Spec Compound Refresh

Compound Refresh

Maintain the quality of docs/solutions/ over time. This workflow reviews existing learnings against the current codebase, then refreshes any derived pattern docs that depend on them.

Workflow Contract Summary

• 输入： docs/solutions/、CONCEPTS.md、可选 scope hint，以及当前 source/test/doc evidence。
• 输出： Keep/Update/Consolidate/Replace/Delete/Stale 分类、已应用的知识维护变更和完整 Applied/Recommended 报告。
• 硬出口： source truth、目标 repo、写入范围或语义分类无法确认时不得把猜测写成 current knowledge；headless 只能把歧义标 stale。
• 权威： 当前代码与验证证据优先于旧 learning；本地 mutation、commit 和 landing 分别需要独立授权，mode:headless 只改变交互方式。
• Refresh is current-source anchored: re-read the defining source refs before updating a learning, retain observed revision/freshness and limitations, and mark the item stale when the source cannot be confirmed. Historical cache, session transcript, or provider output is advisory and never a substitute for the current source.
• 消费者： 项目维护者，以及读取 docs/solutions//CONCEPTS.md 的规划、实现、调试和审查 workflow。

Mode Detection

Check whether the invocation arguments supplied by the current host contain the exact token mode:headless. If present, strip only that token (preserving the remainder, quoted paths, and token order as the scope hint) and run in headless mode.

Resolve mutation and landing authority

Record three independent run-local facts before any write or Git action:

mutation_authorization: authorized | missing
commit_authorization: authorized | missing
landing_authorization: authorized | missing

A direct request to refresh or maintain docs/solutions/ authorizes only the bounded local document mutations described by this workflow. Set commit_authorization only when the current user or visible upstream handoff separately requests a commit. Set landing_authorization only when push or PR creation/update is separately explicit. mode:headless, a feature branch, writable permissions, successful edits, or a clean tree do not grant commit, branch creation, push, or PR authority. Without commit authorization, preserve verified edits as uncommitted work; without landing authorization, do not push or open a PR.

Headless mode rules

• Skip all user questions. Never pause for input.
• Process all docs in scope. No scope narrowing questions — if no scope hint was provided, process everything.
• Attempt all safe actions: Keep (no-op), Update (fix references), Consolidate (merge and delete subsumed doc), auto-Delete (unambiguous criteria met), Replace (when evidence is sufficient). If a write succeeds, record it as applied. If a write fails (e.g., permission denied), record the action as recommended in the report and continue — do not stop or ask for permissions.
• Mark as stale when uncertain. If classification is genuinely ambiguous (Update vs Replace vs Consolidate vs Delete) or Replace evidence is insufficient, mark as stale with status: stale, stale_reason, and stale_date in the frontmatter. If even the stale-marking write fails, include it as a recommendation.
• Use conservative confidence. In interactive mode, borderline cases get a user question. In headless mode, borderline cases get marked stale. Err toward stale-marking over incorrect action.
• Always generate a report. The report is the primary deliverable. It has two sections: Applied (actions that were successfully written) and Recommended (actions that could not be written, with full rationale so a human can apply them or run the skill interactively). The report structure is the same regardless of what permissions were granted — the only difference is which section each action lands in.

CONCEPTS.md bootstrap requests

If invoked specifically to create or bootstrap CONCEPTS.md (e.g., "create a CONCEPTS.md", "build the concept map", "set up shared vocabulary"), the intent is ambiguous between two jobs — building the vocabulary file and running a docs/solutions refresh — so disambiguate before proceeding. Use the platform's blocking question tool: AskUserQuestion in Claude Code (call ToolSearch with select:AskUserQuestion first if its schema isn't loaded), request_user_input in Codex. Fall back to numbered options in chat only when no blocking tool exists in the harness or the call errors (e.g., Codex edit modes) — not because a schema load is required. Never silently skip the question. Two options:

1. Create CONCEPTS.md (build the concept map) — seed the repo-wide concept map; skip only the docs/solutions classification phases (Phases 0–4). Read references/concepts-vocabulary.md and follow its Seed goal and Scope of a seed (repo-wide) rules: seed the project's core domain nouns from the declared domain model (schema, core types, primary models, top-level domain docs), each meeting the qualifying bar, the codebase setting the count. Write the preamble (see Phase 4.5), cluster per the organization rules, and run the Discoverability Check so AGENTS.md/CLAUDE.md surface the new file. Then enter the authority-aware Phase 5 closeout; without separate commit authorization, leave the verified bootstrap edits uncommitted.
2. Run a refresh cycle — proceed with the normal refresh flow below; CONCEPTS.md is seeded (if absent) and reconciled as part of Phase 4.5.

In headless mode there is no user to ask: default to the refresh cycle (vocabulary is seeded and reconciled within Phase 4.5 regardless) and note in the report that a standalone repo-wide bootstrap was not run.

Interaction Principles

These principles apply to interactive mode only. In headless mode, skip all user questions and apply the headless mode rules above.

Follow the same interaction style as spec-brainstorm:

• Ask questions one at a time — use the platform's blocking question tool: AskUserQuestion in Claude Code (call ToolSearch with select:AskUserQuestion first if its schema isn't loaded), request_user_input in Codex. Fall back to numbered options in plain text only when no blocking tool exists in the harness or the call errors (e.g., Codex edit modes) — not because a schema load is required. Never silently skip the question
• Prefer multiple choice when natural options exist
• Start with scope and intent, then narrow only when needed
• Do not ask the user to make decisions before you have evidence
• Lead with a recommendation and explain it briefly

The goal is not to force the user through a checklist. The goal is to help them make a good maintenance decision with the smallest amount of friction.

Refresh Order

Refresh in this order:

1. Review the relevant individual learning docs first
2. Note which learnings stayed valid, were updated, were consolidated, were replaced, or were deleted
3. Then review any pattern docs that depend on those learnings

Why this order:

• learning docs are the primary evidence
• pattern docs are derived from one or more learnings
• stale learnings can make a pattern look more valid than it really is

If the user starts by naming a pattern doc, you may begin there to understand the concern, but inspect the supporting learning docs before changing the pattern.

Maintenance Model

For each candidate artifact, classify it into one of five outcomes:

Core Rules

1. Evidence informs judgment. The signals below are inputs, not a mechanical scorecard. Use engineering judgment to decide whether the artifact is still trustworthy.
2. Prefer no-write Keep. Do not update a doc just to leave a review breadcrumb.
3. Match docs to reality, not the reverse. When current code differs from a learning, update the learning to reflect the current code. The skill's job is doc accuracy, not code review — do not ask the user whether code changes were "intentional" or "a regression." If the code changed, the doc should match. If the user thinks the code is wrong, that is a separate concern outside this workflow.
4. Be decisive, minimize questions. When evidence is clear (file renamed, class moved, reference broken), apply the update. In interactive mode, only ask the user when the right action is genuinely ambiguous. In headless mode, mark ambiguous cases as stale instead of asking. The goal is automated maintenance with human oversight on judgment calls, not a question for every finding.
5. Avoid low-value churn. Do not edit a doc just to fix a typo, polish wording, or make cosmetic changes that do not materially improve accuracy or usability.
6. Use Update only for meaningful, evidence-backed drift. Paths, module names, related links, category metadata, code snippets, and clearly stale wording are fair game when fixing them materially improves accuracy.
7. Use Replace only when there is a real replacement. That means either:
   • the current conversation contains a recently solved, verified replacement fix, or
   • the user has provided enough concrete replacement context to document the successor honestly, or
   • the codebase investigation found the current approach and can document it as the successor, or
   • newer docs, pattern docs, PRs, or issues provide strong successor evidence.
8. Delete when the code is gone, and only after checking for inbound links. If the referenced code, controller, or workflow no longer exists in the codebase and no successor can be found, delete the file — don't default to Keep just because the general advice is still "sound." When in doubt between Keep and Delete, ask the user (in interactive mode) or mark as stale (in headless mode). Inbound links inform classification, not cleanup: cleanup is always mechanical, but decorative citations (principle stated inline) allow Delete, while substantive citations (citing doc relies on the cited doc) signal Replace. The auto-delete case is missing code, no matching successor, and citations absent or decorative.
9. Evaluate document-set design, not just accuracy. In addition to checking whether each doc is accurate, evaluate whether it is still the right unit of knowledge. If two or more docs overlap heavily, determine whether they should remain separate, be cross-scoped more clearly, or be consolidated into one canonical document. Redundant docs are dangerous because they drift silently — two docs saying the same thing will eventually say different things.
10. Delete, don't archive. There is no _archived/ directory. When a doc is no longer useful, delete it. Git history preserves every deleted file — that is the archive. A dedicated archive directory creates problems: archived docs accumulate, pollute search results, and nobody reads them. If someone needs a deleted doc, git log --diff-filter=D -- docs/solutions/ will find it.

Scope Selection

Start by discovering learnings and pattern docs under docs/solutions/.

Exclude:

• README.md
• docs/solutions/_archived/ (legacy — if this directory exists, flag it for cleanup in the report)

Find all .md files under docs/solutions/, excluding README.md files and anything under _archived/. If an _archived/ directory exists, note it in the report as a legacy artifact that should be cleaned up (files either restored or deleted).

If invocation arguments remain, use them to narrow scope before proceeding. Try these matching strategies in order, stopping at the first that produces results:

1. Directory match — check if the argument matches a subdirectory name under docs/solutions/ (e.g., performance-issues, database-issues)
2. Frontmatter match — search module, component, or tags fields in learning frontmatter for the argument
3. Filename match — match against filenames (partial matches are fine)
4. Content search — search file contents for the argument as a keyword (useful for feature names or feature areas)

If no matches are found, report that and ask the user to clarify. In headless mode, when a scope hint was provided but matched nothing, report the miss in the summary and exit without widening to all docs — do not silently fall back to processing everything. (The "process everything" rule from Headless mode rules applies only when no scope hint was provided.)

If no candidate docs are found, report:

No candidate docs found in docs/solutions/.
Run `spec-compound` after solving problems to start building your knowledge base.

Phase 0: Assess and Route

Before asking the user to classify anything:

1. Discover candidate artifacts
2. Estimate scope
3. Choose the lightest interaction path that fits

Route by Scope

Broad Scope Triage

When scope is broad (9+ candidate docs), do a lightweight triage before deep investigation:

1. Inventory — read frontmatter of all candidate docs, group by module/component/category
2. Impact clustering — identify areas with the densest clusters of learnings + pattern docs. A cluster of 5 learnings and 2 patterns covering the same module is higher-impact than 5 isolated single-doc areas, because staleness in one doc is likely to affect the others.
3. Spot-check drift — for each cluster, check whether the primary referenced files still exist. Missing references in a high-impact cluster = strongest signal for where to start.
4. Recommend a starting area — present the highest-impact cluster with a brief rationale and ask the user to confirm or redirect. In headless mode, skip the question and process all clusters in impact order.

Example:

Found 24 learnings across 5 areas.

The auth module has 5 learnings and 2 pattern docs that cross-reference
each other — and 3 of those reference files that no longer exist.
I'd start there.

1. Start with auth (recommended)
2. Pick a different area
3. Review everything

Do not ask action-selection questions yet. First gather evidence.

Phase 1: Investigate Candidate Learnings

For each learning in scope, read it, cross-reference its claims against the current codebase, and form a recommendation.

A learning has several dimensions that can independently go stale. Surface-level checks catch the obvious drift, but staleness often hides deeper:

• References — do the file paths, class names, and modules it mentions still exist or have they moved?
• Recommended solution — does the fix still match how the code actually works today? A renamed file with a completely different implementation pattern is not just a path update.
• Code examples — if the learning includes code snippets, do they still reflect the current implementation?
• Related docs — are cross-referenced learnings and patterns still present and consistent?
• Auto memory (Claude Code only) — does the injected auto-memory block in your system prompt contain entries in the same problem domain? Scan that block directly. If the block is absent, skip this dimension. A memory note describing a different approach than what the learning recommends is a supplementary drift signal.
• Overlap — while investigating, note when another doc in scope covers the same problem domain, references the same files, or recommends a similar solution. For each overlap, record: the two file paths, which dimensions overlap (problem, solution, root cause, files, prevention), and which doc appears broader or more current. These signals feed Phase 1.75 (Document-Set Analysis).
• Vocabulary — note domain terms the learning cites (entities, named processes, status concepts with project-specific meaning). For each term: does it appear in CONCEPTS.md? If yes, does the definition still match how the code uses the term? If no, flag the term for Phase 4.5 to add or bootstrap. Do not edit CONCEPTS.md during investigation — just collect the signal centrally.

Match investigation depth to the learning's specificity — a learning referencing exact file paths and code snippets needs more verification than one describing a general principle.

Drift Classification: Update vs Replace

The critical distinction is whether the drift is cosmetic (references moved but the solution is the same) or substantive (the solution itself changed):

• Update territory — file paths moved, classes renamed, links broke, metadata drifted, but the core recommended approach is still how the code works. spec-compound-refresh fixes these directly.
• Replace territory — the recommended solution conflicts with current code, the architectural approach changed, or the pattern is no longer the preferred way. This means a new learning needs to be written. An authorized replacement subagent may draft the successor following spec-compound's document format, using the investigation evidence already gathered, but it never writes the tracked successor. Without authorized dispatch, the orchestrator composes the replacement inline or serially. In every path the orchestrator is the sole tracked-file writer, and the successor must pass the same source_refs / invalidation_condition promotion gate as a new spec-compound learning.

The boundary: if you find yourself rewriting the solution section or changing what the learning recommends, stop — that is Replace, not Update.

Memory-sourced drift signals are supplementary, not primary. A memory note describing a different approach does not alone justify Replace or Delete. Use memory signals to:

• Corroborate codebase-sourced drift (strengthens the case for Replace)
• Prompt deeper investigation when codebase evidence is borderline
• Add context to the evidence report ("(auto memory [claude]) notes suggest approach X may have changed since this learning was written")

In headless mode, memory-only drift (no codebase corroboration) should result in stale-marking, not action.

Judgment Guidelines

Three guidelines that are easy to get wrong:

1. Contradiction = strong Replace signal. If the learning's recommendation conflicts with current code patterns or a recently verified fix, that is not a minor drift — the learning is actively misleading. Classify as Replace.
2. Age alone is not a stale signal. A 2-year-old learning that still matches current code is fine. Only use age as a prompt to inspect more carefully.
3. Check for successors before deleting. Before recommending Replace or Delete, look for newer learnings, pattern docs, PRs, or issues covering the same problem space. If successor evidence exists, prefer Replace over Delete so readers are directed to the newer guidance.

Phase 1.5: Investigate Pattern Docs

After reviewing the underlying learning docs, investigate any relevant pattern docs under docs/solutions/patterns/.

Pattern docs are high-leverage — a stale pattern is more dangerous than a stale individual learning because future work may treat it as broadly applicable guidance. Evaluate whether the generalized rule still holds given the refreshed state of the learnings it depends on.

A pattern doc with no clear supporting learnings is a stale signal — investigate carefully before keeping it unchanged.

Phase 1.75: Document-Set Analysis

After investigating individual docs, step back and evaluate the document set as a whole. The goal is to catch problems that only become visible when comparing docs to each other — not just to reality.

Overlap Detection

For docs that share the same module, component, tags, or problem domain, compare them across these dimensions:

• Problem statement — do they describe the same underlying problem?
• Solution shape — do they recommend the same approach, even if worded differently?
• Referenced files — do they point to the same code paths?
• Prevention rules — do they repeat the same prevention bullets?
• Root cause — do they identify the same root cause?

High overlap across 3+ dimensions is a strong Consolidate signal. The question to ask: "Would a future maintainer need to read both docs to get the current truth, or is one mostly repeating the other?"

Supersession Signals

Detect "older narrow precursor, newer canonical doc" patterns:

• A newer doc covers the same files, same workflow, and broader runtime behavior than an older doc
• An older doc describes a specific incident that a newer doc generalizes into a pattern
• Two docs recommend the same fix but the newer one has better context, examples, or scope

When a newer doc clearly subsumes an older one, the older doc is a consolidation candidate — its unique content (if any) should be merged into the newer doc, and the older doc should be deleted.

Canonical Doc Identification

For each topic cluster (docs sharing a problem domain), identify which doc is the canonical source of truth:

• Usually the most recent, broadest, most accurate doc in the cluster
• The one a maintainer should find first when searching for this topic
• The one that other docs should point to, not duplicate

All other docs in the cluster are either:

• Distinct — they cover a meaningfully different sub-problem and have independent retrieval value. Keep them separate.
• Subsumed — their unique content fits as a section in the canonical doc. Consolidate.
• Redundant — they add nothing the canonical doc doesn't already say. Delete.

Retrieval-Value Test

Before recommending that two docs stay separate, apply this test: "If a maintainer searched for this topic six months from now, would having these as separate docs improve discoverability, or just create drift risk?"

Separate docs earn their keep only when:

• They cover genuinely different sub-problems that someone might search for independently
• They target different audiences or contexts (e.g., one is about debugging, another about prevention)
• Merging them would create an unwieldy doc that is harder to navigate than two focused ones

If none of these apply, prefer consolidation. Two docs covering the same ground will eventually drift apart and contradict each other — that is worse than a slightly longer single doc.

Cross-Doc Conflict Check

Look for outright contradictions between docs in scope:

• Doc A says "always use approach X" while Doc B says "avoid approach X"
• Doc A references a file path that Doc B says was deprecated
• Doc A and Doc B describe different root causes for what appears to be the same problem

Contradictions between docs are more urgent than individual staleness — they actively confuse readers. Flag these for immediate resolution, either through Consolidate (if one is right and the other is a stale version of the same truth) or through targeted Update/Replace.

Subagent Strategy

Before any investigation or replacement dispatch, record:

worker_dispatch_authorization: authorized | missing
capability_probe: not_applicable | attempted | unavailable
worker_dispatch_capability: available | missing | unknown
worker_context_isolation: isolated | inherited | unknown
worker_model_override: supported | unsupported | unknown
worker_bounded_parallelism: supported | unsupported | unknown

workflow invocation does not authorize dispatch。只有当前用户或可见 upstream handoff 明确请求 subagent、delegated work、persona 或 parallel work 时才可派发；headless mode、scope size、permission settings 或本 Skill 被调用都不是授权。缺授权时不得探测 tool schema，固定为 capability_probe: not_applicable + worker_dispatch_capability: unknown，使用 main-thread/serial fallback 并记录 dispatch_authorization_missing。只有授权后才把 current-session registry/schema 作为 provider_untrusted evidence 检查：确认缺失时记录 subagent_capability_missing；surface 不可用、schema 不完整或候选不唯一时记录 worker_capability_unproven，均使用同一 fallback。隔离、模型覆盖和有界并发只取 live facts；required isolation 未满足时保持依赖 gate 打开，model unknown 时继承，parallelism unknown 时串行。记录 worker_dispatch_outcome。Inline fallback 不得声称 independent investigation coverage。

When dispatch is authorized and available, use subagents for context isolation when investigating multiple artifacts — not just because the task sounds complex. Otherwise choose the matching main-thread or serial approach with the same evidence contract:

When spawning an authorized subagent, omit the mode parameter so the user's configured permission settings apply; those settings are execution conditions, not authorization. Include this instruction in its task prompt:

Use dedicated file search and read tools (Glob, Grep, Read) for all investigation. Do NOT use shell commands (ls, find, cat, grep, test, bash) for file operations. This avoids permission prompts and is more reliable.

Also scan the "user's auto-memory" block injected into your system prompt (Claude Code only). Check for notes related to the learning's problem domain. Report any memory-sourced drift signals separately from codebase-sourced evidence, tagged with "(auto memory [claude])" in the evidence section. If the block is not present in your context, skip this check.

There are two subagent roles:

1. Investigation subagents — read-only. They must not edit files, create successors, or delete anything. Each returns: file path, evidence, recommended action, confidence, and open questions. These can run in parallel when artifacts are independent.
2. Replacement subagents — draft one candidate successor and return its content or a run-local scratch reference. They never write tracked files, stage, commit, or delete. These run one at a time, sequentially; the orchestrator validates the draft and performs the tracked successor write, deletion, and metadata updates.

The orchestrator merges investigation results, detects contradictions, coordinates replacement subagents, and performs all deletions/metadata edits centrally. In interactive mode, it asks the user questions on ambiguous cases. In headless mode, it marks ambiguous cases as stale instead. If two artifacts overlap or discuss the same root issue, investigate them together rather than parallelizing.

Phase 2: Classify the Right Maintenance Action

After gathering evidence, assign one recommended action.

Keep

The learning is still accurate and useful. Do not edit the file — report that it was reviewed and remains trustworthy. Only add last_refreshed if you are already making a meaningful update for another reason.

Update

The core solution is still valid but references have drifted (paths, class names, links, code snippets, metadata). Apply the fixes directly.

Consolidate

Choose Consolidate when Phase 1.75 identified docs that overlap heavily but are both materially correct. This is different from Update (which fixes drift in a single doc) and Replace (which rewrites misleading guidance). Consolidate handles the "both right, one subsumes the other" case.

When to consolidate:

• Two docs describe the same problem and recommend the same (or compatible) solution
• One doc is a narrow precursor and a newer doc covers the same ground more broadly
• The unique content from the subsumed doc can fit as a section or addendum in the canonical doc
• Keeping both creates drift risk without meaningful retrieval benefit

When NOT to consolidate (apply the Retrieval-Value Test from Phase 1.75):

• The docs cover genuinely different sub-problems that someone would search for independently
• Merging would create an unwieldy doc that harms navigation more than drift risk harms accuracy

Consolidate vs Delete: If the subsumed doc has unique content worth preserving (edge cases, alternative approaches, extra prevention rules), use Consolidate to merge that content first. If the subsumed doc adds nothing the canonical doc doesn't already say, skip straight to Delete.

The Consolidate action is: merge unique content from the subsumed doc into the canonical doc, then delete the subsumed doc. Not archive — delete. Git history preserves it.

Replace

Choose Replace when the learning's core guidance is now misleading — the recommended fix changed materially, the root cause or architecture shifted, or the preferred pattern is different.

The user may have invoked the refresh months after the original learning was written. Do not ask them for replacement context they are unlikely to have — use agent intelligence to investigate the codebase and synthesize the replacement.

Evidence assessment:

By the time you identify a Replace candidate, Phase 1 investigation has already gathered significant evidence: the old learning's claims, what the current code actually does, and where the drift occurred. Assess whether this evidence is sufficient to write a trustworthy replacement:

• Sufficient evidence — you understand both what the old learning recommended AND what the current approach is. The investigation found the current code patterns, the new file locations, the changed architecture. → Proceed to write the replacement (see Phase 4 Replace Flow).
• Insufficient evidence — the drift is so fundamental that you cannot confidently document the current approach. The entire subsystem was replaced, or the new architecture is too complex to understand from a file scan alone. → Mark as stale in place:
  • Add status: stale, stale_reason: [what you found], stale_date: YYYY-MM-DD to the frontmatter
  • Report what evidence you found and what is missing
  • Recommend the user run spec-compound after their next encounter with that area, when they have fresh problem-solving context

Delete

Choose Delete when:

• The code or workflow no longer exists and the problem domain is gone
• The learning is obsolete and has no modern replacement worth documenting
• The learning is fully redundant with another doc (use Consolidate if there is unique content to merge first)
• There is no meaningful successor evidence suggesting it should be replaced instead

Action: delete the file. No archival directory, no metadata — just delete it. Git history preserves every deleted file if recovery is ever needed.

Before deleting: check if the problem domain is still active

When a learning's referenced files are gone, that is strong evidence — but only that the implementation is gone. Before deleting, reason about whether the problem the learning solves is still a concern in the codebase:

• A learning about session token storage where auth_token.rb is gone — does the application still handle session tokens? If so, the concept persists under a new implementation. That is Replace, not Delete.
• A learning about a deprecated API endpoint where the entire feature was removed — the problem domain is gone. That is Delete.

Do not search mechanically for keywords from the old learning. Instead, understand what problem the learning addresses, then investigate whether that problem domain still exists in the codebase. The agent understands concepts — use that understanding to look for where the problem lives now, not where the old code used to be.

Before deleting: check for inbound links

A doc that other files cite is load-bearing in a way the doc itself does not announce. Before classifying as Delete, search the repo's markdown content (other docs, plans, instruction files, READMEs) for citations of the file — not source code, where citations are rare and only appear in comments. The filename slug is usually unique enough that one query covers all citation sites.

Search efficiently:

• Prefer the platform's native content-search tool (e.g., Grep in Claude Code) over shell. Drop to shell when materially better for the case.
• Search the filename slug (without .md); narrow to the full path only if matches are noisy.
• Read context lines around each match (e.g., Grep's -B/-A), not whole files.

Inbound links inform the classification, not the cleanup. Removing a citation is always mechanical (drop the parenthetical, the bare entry, or the deferring clause). The judgment is upstream: given these citations, is Delete still right, or is Replace closer to right?

Classify each citation by what it does in its citing context:

• Decorative — principle stated inline, citation is a "see also" pointer or bare attribution. Delete is fine; clean up citations in the same refresh change set.
• Substantive — citing doc relies on the cited doc to provide content not stated inline (e.g., "see X for details on Y" with no inline Y). Signal Replace — write a successor at the same path, or Keep with narrowed scope if the doc's actual content is broader than its title implies.
• Mixed or unclear — stale-mark.

In headless mode, Delete + decorative cleanup is fine. Any substantive citation, or any genuine ambiguity, downgrades to stale-marking — writing a Replace successor is judgment-heavy and should not happen unattended.

Auto-delete only when all three hold:

• The implementation is gone (or fully superseded by a clearly better successor, or the doc is plainly redundant).
• The problem domain is gone — the app no longer deals with what the learning addresses.
• Inbound links are absent or unambiguously decorative.

If any condition fails, classify as Replace, Update, Consolidate, or stale-mark per the rules above. Do not delete a learning whose problem domain is still active or whose principles are cited substantively — fill the gap with a replacement instead.