team-boot

Overview

Assembles team AI directives context and injects it into the system prompt at session start. The CDR index lists all available team context modules with descriptors — read full module files on demand when a task matches.

team-boot injects only the always-relevant layer: constitution titles, the compact CDR index, the Class Boots catalog, the skills registry, and MCP servers. The four record classes (ADR/PDR/ChDR/CDR deep-dive) plus technology selection load on demand through the class boots below — each pairs its index injection with decision capture.

Discovery is native: the injected CDR index is matched by the LLM against the current task on its own, per prompt, with no skill invocation. /team-discover is a separate, user-invoked command for explicit structured re-discovery (e.g., starting a complex feature) and is not part of the bootstrap loop.

Fast path: if the team AI directives context (constitution, CDR index, Class Boots catalog) is already in your system prompt or first user message, the event hook already ran — do nothing.

Class Boots

Invoke a class boot when a task or decision matches its row — invoke it at the START of the matching task, before planning the todo list and before implementation, so the class context informs planning. Never defer to session end. When no local memory index exists in the current working directory and the directory sits inside a workspace (detected via a .gitmodules marker in an ancestor), the boot reads the workspace root's docs/adlc/memory/ index instead. Each boot emits its class context section and its own searched line (_Searched N <class> records, K matched._), and carries the full detection and capture guidance for its class. 0 rows matched → the boot emits its section heading + the searched line only — no table.

Event hook (automatic)

The session_start and session_compact event hooks (declared in .events.json, wired by adlc-cli) run scripts/bash/boot.sh (POSIX) or scripts/powershell/boot.ps1 (Windows), which reads .adlc/init-options.json, assembles the context block (constitution, CDR.md index table, .skills.json), and outputs it to stdout. The plugin caches the result and pushes it into the system prompt on every step (idempotent — same cached content, no accumulation). The session_compact declaration makes post-compaction re-injection part of the contract: when a harness summarizes history, the generated plugin re-runs the handler and the dedup guard prevents double-injection. Agents whose adapters don't map session_compact yet skip it — adlc-cli owns those mappings.

Manual fallback (agents without event support)

1. Read .adlc/init-options.json from the current working directory. Do NOT walk up parent directories. Do NOT use glob, find, or any file-search tool to locate it.
2. If unconfigured (missing, null, or path doesn't exist): invoke the team-setup skill.
3. If the factory-setup skill is installed in this session and no .adlc/issues-provider.yml resolves in the repo root: ask the user whether to run /factory-setup now (provider pick is questions-driven, same as team-setup modes). On decline, skip for this session — there is deliberately no persistent opt-out marker.
4. If configured: assemble the context and follow the Class Boots catalog above. Full walkthrough in references/manual-fallback.md.

Decision Capture

Detect decisions as they emerge and suggest the matching capture skill to the user — never write drafts or touch .adlc/drafts/. The capture skill writes the draft when the user runs it; the clarify skill gates it afterwards.

Detection Triggers

Proportionality gate and trust model in references/decision-capture.md: match documentation depth to how non-obvious the decision is, and synthesize project knowledge — never transcribe instructions.

Team Context & Decisions (every response)

Every response carries ONE merged section — grounding context rows and owed decision rows in the same 6-column table:

## Team Context & Decisions

| ID | Name | Type | Rel | Status | Clarify |
|--|--|--|--|--|--|
| CDR-YYYY-NNN | <name> | <type> | <relevance> | in use | — |
| — | <decision> | <ADR/PDR/CDR/ChDR/Eval> | <trigger> | pending | <clarify skill> |

_Scope: N CDRs · A ADRs · P PDRs · C ChDRs · E evals · M skills — J rows shown · Unrecorded: N pending · Unclarified: M captured drafts._

Status in use + Clarify — = grounding context (accepted records only). Status pending/captured/clarified/handed off + Clarify skill = owed decisions. No draft row may carry Status in use — drafts appear only as pending decision rows. J = all rows shown.

• Detect: match session decisions against triggers above.
• Classify: assign record type (ADR/PDR/CDR/ChDR).
• Suggest: name the capture skill in one line — _Suggested capture: <decision name> → run <capture skill>._ Never write drafts or touch .adlc/drafts/; the capture skill writes the draft when the user runs it.
• Track: update the table row (ID = draft ID or —, Status = pending/captured/clarified/handed off, Clarify = matching clarify skill).
• Surface: mirror each detected decision as a task-list todo (suggest capture → matching clarify skill at session end). After code-modifying tasks, add a trailing todo to sweep Team Context & Decisions until Unrecorded: 0 pending · Unclarified: 0 drafts (a draft leaves Unclarified only via its clarify skill or an explicit user handoff to a named clarify or execute skill).
• Session-end: before closing, deliver the clarify prompt naming each uncaptured decision (ID + capture skill) and each captured draft (ID + clarify skill); if the user defers clarify, mark those rows handed off.

Capture goes through the class capture skills (/architect-specify, /product-specify, /change-specify, /change-init for mined history, /team-levelup, /evals-specify) — the boot suggests, the user runs. The clarify skills (/architect-clarify, /product-clarify, /change-clarify, /evals-clarify) review whatever drafts exist at session end.

Failure Handling

• Missing index + missing records → emit the section heading + _Scope: 0 CDRs · 0 ADRs · 0 PDRs · 0 ChDRs · 0 evals · 0 skills — 0 rows shown._ only — no table — and continue the user's task; never block.
• Unparseable index rows → skip malformed rows, note the skip count.

Verification

• Team Context & Decisions emitted with _Scope: N CDRs · A ADRs · P PDRs · C ChDRs · E evals · M skills — J rows shown · Unrecorded: N pending · Unclarified: M captured drafts._ (J = all rows shown; Status=in use rows are accepted records only, decision rows carry a clarify skill).
• Class boots fired at task start (before todo planning), never deferred; workspace-root fallback engaged when CWD has no local memory.
• Detected decisions added as table rows with the matching clarify skill.
• No drafts written by the boot; each detected decision carries a _Suggested capture: … line naming its capture skill.
• Detected decisions mirrored as task-list todos; trailing ledger-sweep todo added after code-modifying tasks; sweep closed only at Unrecorded: 0 pending · Unclarified: 0 drafts with the session-end clarify prompt delivered.

Unconfigured projects

Invoke team-setup to configure team AI directives for this project.

If factory skills are installed but no provider config resolves, ask whether to run factory-setup.