with justification, do not disable the rule inline. Fix it at the configuration level.                          |
+
 ## Examples
 
 ### Example: Forbidden import detected

package/dist/agents/skills/claude-code/cleanup-dead-code/SKILL.md

@@ -191,6 +191,17 @@ Code behind feature flags or environment checks may appear dead in the default c
 - No dynamically-imported, type-only, or side-effect code was accidentally deleted
 - Each cleanup commit is atomic and has a descriptive message explaining what was removed and why
 
+## Rationalizations to Reject
+
+These are common rationalizations that sound reasonable but lead to incorrect results. When you catch yourself thinking any of these, stop and follow the documented process instead.
+
+| Rationalization                                                                                        | Why It Is Wrong                                                                                                                                                 |
+| ------------------------------------------------------------------------------------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| "This export has zero static imports, so it is definitely dead and safe to remove"                     | Zero static imports does not mean zero consumers. Dynamic imports, type-only imports, side-effect imports, and package entry points all create false positives. |
+| "I removed the dead code and the tests pass, so I do not need to run harness validate and check-deps"  | Both harness validate and harness check-deps must pass after every cleanup. Dead code removal can introduce dependency violations.                              |
+| "The convergence loop found new dead code after my fixes, but it is probably just noise from the tool" | Removing dead code creates more dead code. The convergence loop exists to catch these cascades. If the issue count decreased, loop back.                        |
+| "The entropy report has 60 items but I can clean them all up in one pass to be thorough"               | When the report is very large (>50 items), pick the highest-confidence dead code first. Attempting everything at once risks compound errors.                    |
+
 ## Examples
 
 ### Example: Removing unused utility functions

package/dist/agents/skills/claude-code/detect-doc-drift/SKILL.md

@@ -127,6 +127,15 @@ Group findings by documentation file so that fixes can be applied file-by-file.
 - No documentation references deleted files, functions, or features
 - Drift findings are prioritized and assigned to the appropriate fix cycle
 
+## Rationalizations to Reject
+
+| Rationalization                                                           | Why It Is Wrong                                                                                                                      |
+| ------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------ |
+| "The docs are close enough -- a renamed function is obvious from context" | Renamed references in AGENTS.md cause AI agents to hallucinate about non-existent code. Precision matters.                           |
+| "We only changed internal code, so the docs do not need checking"         | Internal API docs with wrong signatures waste developer debugging time. Changed-behavior-not-reflected drift is High priority.       |
+| "There are too many findings to deal with right now, so skip the scan"    | The escalation protocol exists for this case: focus on Critical and High items, create a tracking issue for the rest.                |
+| "We can rely on code review to catch stale docs"                          | Code reviewers focus on code correctness, not documentation cross-references. harness check-docs catches what humans routinely miss. |
+
 ## Examples
 
 ### Example: Renamed function detected

package/dist/agents/skills/claude-code/enforce-architecture/SKILL.md

@@ -272,21 +272,11 @@ These apply to ALL skills. If you catch yourself doing any of these, STOP.
 
 ## Rationalizations to Reject
 
-### Universal
-
-These reasoning patterns sound plausible but lead to bad outcomes. Reject them.
-
-- **"It's probably fine"** — "Probably" is not evidence. Verify before asserting.
-- **"This is best practice"** — Best practice in what context? Cite the source and
-  confirm it applies to this codebase.
-- **"We can fix it later"** — If it is worth flagging, it is worth documenting now
-  with a concrete follow-up plan.
-
-### Domain-Specific
-
-- **"The violation is minor — just one import"** — One violation sets a precedent. Enforce the constraint or document an explicit exception with rationale.
-- **"It works, so the architecture must be fine"** — Working code with bad architecture is technical debt with compound interest. Correct function does not excuse structural violations.
-- **"This is a legacy module, different rules apply"** — Legacy does not mean exempt. Either the constraint applies or it needs an explicit documented exception.
+| Rationalization                                  | Reality                                                                                                                              |
+| ------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------ |
+| "The violation is minor — just one import"       | One violation sets a precedent. Enforce the constraint or document an explicit exception with rationale.                             |
+| "It works, so the architecture must be fine"     | Working code with bad architecture is technical debt with compound interest. Correct function does not excuse structural violations. |
+| "This is a legacy module, different rules apply" | Legacy does not mean exempt. Either the constraint applies or it needs an explicit documented exception.                             |
 
 ## Escalation
 

package/dist/agents/skills/claude-code/harness-accessibility/SKILL.md

@@ -261,6 +261,16 @@ A11Y-030 [info] Hardcoded color value not from design token set
 - `A11Y-031`: Contrast failure -- fix requires choosing a darker color. Escalate to design tokens or get human input on replacement color.
 - `A11Y-001`: The `alt=""` fix assumes decorative. If the icon conveys meaning, human must write descriptive alt text.
 
+## Rationalizations to Reject
+
+| Rationalization                                                                                                                         | Reality                                                                                                                                                                                                                                                                        |
+| --------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| "The contrast ratio is 4.4:1 — that's essentially 4.5:1 and the difference is imperceptible. I'll mark it passing."                     | WCAG AA requires exactly 4.5:1 for normal text. 4.4:1 fails. There is no rounding or visual perception exception in the standard. Flag the failure with the actual ratio and let the user decide how to remediate.                                                             |
+| "This `<div onClick>` already has good visual styling — adding `role='button'` and a keyboard handler is unnecessary clutter."          | A clickable `<div>` without `role="button"` and `onKeyDown` is inaccessible to keyboard-only users and screen reader users. Visual styling has no bearing on ARIA semantics or keyboard reachability. This is A11Y-012, always flagged.                                        |
+| "The automated fix for this `<img>` alt attribute is obvious — I'll apply it without showing the diff since it's just adding `alt=''`." | Every automated fix must be presented as a before/after diff before being written to disk. This is a hard gate. The correct alt value for non-decorative images requires human judgment, and even `alt=""` makes a semantic claim about decorativeness that must be confirmed. |
+| "I18n is enabled, so I'll skip the `lang` and `dir` attribute checks entirely — harness-i18n will catch them."                          | Deferral to harness-i18n is conditional on `i18n.enabled: true` in config. If i18n is not configured, these checks remain part of this skill's scan. Always read the config before skipping any check category.                                                                |
+| "There are 15 findings in this component — I'll fix the easy ones automatically and leave the rest without reporting them explicitly."  | All findings must be reported, regardless of whether they are auto-fixable. The report is the primary deliverable of the REPORT phase. Selectively reporting only fixable violations hides the full accessibility debt from the team.                                          |
+
 ## Gates
 
 These are hard stops. Violating any gate means the process has broken down.

package/dist/agents/skills/claude-code/harness-api-design/SKILL.md

@@ -332,21 +332,11 @@ These apply to ALL skills. If you catch yourself doing any of these, STOP.
 
 ## Rationalizations to Reject
 
-### Universal
-
-These reasoning patterns sound plausible but lead to bad outcomes. Reject them.
-
-- **"It's probably fine"** — "Probably" is not evidence. Verify before asserting.
-- **"This is best practice"** — Best practice in what context? Cite the source and
-  confirm it applies to this codebase.
-- **"We can fix it later"** — If it is worth flagging, it is worth documenting now
-  with a concrete follow-up plan.
-
-### Domain-Specific
-
-- **"It's an internal API, breaking changes are fine"** — Internal consumers break too. Version the change or coordinate the migration explicitly.
-- **"The field name is obvious enough"** — API field names are a public contract. Follow existing naming conventions and document the semantics.
-- **"Nobody uses that endpoint anyway"** — Verify with access logs or usage data. Assumptions about usage without evidence lead to silent breakages.
+| Rationalization                                   | Reality                                                                                                   |
+| ------------------------------------------------- | --------------------------------------------------------------------------------------------------------- |
+| "It's an internal API, breaking changes are fine" | Internal consumers break too. Version the change or coordinate the migration explicitly.                  |
+| "The field name is obvious enough"                | API field names are a public contract. Follow existing naming conventions and document the semantics.     |
+| "Nobody uses that endpoint anyway"                | Verify with access logs or usage data. Assumptions about usage without evidence lead to silent breakages. |
 
 ## Escalation
 

package/dist/agents/skills/claude-code/harness-architecture-advisor/SKILL.md

@@ -309,21 +309,11 @@ These apply to ALL skills. If you catch yourself doing any of these, STOP.
 
 ## Rationalizations to Reject
 
-### Universal
-
-These reasoning patterns sound plausible but lead to bad outcomes. Reject them.
-
-- **"It's probably fine"** — "Probably" is not evidence. Verify before asserting.
-- **"This is best practice"** — Best practice in what context? Cite the source and
-  confirm it applies to this codebase.
-- **"We can fix it later"** — If it is worth flagging, it is worth documenting now
-  with a concrete follow-up plan.
-
-### Domain-Specific
-
-- **"This will be easier to maintain"** — Easier for whom, and compared to what? Cite the maintenance burden with evidence from the codebase.
-- **"It's the modern approach"** — Modernity is not a design criterion. Fitness for purpose is. State the specific benefit.
-- **"Other teams do it this way"** — Other teams have different constraints. Evaluate the option on this codebase's specific merits.
+| Rationalization                   | Reality                                                                                             |
+| --------------------------------- | --------------------------------------------------------------------------------------------------- |
+| "This will be easier to maintain" | Easier for whom, and compared to what? Cite the maintenance burden with evidence from the codebase. |
+| "It's the modern approach"        | Modernity is not a design criterion. Fitness for purpose is. State the specific benefit.            |
+| "Other teams do it this way"      | Other teams have different constraints. Evaluate the option on this codebase's specific merits.     |
 
 ## Escalation
 

package/dist/agents/skills/claude-code/harness-auth/SKILL.md

@@ -307,21 +307,11 @@ These apply to ALL skills. If you catch yourself doing any of these, STOP.
 
 ## Rationalizations to Reject
 
-### Universal
-
-These reasoning patterns sound plausible but lead to bad outcomes. Reject them.
-
-- **"It's probably fine"** — "Probably" is not evidence. Verify before asserting.
-- **"This is best practice"** — Best practice in what context? Cite the source and
-  confirm it applies to this codebase.
-- **"We can fix it later"** — If it is worth flagging, it is worth documenting now
-  with a concrete follow-up plan.
-
-### Domain-Specific
-
-- **"No one would guess this token format"** — Security by obscurity. Tokens must be cryptographically secure regardless of format predictability.
-- **"This is an internal service, auth is less critical"** — Internal services are lateral movement targets. Authenticate all service boundaries.
-- **"The frontend validates permissions, so the backend doesn't need to"** — Client-side checks are bypassable. Server-side authorization is the only real enforcement.
+| Rationalization                                                      | Reality                                                                                             |
+| -------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------- |
+| "No one would guess this token format"                               | Security by obscurity. Tokens must be cryptographically secure regardless of format predictability. |
+| "This is an internal service, auth is less critical"                 | Internal services are lateral movement targets. Authenticate all service boundaries.                |
+| "The frontend validates permissions, so the backend doesn't need to" | Client-side checks are bypassable. Server-side authorization is the only real enforcement.          |
 
 ## Escalation
 

package/dist/agents/skills/claude-code/harness-autopilot/SKILL.md

@@ -720,6 +720,16 @@ INIT → ASSESS → PLAN → APPROVE_PLAN → EXECUTE → VERIFY → REVIEW →
 - Checkpoint commits fire after every passing checkpoint; recovery commits fire on retry budget exhaustion
 - Rigor level persists across session resume — set once during INIT, never changed mid-session
 
+## Rationalizations to Reject
+
+| Rationalization                                                                                     | Reality                                                                                                                                                                                                                 |
+| --------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| "This phase is low complexity, so I can skip the APPROVE_PLAN gate entirely"                        | Low complexity only means auto-approval when no concern signals fire. If the planner flagged concerns, produced a complexity override, or the task count exceeds 15, the gate pauses regardless of the spec annotation. |
+| "I can write the planning logic inline instead of dispatching to the harness-planner persona agent" | The Iron Law is explicit: autopilot delegates, never reimplements. Using a general-purpose agent or inlining planning logic bypasses the harness methodology.                                                           |
+| "The retry budget is exhausted but I can try one more approach before stopping"                     | The 3-attempt retry budget exists because each failed attempt degrades context and compounds risk. Exceeding the budget without human input turns a recoverable failure into an unrecoverable one.                      |
+| "I will skip the scratchpad since keeping research in conversation is faster"                       | Scratchpad is gated by rigor level. At standard or thorough, bulky research (>500 words) must go to scratchpad to keep agent conversation focused on decisions.                                                         |
+| "The plan auto-approved, so I can skip recording the decision in the decisions array"               | Every plan approval -- auto or manual -- must be recorded with its signal evaluation. The decisions array is the audit trail that explains why a plan was approved.                                                     |
+
 ## Examples
 
 ### Example: 3-Phase Security Scanner

package/dist/agents/skills/claude-code/harness-brainstorming/SKILL.md

@@ -326,6 +326,15 @@ These patterns make requirements testable and unambiguous. Apply them when the o
 - `harness validate` passes after the spec is written
 - If scope was too large, it was decomposed into sub-projects with the human's approval
 
+## Rationalizations to Reject
+
+| Rationalization                                                                    | Reality                                                                                                                                                                                     |
+| ---------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| "I already understand the problem well enough to skip the question phase"          | The Gates section is explicit: you must ask at least one clarifying question before proposing approaches. Skipping questions means making untested assumptions.                             |
+| "There is only one viable approach, so presenting alternatives would be contrived" | The gate requires at least 2 approaches with tradeoffs. A single approach is a recommendation disguised as a decision -- the human has no real choice.                                      |
+| "I will draft the full spec and present it for review all at once to save time"    | Section-dump specs are explicitly forbidden. Presenting section by section with feedback between each catches misunderstandings early.                                                      |
+| "This future capability is low-cost to include now, so we should build it in"      | YAGNI is a gate, not a suggestion. Every capability must trace to a stated requirement. "We might need this later" is the exact rationalization that turns focused specs into bloated ones. |
+
 ## Examples
 
 ### Example: Designing a Notification System

package/dist/agents/skills/claude-code/harness-caching/SKILL.md

@@ -295,6 +295,16 @@ async function getProducts(): Promise<Product[]> {
 }
 ```
 
+## Rationalizations to Reject
+
+| Rationalization                                                                         | Reality                                                                                                                                                                                                                                                                                                                               |
+| --------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| "Cache invalidation is complex — a short TTL means we're always fresh enough"           | Short TTLs without event-based invalidation guarantee staleness windows proportional to the TTL. A 60-second TTL on pricing data means a price change takes up to 60 seconds to propagate. For correctness-sensitive data, TTL alone is not a strategy — it is a staleness budget.                                                    |
+| "Redis is fast — we don't need to worry about the cache stampede, it'll resolve itself" | Cache stampedes do not resolve themselves; they resolve after they have already caused a database overload. With 200 requests per second and a 450ms rebuild time, a stampede generates ~90 concurrent database queries at once. The database can fail before a single cache entry is rebuilt.                                        |
+| "We cache the whole user object so we have everything in one key"                       | Caching composite objects creates broad invalidation requirements. A user profile update invalidates a cache entry that includes unrelated fields like payment methods or notification preferences. Cache keys scoped to specific sub-resources reduce invalidation blast radius.                                                     |
+| "The cache failure is caught and logged — users will just get a slower response"        | Catching the error is not the same as handling it. If cache failure causes a 30-second database query on every request, "slower response" becomes "timeout" becomes "circuit breaker opens" becomes "service unavailable." Failure handling must degrade to database reads within the acceptable latency budget, not unconditionally. |
+| "We namespace by feature name — there's no collision risk"                              | Feature names are not unique across services. Two services that share a Redis instance and both cache `user:123` without a service prefix will corrupt each other's data silently. Namespace prefixes must include the service and the data schema version, not just the feature name.                                                |
+
 ## Gates
 
 - **No unbounded caches.** Every cache (in-memory, Redis, Memcached) must have either a `maxSize`/`maxmemory` limit or a TTL on every key. An unbounded cache will grow until it causes memory exhaustion. WHERE a cache has no eviction policy configured, THEN the skill must halt and require one before proceeding.

package/dist/agents/skills/claude-code/harness-chaos/SKILL.md

@@ -280,6 +280,16 @@ Result: PASSED - System maintained availability throughout.
         Zero 5xx errors observed. No data loss.
 ```
 
+## Rationalizations to Reject
+
+| Rationalization                                                                           | Reality                                                                                                                                                                                                                                                                        |
+| ----------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
+| "Our circuit breakers are already tested in unit tests — we don't need chaos experiments" | Unit tests verify that circuit breaker code executes. Chaos experiments verify that the circuit breaker actually opens under real load, that the fallback produces an acceptable user-facing response, and that monitoring detects the transition. These are different things. |
+| "We can't run chaos experiments in production — it's too risky"                           | Avoiding chaos experiments does not reduce risk — it defers the discovery of failure modes to real incidents. Chaos experiments in staging with defined abort criteria and short durations are lower risk than discovering failure modes at 2am during a real outage.          |
+| "The experiment passed in staging so we know it'll work in production"                    | Staging differences in traffic volume, data distribution, and infrastructure scale can mask failure modes. Staging experiments validate the mechanism; production experiments (with tightly scoped blast radius) validate the system under real conditions.                    |
+| "We injected the fault and the system recovered — the experiment is done"                 | Recovery alone does not validate resilience. The experiment must also confirm: detection time (did monitoring catch it?), recovery time (did it meet the SLA?), and no data loss or corruption. A system that recovers after 10 minutes of data inconsistency has not passed.  |
+| "We have runbooks for these failure modes, so game days aren't necessary"                 | A runbook that has never been executed under pressure is a hypothesis, not a procedure. Game days reveal whether runbooks are complete, whether on-call engineers can execute them accurately under stress, and whether the estimated recovery times are realistic.            |
+
 ## Gates
 
 - **No chaos experiments without abort criteria.** Every experiment must define conditions under which it is immediately terminated. Running an experiment that you cannot stop is reckless, not engineering.

package/dist/agents/skills/claude-code/harness-code-review/SKILL.md

@@ -831,21 +831,11 @@ These apply to ALL skills. If you catch yourself doing any of these, STOP.
 
 ## Rationalizations to Reject
 
-### Universal
-
-These reasoning patterns sound plausible but lead to bad outcomes. Reject them.
-
-- **"It's probably fine"** — "Probably" is not evidence. Verify before asserting.
-- **"This is best practice"** — Best practice in what context? Cite the source and
-  confirm it applies to this codebase.
-- **"We can fix it later"** — If it is worth flagging, it is worth documenting now
-  with a concrete follow-up plan.
-
-### Domain-Specific
-
-- **"The tests pass, so the logic must be correct"** — Tests can be incomplete. Review the logic independently of test results.
-- **"This is how it was done elsewhere in the codebase"** — Existing patterns can be wrong. Evaluate the pattern on its merits, not just its precedent.
-- **"It's just a refactor, low risk"** — Refactors change behavior surfaces. Review them with the same rigor as feature changes.
+| Rationalization                                     | Reality                                                                                     |
+| --------------------------------------------------- | ------------------------------------------------------------------------------------------- |
+| "The tests pass, so the logic must be correct"      | Tests can be incomplete. Review the logic independently of test results.                    |
+| "This is how it was done elsewhere in the codebase" | Existing patterns can be wrong. Evaluate the pattern on its merits, not just its precedent. |
+| "It's just a refactor, low risk"                    | Refactors change behavior surfaces. Review them with the same rigor as feature changes.     |
 
 ## Escalation
 

package/dist/agents/skills/claude-code/harness-codebase-cleanup/SKILL.md

@@ -187,6 +187,17 @@ Remaining findings: 3 (require human action)
      Suggested: Deprecate with @deprecated JSDoc tag, remove in next major version
 ```
 
+## Rationalizations to Reject
+
+These are common rationalizations that sound reasonable but lead to incorrect results. When you catch yourself thinking any of these, stop and follow the documented process instead.
+
+| Rationalization                                                                                               | Why It Is Wrong                                                                                                                                  |
+| ------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------ |
+| "This dead export is in a high-churn file but the removal is clearly safe"                                    | High-churn files have more hidden consumers. Safe findings in the top 10% by churn are downgraded to probably-safe, requiring explicit approval. |
+| "The convergence loop is not reducing findings quickly enough, so I will apply unsafe fixes to make progress" | Unsafe findings are never auto-fixed, regardless of convergence pressure. Each requires human judgment.                                          |
+| "The verification gate failed on a probably-safe fix, but I am confident the fix is correct"                  | When verification fails after a fix batch, the entire batch must be reverted and all findings reclassified as unsafe.                            |
+| "I will skip the hotspot context phase since it adds time and the churn data is just supplementary"           | The hotspot map drives safety classification accuracy. Without it, safe fixes in high-churn areas are not downgraded.                            |
+
 ## Examples
 
 ### Example: Post-Refactoring Cleanup

package/dist/agents/skills/claude-code/harness-compliance/SKILL.md

@@ -288,6 +288,16 @@ Phase 4: REPORT
       7. Add automated HIPAA compliance regression tests to CI pipeline
 ```
 
+## Rationalizations to Reject
+
+| Rationalization                                                                 | Reality                                                                                                                                                                                                                                       |
+| ------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| "We're not in the EU so GDPR doesn't apply to us"                               | GDPR applies to any organization that processes data of EU residents, regardless of where the organization is based. If a single EU user can sign up, GDPR scope must be assessed.                                                            |
+| "Our lawyers will handle the compliance questions — just document what we have" | Legal review and technical implementation are distinct. Lawyers cannot attest that Article 17 deletion cascades to S3 and Segment. The technical implementation must be audited separately.                                                   |
+| "We already did a SOC2 audit last year — this codebase is the same"             | SOC2 Type II assesses controls over time. Adding a new data store, third-party processor, or API endpoint can invalidate previous control attestations. Audits are point-in-time snapshots, not permanent certificates.                       |
+| "The audit isn't for three months — we can fix the gaps before then"            | Gaps found now require implementation, testing, and evidence collection time. Auditors expect evidence of sustained control operation, not freshly deployed fixes. A gap fixed the week before an audit is still a finding.                   |
+| "That field is technically a username, not PII"                                 | Data classification cannot be done by naming convention. A username combined with any other identifying field (email, IP, phone) is PII under GDPR. Classification must be based on the realistic re-identification risk, not the field name. |
+
 ## Gates
 
 - **No compliance report without data classification.** A compliance audit that does not inventory and classify data fields is incomplete. The classification matrix must be produced before controls can be meaningfully assessed. Without knowing what data exists and where, control checks are theoretical.

package/dist/agents/skills/claude-code/harness-containerization/SKILL.md

@@ -269,6 +269,16 @@ Phase 4: VALIDATE
   Result: PASS -- well-configured container setup
 ```
 
+## Rationalizations to Reject
+
+| Rationalization                                                                       | Reality                                                                                                                                                                                                                                                                                                           |
+| ------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| "We use `latest` in production because we always want the most recent image"          | `latest` is a mutable pointer. A deployment at 9am and a rollback at 9pm may pull different image contents with the same tag. Immutable tags (semver or digest) are required for reproducible deployments and reliable rollbacks.                                                                                 |
+| "The container runs as root because it needs to bind to port 80"                      | Binding to a privileged port requires elevated capability, not full root access. The correct solution is to run the container as a non-root user and use `--cap-add NET_BIND_SERVICE`, or to expose a high port and use a load balancer or ingress for port translation.                                          |
+| "We don't set resource limits because we want the container to use whatever it needs" | Containers without memory limits can exhaust node memory, triggering OOM kills of other containers on the same node. Kubernetes uses resource requests for scheduling and limits for safety; omitting limits transfers the risk of one container onto the entire node.                                            |
+| "Our image is 2GB but it only takes a few seconds to pull in our CI"                  | Image size multiplies across every developer pull, every CI run, and every Kubernetes pod startup. A 2GB image that takes 3 seconds to pull in CI with a warm cache takes 90 seconds on a cold node during an autoscaling event at peak traffic.                                                                  |
+| "We don't need liveness and readiness probes — the container exits if it crashes"     | Process exit is a coarse health signal. A process that is running but deadlocked, stuck in an infinite retry loop, or unable to connect to its database will never exit. Kubernetes will continue routing traffic to it. Readiness probes prevent traffic routing to unhealthy containers that are still running. |
+
 ## Gates
 
 - **No `latest` tag in production manifests.** Production Kubernetes manifests or compose files using `latest` image tags are blocking findings. Immutable tags or digests are required.

package/dist/agents/skills/claude-code/harness-data-pipeline/SKILL.md

@@ -259,6 +259,16 @@ Phase 4: DOCUMENT
   Quality Report: FAIL (2 errors requiring immediate attention)
 ```
 
+## Rationalizations to Reject
+
+| Rationalization                                                                                                                   | Reality                                                                                                                                                                                                                                                                                            |
+| --------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| "The pipeline failed halfway through — we'll just re-run it and it'll pick up where it left off."                                 | A non-idempotent pipeline that is re-run from the middle writes duplicate records for the portion that succeeded before failure. The correct fix is to make the pipeline idempotent (MERGE, upsert, or delete-then-insert) so re-runs are always safe, not to assume partial re-runs are harmless. |
+| "The model has no dbt tests yet, but it's only used in one dashboard — low risk."                                                 | Every untested model is a silent data quality failure waiting to reach a stakeholder. Revenue and user-facing models require test coverage regardless of how few consumers they have today. The number of consumers grows; the coverage does not add itself retroactively.                         |
+| "We're still figuring out the schema — we'll add data contracts once the model stabilizes."                                       | Contracts are most valuable during schema evolution, not after it. An unstable schema without a contract lets breaking changes propagate undetected to downstream consumers. Add the contract as the model is defined; update it explicitly as the schema changes. That explicitness is the value. |
+| "Circular dependency detection is handled by the orchestrator — I don't need to check for it during design."                      | Orchestrators detect circular dependencies at runtime, after the DAG has been deployed. Static analysis during design catches them before deployment, before the pipeline fails at 3am, and before engineers have to diagnose a graph cycle under pressure. Detect them early.                     |
+| "The freshness check is too strict — it keeps alerting because the upstream source is occasionally delayed. I'll just remove it." | A freshness check that fires too often has the wrong threshold. Removing it means stale data reaches analysts silently. Adjust the `warn_after` and `error_after` thresholds to match the source's actual SLA, and escalate if the source cannot meet its own SLA.                                 |
+
 ## Gates
 
 - **No approving non-idempotent production pipelines.** If a pipeline writes data without MERGE, upsert, or delete-then-insert patterns, it is flagged as an error. Non-idempotent pipelines cause data duplication on re-runs.

package/dist/agents/skills/claude-code/harness-data-validation/SKILL.md

@@ -328,6 +328,16 @@ await producer.send({ topic: 'order-events', messages: [{ value: JSON.stringify(
 const event = orderPlacedSchema.parse(JSON.parse(message.value));
 ```
 
+## Rationalizations to Reject
+
+| Rationalization                                                                                                                 | Reality                                                                                                                                                                                                                                                                                                                                                                           |
+| ------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| "TypeScript already types the request body — we don't need runtime Zod validation on top of that."                              | TypeScript types are erased at runtime. `req.body as CreateUserInput` compiles fine and accepts any payload at runtime. A missing required field, a string where a number is expected, or an injected extra field bypasses TypeScript entirely. Runtime validation is not redundant with types — it is the only enforcement that exists when the application is actually running. |
+| "We trust this internal service — we don't need to validate its message payloads."                                              | Trust boundaries are not about intent; they are about reliability. Internal services change their schemas, deploy independently, and have bugs. A consumer that accepts payloads without validation silently processes malformed data and produces corrupted downstream records. Validate every message that crosses a process boundary, regardless of who sent it.               |
+| "The validation error message just says 'invalid input' — the developer can look at the schema to understand what failed."      | Developers are not the only consumers of validation errors. Frontend applications display them, monitoring systems alert on them, and support teams diagnose them. A message that says `{"field":"email","expected":"string email","received":"null"}` is resolved in seconds. "Invalid input" creates a support ticket.                                                          |
+| "The two services define their own schemas independently but they've been in sync so far — shared contracts are overkill."      | "In sync so far" describes luck, not process. Independent schema definitions diverge at the next feature sprint when one team changes a field name. Shared contracts in a common package make schema drift a compile-time error instead of a runtime mystery. The divergence between `userId` and `customerId` in the same event is exactly what independent definitions produce. |
+| "Environment variable validation at startup is unnecessary — if a variable is missing, the app will fail when it's first used." | Failing at the first usage of a missing variable produces a cryptic error deep in the call stack, often after the app has been running for minutes and has processed real requests. Failing at startup produces a clear error with the variable name, before any requests are served. Fast failure is always better than deferred failure.                                        |
+
 ## Gates
 
 - **No type assertions on external data.** WHERE `as` is used to cast data from an API response, message payload, request body, or `JSON.parse` result, THEN the skill must flag it as a trust boundary violation. Type assertions bypass runtime validation entirely. The only acceptable pattern is runtime validation followed by type inference.

package/dist/agents/skills/claude-code/harness-database/SKILL.md

@@ -286,21 +286,11 @@ These apply to ALL skills. If you catch yourself doing any of these, STOP.
 
 ## Rationalizations to Reject
 
-### Universal
-
-These reasoning patterns sound plausible but lead to bad outcomes. Reject them.
-
-- **"It's probably fine"** — "Probably" is not evidence. Verify before asserting.
-- **"This is best practice"** — Best practice in what context? Cite the source and
-  confirm it applies to this codebase.
-- **"We can fix it later"** — If it is worth flagging, it is worth documenting now
-  with a concrete follow-up plan.
-
-### Domain-Specific
-
-- **"The table is small, we don't need an index"** — Tables grow. Plan for the steady state, not the current row count.
-- **"The ORM handles this for us"** — ORMs generate SQL that may not match your performance expectations. Review the generated queries for correctness and efficiency.
-- **"We can always add a migration later"** — Schema changes in production have operational cost. Design the schema thoughtfully now rather than migrating repeatedly.
+| Rationalization                              | Reality                                                                                                                          |
+| -------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------- |
+| "The table is small, we don't need an index" | Tables grow. Plan for the steady state, not the current row count.                                                               |
+| "The ORM handles this for us"                | ORMs generate SQL that may not match your performance expectations. Review the generated queries for correctness and efficiency. |
+| "We can always add a migration later"        | Schema changes in production have operational cost. Design the schema thoughtfully now rather than migrating repeatedly.         |
 
 ## Escalation
 

package/dist/agents/skills/claude-code/harness-debugging/SKILL.md

@@ -298,6 +298,15 @@ Update the session status to `resolved`.
 - Debug session file is complete with investigation log, hypotheses, and resolution
 - Learnings were captured for future reference
 
+## Rationalizations to Reject
+
+| Rationalization                                                                   | Reality                                                                                                                                     |
+| --------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------- |
+| "I have a strong hunch about what is wrong, so I will jump straight to fixing it" | Phase 1 INVESTIGATE must be completed before ANY fix code is written. You are guessing, not debugging.                                      |
+| "I changed two things and the bug is gone, so the fix must be correct"            | One variable at a time is a gate. Changing multiple things simultaneously means you do not know which change fixed it.                      |
+| "This is my third attempt but I feel close, so one more try before escalating"    | After 3 failed fix attempts, the gate requires you to question the architecture. The problem is likely not where you think it is.           |
+| "A try-catch that swallows the error prevents the crash, so the bug is fixed"     | Symptom suppression is explicitly listed as a bad fix. Wrapping the failure in a try-catch addresses what the bug did, not why it happened. |
+
 ## Examples
 
 ### Example: API Endpoint Returns 500 Instead of 400