Technical · speaker script
00:00–01:00 · SLIDE 40
Open the toolbox. Choose a small working set.
SAY The current source has 96 skills, 69 roles and 33 hook scripts. The point is a smaller set of overlapping commands, with useful methods retained.
DO Read the three counts once. Distinguish a source inventory from tools actually loaded or executed.
THEN What matters is how a task moves through that library.
This module follows source 49f2d15260f096213086f02dfeb1fae6cbe62d45. Its 0.12.0 changelog is unreleased. The current census is 96 canonical skills, 69 agent roles and 33 hook scripts. Generated wrappers are not extra canonical skills, and a file on disk is not proof that a host discovered or executed it.
The consolidation changed entrypoints. Inspection now uses inspect; verification uses verify; investigation uses diagnose; continuity uses pause and compatible resume. It retained useful capabilities, original task authority, saved work and provider contracts. The old 127-skill library and roughly 80 aspiration are historical context, not today's delivered count or ADR-0034's numeric target.
A smaller inventory alone proves no productivity gain. Choose the working set needed for one outcome and verify the result.
01:00–02:30 · SLIDE 41
Start with the loop. Add depth on demand.
SAY One loop; explicit methods when the work needs more depth.
DO Walk the four central phases, then give inspect, verify and web-session as consolidation examples.
THEN The method stays coherent when ownership is clear.
The nine-phase lifecycle remains. PLAN preserves original requirements and task IDs; BUILD groups connected short leaves into a bounded package with one implementer and evidence for every leaf. REVIEW checks specification, quality and applicable mandatory controls. CAPTURE reconciles reviewed knowledge and the handoff. A failed leaf keeps its package open.
Use inspect for plan or repository lenses; verify is read-only by default and repairs need their actual authority; diagnose owns scoped investigation; cross-check uses a real independent reviewer. TA, DA, SC, DH and TQ remain, as does full-engineering-pass.
web-session brings browse, scrape, open and cookies behind explicit modes without replacing their actual providers. Frontend advice, variants and built-UI review remain under the design owners; generate-docs, generate-web and generate-pdf retain their outputs. Document-format acceptance remains staged. PDF printing is not PDF content or visual verification.
02:30–04:00 · SLIDE 42
Clear ownership. Independent challenge.
SAY A role carries expertise. Independent review needs a separate, attributable reviewer.
DO Follow preparation, implementation and challenge. Separate the MARS review panel from the Swarm implementation diagram.
THEN All of that work needs a place to leave its evidence.
One implementer owns a bounded package. Reviews cover every original leaf and their integration. Substantive work needs actual independent review; a role name, another model name or a builder's self-review cannot manufacture it. Mechanical work can use the documented coordinator review.
The shared Review Method gives single REVIEW and each MARS participant the same frozen packet body: acceptance, standing questions, evidence levels, severity and coverage. MARS is an optional multi-model second look with consent and observed host capability. Reviewers work blind first; at most one challenge round follows contested claims. The coordinator judges source evidence rather than voting and preserves dissent. MARS itself never grants release clearance.
Swarm is different: it schedules implementation packages. One coordinator owns shared state and integration. Concurrent writers need disjoint scopes and attributable worktrees, patches or equivalent isolation. Without that, replay the briefs serially. Lane reviews never replace the final integrated review.
04:00–06:00 · SLIDE 43
Four places. Different responsibilities.
SAY Separate reusable source, configured home, project knowledge and local evidence.
DO Toggle Bare install and Repo kit. Point out which records travel through Git and which do not.
THEN Now follow those records through one cycle.
LINTEL_SOURCE_ROOT identifies trusted helpers and templates; LINTEL_REPO_ROOT is the working consumer. The configuration home is explicit. Bare installation commonly uses ~/.lintel; repository adapters default to .claude/runtime/lintel-home when no home is configured. Their shared source bundle is .github/lintel. That repository kit installs no hook bundle.
Commit reviewed project plans, decisions, lessons and handoffs intentionally. Runtime contains local cycle state, checkpoints, review/QA records, audit observations and MARS panels. A clone does not inherit ignored evidence or private packs.
Review packet inputs can live under runtime/review/<id>; strict decision records use runtime/reviews. MARS uses runtime/mars/<panel>/inputs and records. Profile contexts keep versioned current-profile.json records under the configured home. A receipt records a local operation; it does not establish host activation, independent review or enforcement.
06:00–08:00 · SLIDE 44
Each phase leaves a usable record.
SAY The record carries the selected work and its evidence into the next phase.
DO Start at PLAN. Select BUILD, REVIEW and CAPTURE. Use other tabs only if asked about paths.
THEN The same records let a fresh session find the unfinished step.
Start an explicit cycle identity and verified profile context before phase writes. The ledger carries the original map, artifact paths and policy reference. The work map indexes authority; it is not a duplicate backlog. Spec Kit keeps its original task IDs and files.
BUILD records evidence per leaf and reviews each bounded package for spec compliance and then quality. One failed leaf keeps the package open and preserves useful completed evidence. BLOCKED returns to BUILD with the exact unresolved action.
REVIEW sends the shared method packet and checks coverage, then persists content-bound v2 decisions and immutable QA obligations. A later rejecting or malformed applicable decision blocks an older pass. Missing mandatory validation or independent corroboration stays open. A complete report can still contain a failing verdict; a checker exit code alone is not a PASS.
SHIP consumes current review and QA within existing authority. CAPTURE reaffirms actual outcomes and records warranted lessons or decisions. It cannot approve missing work or authorize a deployment.
08:00–09:30 · SLIDE 45
Resume the work. Keep its identity.
SAY Pause saves continuity. Resume reads it alongside the original work and current evidence.
DO Point to the committed map, then the original artifacts. Name the missing-review boundary.
THEN Continuity helps; the actual client still decides which controls can run.
Use pause to reserve a collision-safe, owned checkpoint. The compatible filename still ends in -context-save.md; old saved data is not renamed or deleted. Verify nonempty readback before claiming it was saved. resume --from <checkpoint> reads continuity context; it does not restore code, change branches or grant BUILD or SHIP authority. A bare phase or selected job-step keeps the existing override meaning; use an explicit path such as ./BUILD if a checkpoint name collides.
Explicitly selected work wins over an unrelated local ledger or checkpoint. A fresh clone follows an unambiguous committed work map and its original artifacts, then reconciles actual checkout and acceptance evidence. It does not choose the newest file or fabricate lost evidence.
Resume verifies the original cycle, target and profile generation/digest. Missing or changed policy needs explicit reconciliation. STARTING, BLOCKED and NEEDS_CONTEXT return to the same unfinished phase. A truncated DONE cannot advance.
09:30–11:30 · SLIDE 46
Hooks react where they are registered.
SAY Thirty-three scripts have different jobs. Nine core registrations use the Claude Code protocol.
DO Toggle core and optional. Use one blocker and one warning; leave the full inventory in the reference.
THEN The same source reaches several surfaces through different discovery routes.
The census remains 33 scripts: 9 registered core across 5 event types, plus 24 optional or manual scripts. Registration is not a firing test. The native bundle uses the Claude Code protocol; repository adapters install no hooks. Copilot's own hook facility does not mean Lintel has translated this bundle.
The two Git scanners deny supported commit/push requests on detected patterns or unavailable required collection/scanning. Warning hooks surface advice; they are not permission controls. Stop does not force continuation or write a checkpoint. The optional frozen-zone warning now uses code-freeze's shared validated state under repository runtime; it does not lock the filesystem. Context-bloat advice uses pause and resume.
hooks-status reads receipts, while doctor --hooks-only diagnoses byte and activation state; neither establishes universal enforcement. Clean/no-record runs remain unobserved. The event catalog has 23 categories and 66 kinds, not 66 platform events.
Source 1 · Source 2 · Source 3 · Source 4 · Source 5 · Source 6
11:30–13:30 · SLIDE 47
One method. Four client families.
SAY Copilot, Claude, Codex and Cursor remain supported. Verify the exact surface you use.
DO Select Copilot, then Codex to show the cloud/manual distinction. Avoid reading all paths aloud.
THEN Portable discovery still needs the right project and policy context.
The registry now has 14 records: 13 named surfaces across 4 supported families, plus other as an explicit manual fallback. Twelve records have native-format discovery and two are manual: Codex cloud and other. Four roots serve the supported families. Fifteen starter workflow wrappers include MARS; Copilot retains three native role profiles.
One .github/lintel bundle supplies selected repository routes. The installer does not change host settings, credentials or permissions and installs no hooks. Removed client-family IDs have no compatibility route; other is a manual handoff, not supported integration. Existing supported plugin routes remain.
Vendor documentation, delivered files and observed execution are separate. The registry retains three partial Copilot App observations. MARS has separate live-pilot evidence: RM9 observed four model identities but collected only three valid reports, so the panel remained partial and multi_model_verified:false. That one synthetic subject, without a challenge or live single-review comparison, does not certify every client or model.
13:30–14:30 · SLIDE 48
Carry the policy with the work.
SAY Preferences choose context. Required policy must stay intact from plan through delivery.
DO Move require → resolve → verify. Point to platform permissions and CI as separately configured controls.
THEN Then preserve what the next task can actually use.
profile.yaml holds operator preferences such as role and mode. Required repository or explicit pack selection is different. The effective context binds actual source content, target, generation and digest. A pack name alone cannot establish that the same policy was applied.
Carry the verified reference through selected work, children, review, QA and recovery. Missing required packs or changed policy do not silently fall back to neutral success. A deliberate switch/rebind records a reason, retains prior history and requires affected work to be reconsidered.
Packs provide reviewed standards and context. Platform permissions, CI and live organizational enforcement remain their own configured and verified layer. Required-policy source tests are not an enterprise certification.
14:30–16:00 · SLIDE 49
Keep useful knowledge. Preserve the evidence.
SAY A lesson helps only when later work retrieves and applies it.
DO Walk capture, retrieve and preserve. Distinguish checkpoint continuity from owned-file recovery.
THEN The pilot should verify this whole path, including failures.
lessons-add checks existing entries before choosing add, update, supersede or no-op. Stable L-NNN IDs, a lock and a conditional write preserve ownership and history. There is no active operator-global lessons sink; old global JSONL is read-only legacy data. lessons-surface ranks active literal-keyword matches, then the agent reads and applies the full rule. Stored text or a ranked heading alone does not prove better outcomes.
ADRs record consequential choices and rationale, normally beginning Proposed. CAPTURE reaffirms the original mapped task outcomes and handoff rather than generating a competing backlog. Standalone retrospective and release-summary modes preserve their selected window and publication boundary.
Owned snapshots protect recent copies, the newest five and incomplete operations. Pruning needs explicit eligible IDs and refuses unexpected content. Keep transaction receipts with their snapshots. This is bounded recovery, not a universal cleanup policy or a way to compact a model conversation.
Source 1 · Source 2 · Source 3 · Source 4 · Source 5 · Source 6
16:00–18:00 · SLIDE 50
Prove the route you intend to use.
SAY Discovery, a real task, a cold restart and a failure case make the claim concrete.
DO Read the four checks and finish with accepted outcomes and overhead.
THEN Return to the main story or open the reference for questions.
Pin source and client versions, actual tools, selected profile and permissions. Verify discovery, then run a bounded task with original requirements, acceptance evidence and the required independent review. Restart without the original conversation, select the same work and report the actual unmet step.
Negative cases matter: changed policy, skipped mandatory QA, missing independent reviewer, later rejecting decision, ambiguous work map and edited managed files. For MARS, a changed brief, substituted model or incomplete panel must remain visible. For Swarm, missing attribution means serial work and missing review leaves the lane open.
The committed Universal report now records 109/109 in-scope items and its final 22-job CI run. Document acceptance removed by ADR-0033 was removed, not passed. Current source still labels the development line unreleased and records platform/runtime limitations; this presentation does not rerun or widen those acceptance claims. Historical demonstrations retain their original source pins.
Measure accepted-change time, interventions, defects, rework, recovery and total process cost under comparable conditions. Neither contribution counts, skill counts nor token guesses establish productivity.