/recall — Interactive Recall Pass

Load relevant project decisions and learnings into session context before working. The interactive equivalent of the pipeline's recall passes (design A.1, runbook Phase 0.5).

Why This Exists

Agents self-retrieve at ~3% rate. The pipeline injects recall at fixed points, but interactive sessions bypass it. This skill fills the gap: front-load relevant decisions so they're available without the agent needing to "decide to recall."

Invocation

/recall                     # Topic-based: select section entries relevant to current discussion
/recall <topic>             # Explicit topic: "recall error handling"
/recall deep                # Saturate: tail-recurse aggressively, full 4-pass depth
/recall broad               # Whole files: load entire decision files, not individual sections
/recall all                 # Deep + broad: maximum effort within topic scope
/recall everything          # Full corpus: load all entries (expensive, rare)

Process

Pass 1: Scan and Select

1. Identify the active topic from conversation context (or explicit argument)
2. Read agents/memory-index.md (skip if already in context). Review entries for domain matches.
3. Select targets based on mode:
   • Default/deep: select relevant entry triggers (section-level)
   • Broad/all: select relevant ## agents/decisions/<file>.md headers (file-level)
4. Skip already-loaded content — if a section or file is already in context from a prior /recall, /when, or /how invocation, do not reload it.

Pass 2: Resolve and Load

Resolution method depends on mode:

Section-level (default, deep): Batch-resolve via claudeutils _recall resolve:

claudeutils _recall resolve "when <trigger>" "how <trigger>" ...

File-level (broad, all): Read decision files directly using the Read tool:

Read agents/decisions/testing.md
Read agents/decisions/implementation-notes.md

Batch multiple Read calls in a single message. Line limits are enforced by precommit — no cap logic needed at read time.

Pass 3: Tail-Recurse or Exit

After loading, re-check memory-index (already in context from Pass 1) for targets that became relevant due to newly loaded context (e.g., a decision file references another pattern).

• New targets found → go to Pass 2 with the new targets
• No new targets → stop (saturation reached)

Depth by mode:

• Default: 1 pass (select all relevant entries from index in a single batch). Pass 2 only if loaded content reveals entries not identifiable from index headings alone
• Deep/all: up to 4 passes (aggressive re-scan, include tangential entries)
• Broad: 1 pass (whole files already capture internal connections)
• Everything: no recursion (full corpus loaded in one pass)

Output

After completion, summarize what was loaded:

Recalled N entries from M decision files:
- <Entry Name> (source: decisions/<file>.md)
- <Entry Name> (source: decisions/<file>.md)
- ...

If tail-recursion added entries beyond the initial selection, note which pass found them.

Cumulative Behavior

Each /recall invocation builds on prior ones within the session:

• Check which entries are already in context (by their heading + source path visible in prior conversation)
• On re-invocation, skip already-loaded entries
• New topic → new relevant entries selected → loaded on top of existing context
• The conversation window IS the accumulation mechanism (skill continuation)

Modes

/recall (default): Select section-level entries relevant to current discussion topic. Single pass typical — tail-recurse only if loaded content reveals clearly relevant entries not yet loaded.

/recall <topic>: Override topic detection. Select entries matching the explicit topic instead of inferring from conversation.

/recall deep: Aggressive saturation. Tail-recurse through the full 4-pass depth, including tangentially related entries on each re-scan. Use when the topic has deep dependency chains across decision files.

/recall broad: Whole-file loading. Instead of resolving individual section entries, identify relevant decision files from memory-index's ## agents/decisions/<file>.md headers and Read each file directly. Skip files already Read whole in a prior /recall broad or /recall all — check the output summary from prior invocations. Prior section-level loads do not count as "already loaded" (whole-file read captures content beyond the loaded sections, justifying the re-read cost). Use when the topic spans many entries within the same files — reading the whole file captures connective tissue between entries that section-level loading misses.

/recall all: Deep + broad combined. Maximum effort within topic scope. Identify relevant files, load them whole, then tail-recurse to discover and load additional files revealed by the content. The "give me everything on this topic" mode.

/recall everything: Full corpus dump. Bypass topic selection entirely — load every decision file in memory-index. Expensive (fills context). Use only when the task touches many domains or when genuinely unsure what's relevant. Not tail-recursive (nothing to discover — everything is loaded).

Relationship to /when and /how

Skill	Scope	Trigger
/when	Single entry by trigger	Agent recognizes a specific situation
/how	Single entry by trigger	Agent needs a specific procedure
/recall	Multiple entries by topic	Front-load before work begins

/recall uses the same resolution infrastructure (claudeutils _recall resolve) but selects multiple entries by topic relevance rather than resolving a single known trigger. It is the batch, proactive counterpart to /when and /how's individual, reactive lookups.