Reporter Notes

Daily Article Notes — 2026-03-24

Chosen angle

**The next CLI UX battle is agent forensics**

Thesis: both Codex and Gemini CLI are adding the plumbing needed to reconstruct agent behavior after the fact. Codex is measuring turns as debuggable cost/action units; Gemini is exposing how context was discovered and attaching trace identifiers to edit offers.

Candidate angles considered

1. **Agent observability moves inside the runtime**

• Strong evidence, but a bit generic.

2. **Context becomes measurable**

• Good, but too close to the 2026-03-14 context-sensor piece.

3. **The next UX battle is agent forensics** ✅

• Freshest frame, best reader payoff, least overlap.

Why this is different from yesterday

Yesterday was about execution context becoming first-class (forks vs threads, explicit loop context).

Today is about **post-hoc reconstruction**: cost, tool activity, discovered context, and traceable edit offers.

Grounding evidence

Codex

• Repo: openai/codex
• GitHub repo metadata via gh repo view
• pushedAt: 2026-03-24T05:00:35Z
• stars: 67170
• Commit e07eaff0d32c — feat: add metric for per-turn tool count and add tmp_mem flag (#13456)
• emits codex.turn.tool.call
• tags metrics with tmp_mem_enabled
• Commit 49634b7f9c07 — add metric for per-turn token usage (#13454)
• snapshots total token usage at turn start
• diffs totals at turn end to emit per-turn histograms
• Code evidence from codex-rs/core/src/tasks/mod.rs
• lines ~243-301 emit:
• codex.turn.tool.call
• codex.turn.token_usage with token_type=total/input/cached_input/output/reasoning_output
• all tagged with tmp_mem_enabled

Gemini CLI

• Repo: google-gemini/gemini-cli
• GitHub repo metadata via gh repo view
• pushedAt: 2026-03-24T04:59:00Z
• stars: 98849
• Commit 7b4a822b0eba — feat(core): instrument file system tools for JIT context discovery (#22082)
• adds discoverJitContext()
• appends --- Newly Discovered Project Context --- to tool outputs
• wired into file tools like read-file
• Commit d7d53981f3c1 — feat(core): add trajectoryId to ConversationOffered telemetry (#22214)
• recordConversationOffered(... trajectoryId)
• createConversationOffered(... trajectoryId)
• includes trajectoryId in returned telemetry payload
• Code evidence from:
• packages/core/src/tools/jit-context.ts
• packages/core/src/tools/read-file.ts
• packages/core/src/code_assist/telemetry.ts

gsio evidence

Query: tmp_mem tool count token usage

Top codex hits included:

• e07eaff0d32c — per-turn tool count + tmp_mem flag
• 49634b7f9c07 — per-turn token usage metrics

Query: JIT context trajectory telemetry parallel tools

Top gemini-cli hits included:

• 7b4a822b0eba — shared JIT context discovery utility
• d7d53981f3c1 — trajectoryId on ConversationOffered telemetry
• additional tracing / context / memory commits clustered nearby

Web data used

• GitHub commit pages fetched for:
• openai/codex@e07eaff0d32c...
• google-gemini/gemini-cli@7b4a822b0eba...
• google-gemini/gemini-cli@d7d53981f3c1...
• GitHub repo pages / API metadata via gh repo view

llm review summary (gpt-5.4)

Winner: **The next UX battle is agent forensics**

Reason: freshest framing; avoids overlap with context-sensor and execution-context pieces by centering on after-action explainability.

Sources — 2026-03-24

Repo metadata

• gh repo view openai/codex --json name,description,url,defaultBranchRef,stargazerCount,pushedAt
• gh repo view google-gemini/gemini-cli --json name,description,url,defaultBranchRef,stargazerCount,pushedAt

Git history and code

• git -C projects/agentic-workflows-gemini-vs-codex/repos/codex log --oneline -n 20
• git -C projects/agentic-workflows-gemini-vs-codex/repos/gemini-cli log --oneline -n 30
• git -C .../codex show --stat --summary e07eaff0d32caaec47cc4b0616d1b32c086e904f
• git -C .../codex show --stat --summary 49634b7f9c0702598bbd5a233ddcb8bad5302e79
• git -C .../gemini-cli show --stat --summary 7b4a822b0ebaa44f6fe12dd1acc6f956d39cfc1e
• git -C .../gemini-cli show --stat --summary d7d53981f3c12e5d00963cb9a276732857ec53e8
• sed -n '1,220p' .../gemini-cli/packages/core/src/tools/jit-context.ts
• sed -n '1,220p' .../codex/codex-rs/core/src/tasks/mod.rs
• sed -n '1,220p' .../gemini-cli/packages/core/src/code_assist/telemetry.ts
• targeted line extraction for read-file.ts, telemetry.ts, tasks/mod.rs

gsio

• gsio search q 'tmp_mem tool count token usage' -p openai/codex --output summary
• gsio search q 'JIT context trajectory telemetry parallel tools' -p google-gemini/gemini-cli --output summary

Web fetches

• https://github.com/openai/codex/commit/e07eaff0d32caaec47cc4b0616d1b32c086e904f
• https://github.com/google-gemini/gemini-cli/commit/7b4a822b0ebaa44f6fe12dd1acc6f956d39cfc1e
• https://github.com/google-gemini/gemini-cli/commit/d7d53981f3c12e5d00963cb9a276732857ec53e8

Model review

• llm -m gpt-5.4 with candidate angles A/B/C and overlap constraints from past articles